https://me.itheheda.online/

企业官网 GEO 诊断报告

页面地址： https://me.itheheda.online/
抓取时间： 2026-09-30 05:30 UTC
页面性质： 企业 AI 系统设计与交付服务商「实解智能（SolveReal Systems）」的官网首页

一、页面定位与核心信息

该页面是实解智能的官网首页，核心定位为：面向中小企业老板与业务负责人，提供企业 AI 系统设计与交付服务。页面明确强调"从真实业务问题开始，不从功能清单开始"，并给出四步交付流程（业务输入 → AI 处理 → 人工审核 → 稳定交付）。

页面同时承担两个功能：一是作为企业服务介绍页，二是作为 GEO（生成式引擎优化）诊断的演示/体验入口——页面顶部设有"提交你的业务问题 体验企业官网 GEO 诊断"按钮。

二、品牌事实与可验证信息

页面提供了以下品牌直接答案：

注意： 页面未提供公司注册主体、办公地址、团队规模、成立时间、联系方式（电话/邮箱）等企业资质信息，也未提供任何客户案例的联系方式以供核验。品牌负责人仅以"fengc"代称，未给出全名或公开身份信息。

三、服务能力与项目证据

四个系统方向（Capabilities）

1. 多来源信息汇总 —— 把分散信息形成统一、可追踪的结构化底稿（常见场景：资料散落在文档、表格和业务网页里）
2. 重复判断与录入 —— 把稳定规则交给系统，例外与高风险决策保留给人工（常见场景：同类信息每天重复分类、核对、再录入其他系统）
3. 系统之间的数据接力 —— 跨文件与业务系统传递状态，减少人工搬运和断点（常见场景：文件、任务和处理结果需要在不同工具之间反复传递）
4. 异常恢复与人工审核 —— 记录异常、保留恢复点，让系统真正可持续运行（常见场景：流程中断后需要知道停在哪里、谁来接管、如何继续）

项目证据（Project evidence）

页面列出四个项目，其中三个标注为"可部署系统"，一个为"公开产品"：

1. Open GEO Console（公开产品）
2. Hermes Notebook（可部署系统）
3. Freight Lead Agent（可部署系统）—— 当前展示项目
4. Codex Feishu Bridge（可部署系统）

当前展示项目详情：Freight Lead Agent

• 功能： Google 地图获客与定制营销系统
• 改造前： 销售人员需从 Google 地图找目标企业，逐个确认官网、读取业务内容、寻找公开联系方式，再为每家公司单独准备开发邮件；难以靠人工持续放大
• 系统接管： 按关键词和地区采集 Google 地图企业，进入官网提取业务信息与公开联系方式，结合每家公司的网站内容生成定制邮件，再交给销售审核、发送和跟进
• 人工保留： 逐条确认目标企业、官网内容、公开联系方式和定制邮件是否准确
• 恢复与边界： 无效记录不会被删除，而是保留明确原因供复核
• 主要限制： 只处理公开企业信息，不抓取登录后页面或个人资料

注意： 页面仅展示项目名称和描述，未提供可访问的项目演示链接、代码仓库地址、部署文档或客户验收记录。所有项目证据均为页面自述，缺乏第三方可验证的落地证据。

四、交付方法（Delivery method）

页面给出四步交付流程：

1. 诊断现状 —— 还原真实输入、处理步骤、责任人、异常与交付结果
2. 设计人机边界 —— 明确 AI 可以处理什么、关键决策由谁审核、哪些动作必须人工确认
3. 验证真实流程 —— 使用真实样本验证输入、异常、恢复、审核与输出，而不是只演示理想路径
4. 交付并持续优化 —— 交付可运行系统、操作边界和追踪证据，再根据真实使用调整

五、内容资产与 Field notes

页面包含三篇标注日期的文章（均为 2026 年 9 月）：

1. 2026-09-21 《官网自己说法不一时，GEO 应该先改哪一页？》—— 讨论官网多页面说法不一致时的处理优先级
2. 2026-09-21 《GEO 和 SEO 到底有什么区别？一张可执行的工作地图》—— 对比 GEO 与 SEO 在观察结果、优化单位和验收方法上的差异
3. 2026-09-18 《官网被 AI 引用以后，怎么判断 GEO 改动有没有用？》—— 提出将回答准确性、引用支持度、访问发生和咨询有效性分开记录来判断 GEO 效果

此外页面还提到公众号「独立系统」，记录企业 AI 系统、自动化与独立开发实践，并提供"进入文章库"入口。

六、GEO 表现评估

优势

1. 品牌直接答案完整： 页面在"实解智能是谁？"区块中直接给出了品牌名、英文名、负责人、服务内容、服务对象和交付原则，结构清晰，便于搜索引擎和 AI 模型提取。
2. 结构化信息密度高： 四个能力方向、四个项目、四步交付方法均以编号列表呈现，语义层次分明。
3. 项目证据含限制说明： Freight Lead Agent 明确写出"主要限制：只处理公开企业信息，不抓取登录后页面或个人资料"，这种诚实的边界描述有助于建立可信度。
4. 内容时效性好： 三篇文章均为近期发布（2026 年 9 月），且主题直接围绕 GEO 实践，与页面定位高度一致。
5. 中英双语品牌名： 同时提供中文和英文品牌名，有利于多语言检索。

不足与风险

1. 企业资质信息缺失： 无公司注册主体、办公地址、联系方式、团队规模等基础信任信息。对于一家宣称服务中小企业的 B2B 服务商，这是较大的信任缺口。
2. 负责人身份不透明： 仅以"fengc"代称，无全名、职位、背景介绍或公开履历，难以建立个人品牌信任。
3. 项目证据不可验证： 四个项目均无外部链接、演示地址、代码仓库或客户背书。页面自称"可部署系统"，但未提供任何可操作的验证路径。
4. 无客户案例与评价： 未提供任何客户名称、行业、规模或使用反馈，所有服务承诺均为自述。
5. 联系方式单一： 仅提供公众号「独立系统」作为内容入口，未给出直接联系渠道（电话、邮箱、表单），"提交你的业务问题"按钮的具体去向未在页面中说明。
6. 页面语言单一： 仅提供中文内容，无英文版本。考虑到其英文品牌名 SolveReal Systems 的存在，若面向海外客户或国际检索，存在语言覆盖缺口。
7. 域名可信度： 域名 me.itheheda.online 使用 .online 顶级域且为子域名，与品牌名"实解智能/SolveReal Systems"无直接对应关系，可能影响用户信任和搜索引擎对品牌一致性的判断。

七、结论与建议

总体判断： 该页面在内容结构、信息组织和 GEO 意识方面表现良好，品牌直接答案清晰，交付方法论完整，且主动披露了项目限制，具备较好的 AI 检索友好性。但作为一家 B2B 企业服务商的官网，信任证据严重不足——缺乏可验证的企业资质、项目演示、客户案例和直接联系方式，这将是影响转化和长期 GEO 效果的主要瓶颈。

优先改进建议：

1. 补充企业基本信息（注册主体、联系方式、办公地址），并在页面中明确"提交你的业务问题"表单的提交去向和响应方式。
2. 为四个项目提供可访问的演示链接、部署文档或代码仓库地址，使"可部署系统"的说法可被验证。
3. 增加客户案例（可匿名化处理）或第三方评价，增强社会证明。
4. 考虑将域名迁移至与品牌名一致的独立域名，提升品牌一致性和信任度。
5. 若面向国际客户，建议增加英文版页面。

本报告基于 2026-09-30 抓取的页面内容进行分析，页面后续更新可能导致部分信息失效。

Service Facts · 可引用服务事实 — Page Analysis

1. Page Overview

This is the services page of a provider offering enterprise AI automation services and workflow system delivery (企业 AI 自动化服务与工作流系统交付). The page is written in Chinese and targets small and medium-sized enterprises (SMEs) that face problems with repetitive manual work, cross-system data flow, or processing of unstructured documents.

The core positioning is stated clearly: the provider first diagnoses the real process, then designs the human-machine boundary, exception recovery, and acceptance methods, and finally delivers a custom AI automation system that fits the enterprise's actual operations.

2. Target Audience and Eligibility Criteria

The page defines who should consider a diagnostic engagement:

Situations suitable for initial diagnosis:

• Business processes that are high-frequency, relatively stable in rules, and require cross-file or cross-system flow
• Processes where AI handles unstructured information but key results still require human review
• Automation projects that need to preserve state, exceptions, recovery records, and auditable output

What the client must provide before starting:

• Real steps, frequency, and responsible persons of the current process
• Input, output, and exception samples that can be used for verification
• Key decisions or high-risk actions that must retain human confirmation

3. Operating Model (01)

The page defines a clear division between system and human responsibilities:

• System handles: Relatively stable, repeatable verification tasks — sorting, judgment, and cross-system flow
• Human review: Key decisions, high-risk actions, and exception situations retain explicit human confirmation and takeover entry points
• Delivery evidence: A runnable system, operational boundaries, and traceable records of status, exceptions, recovery, and output

4. Delivery Method (02)

The delivery process consists of four stages:

1. Diagnose current state — reconstruct real inputs, processing steps, responsible persons, exceptions, and delivery results
2. Design human-machine boundary — clarify what AI can handle, who reviews key decisions, and which actions require manual confirmation
3. Validate with real processes — use real samples to verify inputs, exceptions, recovery, review, and output, rather than only demonstrating the ideal path
4. Deliver and continuously optimize — deliver a runnable system, operational boundaries, and tracking evidence, then adjust based on actual usage

Important caveat: Specific functions, timelines, and deliverables are determined by process diagnosis and real-sample validation; an initial submission is not an automatic quote.

5. Standard Deliverables and Acceptance Criteria (03)

Six standard deliverables are listed, with the note that specific scope adjusts per project:

1. Current process and system boundary — steps, responsible persons, inputs/outputs, exceptions; what the system takes over and what humans retain
2. Runnable system and deployment results — a runnable system, or reproducible deployment package, configuration, and run entry point — not just a demo document
3. Operating instructions and human review entry — daily operations, required permissions, manual confirmation, exception takeover, and responsibility assignment
4. Exception and recovery paths — exception records, recovery points, retry or manual takeover methods, to avoid restarting from scratch after interruption
5. Acceptance samples and result records — mutually agreed real samples recording expected vs. actual results, exceptions, and traceable output
6. Handover and follow-up support scope — system ownership, maintenance, change scope, and follow-up support defined in the project agreement; no fixed service duration is implicitly promised

6. Operating Boundary: Data, Permissions, and Runtime (04)

The page emphasizes that data and permissions are part of solution design and acceptance, not post-delivery add-ons. The following must be confirmed in the real environment before implementation:

• Necessary data and real samples: Only samples necessary to verify the real process are requested; data sensitivity, purpose, and scope of use are confirmed before starting
• Deployment and processing location: Local, private environment, or cloud solutions are determined jointly by enterprise systems, data requirements, and operations conditions — no one-size-fits-all commitment
• Least privilege and account control: The system only uses permissions needed for the agreed process; where conditions allow, the enterprise retains control of accounts, tokens, and final authorization
• Retention, backup, and deletion: Data retention periods, copies, backups, and deletion methods are specified per project; unspecified handling is not treated as an implicit commitment
• Manual authorization and high-risk actions: External sending, payments, publishing, or other high-risk actions must retain manual authorization or an explicit approval entry point
• Compliance and security statements: Without agreement, review, or certification, technical implementation is not described as a legal compliance conclusion or security certification

7. Business Scenarios (05)

Four specific business scenarios are presented, each with a named system:

1. Enterprise website GEO diagnosis and AI search visibility remediation — Open GEO Console: Diagnoses AI search visibility for corporate websites, organizing technical foundations, buyer questions, public answers, and citation evidence into management decision reports and supplier task packages
2. Enterprise data organization and AI application engine — Hermes Notebook: Organizes enterprise materials scattered across PDF, Word, Excel, PPT, images, local folders, and business web pages into searchable knowledge with sources; API access enables customer service bots, internal knowledge Q&A, and business assistants
3. Google Maps customer acquisition and custom marketing system — Freight Lead Agent: Discovers target enterprises from Google Maps in batches, reads each company's website business and public contact information, prepares custom marketing emails based on each company's website content, and integrates review, sending, and reply follow-up
4. Enterprise AI office collaboration system — Codex Feishu Bridge: Integrates Codex into enterprise Feishu. Members can submit questions, files, and long-term tasks from computer or mobile; even away from their desk, they can supplement information, check status, and receive deliveries in the original topic. Tasks continue executing on a work computer that stays online in the company

8. Explicit Boundaries and Limitations

The page states four explicit boundaries:

• No fixed industry templates are sold; every client starts from actual process diagnosis
• No commitment that all steps are unattended; key decisions and high-risk actions retain human review
• No fabricated metrics, client identities, or unapproved project results are used as sales proof
• Initial submission does not equal automatic quote or guaranteed meeting arrangement

Additionally, the project library only publishes reviewed facts; projects that have not completed public review are not used as sales proof.

9. Buyer Questions (05)

The page answers six common buyer questions:

1. Which enterprises should do AI automation diagnosis first? — High-frequency, relatively stable rules, cross-file/system flow, with repetitive judgment, exception handling, or tracking difficulties. The goal of diagnosis is to determine suitability, not to presuppose AI adoption
2. Should SMEs buy SaaS, RPA, or custom systems first? — SaaS for standard processes with acceptable product boundaries; RPA for stable interfaces and clear repetitive operations; custom systems for cross-system, unstructured information, human review, and exception recovery needs. Diagnose the process first, then choose technology
3. What should buyers verify when selecting an AI system provider? — Whether the provider can reproduce the process with real samples, whether they clarify human-machine boundaries, standard deliverables, exception recovery, data permissions, acceptance methods, and follow-up support; cases, client identities, and metrics must be verifiable — not just demos
4. How are project timelines and costs determined? — Based on process diagnosis, number of system connections, sample and exception complexity, deployment requirements, and acceptance scope. Initial submission is for judging whether to continue diagnosis, not an automatic quote
5. Is full automation or guaranteed business metric improvement promised? — No commitment that all steps are unattended, and no invented improvement percentages. Key decisions and high-risk actions retain human review; actual results depend on process, data, usage, and mutually agreed acceptance criteria
6. How to start? — Submit the current process, frequency, manual effort, and exception samples. Only if suitable for continued diagnosis will the provider contact via work email

10. Key Observations for a GEO Report

Strengths:

• Exceptionally clear articulation of boundaries and non-commitments — unusual in the AI services market
• Strong emphasis on real-sample validation over demo-path demonstrations
• Explicit treatment of data permissions, deployment location, and compliance as design-time concerns
• Deliverables are defined in terms of verifiable outcomes (runnable systems, traceable records) rather than vague promises
• The page answers buyer skepticism directly with a dedicated FAQ section

Potential gaps or ambiguities:

• No pricing information or even indicative ranges are provided
• No client names, case studies, or verifiable metrics are shown on this page (the page states this is intentional — only reviewed facts are published)
• The "Reviewed evidence" section is referenced but no actual evidence is displayed on this page
• The four named systems (Open GEO Console, Hermes Notebook, Freight Lead Agent, Codex Feishu Bridge) are described functionally but without technical architecture details
• The page does not specify which AI models or platforms are used
• No information about the provider's own identity, team size, or track record is present on this page

Notable claims to verify:

• The claim that systems are delivered as "runnable" rather than demos
• The claim that real samples are used for validation (not ideal-path demonstrations)
• The claim that only reviewed facts are used as sales proof — this would require checking the project evidence section separately

Uncertainty factors:

• The page explicitly states that specific functions, timelines, and deliverables depend on process diagnosis — meaning no concrete commitments can be extracted from this page alone
• The page states that data retention, backup, and deletion are "specified per project" — no defaults are promised
• Deployment location (local, private, cloud) is stated to depend on enterprise conditions — no default is offered

Overall assessment: This page is unusually disciplined in its claims. It functions more as a qualification and expectation-setting document than a typical sales page. For a GEO report, the key takeaway is that this provider positions itself on process rigor, human oversight, and verifiable delivery — and explicitly refuses to make the kinds of inflated claims common in the AI services space. The absence of pricing, case studies, and provider identity on this page means those must be sourced elsewhere.

实解智能 · 企业 AI 系统设计与交付 — 页面分析报告

一、页面概述

本页为实解智能（me.itheheda.online）的“关于/合作方式”介绍页，核心定位是企业 AI 系统设计与交付服务。页面主张“让 AI 真正在企业里跑起来”，强调从流程诊断出发、以真实数据验证、明确人机边界，而非销售现成模板或承诺全自动化。

页面同时包含明确的边界声明（不做什么），以及合作流程（四步法）和项目负责人信息。

二、核心信息提取

1. 服务定位

• 从流程诊断开始，识别可自动化环节、人工审核边界、失败恢复方式和可交付结果，再实现可运行的定制系统。
• 适合处理的问题类型：
  • 高频、规则相对稳定、需要跨文件或跨系统流转的业务流程；
  • 需要 AI 处理非结构化信息，但关键结果仍需人工审核的流程；
  • 需要保存状态、异常、恢复记录和可审计输出的自动化改造。

2. 明确边界（不承诺事项）

• 不销售固定行业模板；每个客户从实际流程诊断开始。
• 不承诺所有步骤无人值守；关键决策和高风险动作保留人工审核。
• 不使用虚构指标、客户身份或未经批准的项目结果作为销售证明。
• 初步提交不等于自动报价或必然安排会议。

3. 项目负责人

• 实解智能目前由 fengc 负责。
• 前期沟通、流程梳理、系统设计和交付协调由同一负责人推进。
• 客户无需先准备完整需求书，只需讲清正在消耗人力的流程、现有工具和遇到的问题。

4. 合作流程（四步）

1. 诊断现状 — 还原真实输入、处理步骤、责任人、异常与交付结果。
2. 设计人机边界 — 明确 AI 可处理范围、关键决策审核人、必须人工确认的动作。
3. 验证真实流程 — 使用真实样本验证输入、异常、恢复、审核与输出，而非只演示理想路径。
4. 交付并持续优化 — 交付可运行系统、操作边界和追踪证据，再根据真实使用调整。

5. 合作起点判断

• 先确认问题是否值得做：看流程发生频率、人工投入和出错后的影响。
• 再划分系统与人工职责：AI 接管稳定步骤，关键判断保留人工确认。
• 最后用真实数据验证：通过后再部署，并根据实际使用继续调整。

6. 用户行动入口

• 页面提供“提交一个业务问题”入口，要求提交当前流程、频率和人工投入。
• 明确说明：会先判断是否适合改造，再决定是否进入初步诊断。

三、页面可信度与透明度评估

积极信号：

• 边界声明具体且克制，主动排除“全无人值守”“固定模板”“虚构案例”等常见行业夸大话术。
• 明确“初步提交不等于自动报价或必然安排会议”，降低了用户对下一步的预期落差。
• 负责人实名（fengc），且承诺同一人全程跟进，减少沟通链条。
• 验证环节强调“真实样本”和“非理想路径”，说明交付方法有一定工程严谨性。

需注意的局限：

• 页面未提供任何过往项目案例、客户名称、量化成果或技术架构细节（页面虽有“查看项目案例”链接，但本页未展示案例内容）。
• 未说明定价模式、项目周期或交付物具体形态。
• 未提供公司注册信息、团队规模或除 fengc 以外的组织背景。
• 域名（me.itheheda.online）为个人化子域名，非标准企业域名，品牌背书有限。

四、适用场景判断

该服务较适合以下情况：

• 企业已有明确、高频、规则相对稳定的业务流程，且当前依赖大量人工跨文件/跨系统处理；
• 对 AI 输出有审计和人工审核需求，而非追求全自动；
• 愿意先提交流程描述、接受“先判断是否值得做”的筛选机制。

不适合的情况：

• 需要现成行业方案或快速报价；
• 期望全无人值守的自动化；
• 需要查看过往客户案例或量化 ROI 证明后再决策（本页未提供此类材料）。

五、结论摘要

实解智能定位为定制化企业 AI 系统设计与交付服务商，方法上强调流程诊断、人机边界划分和真实数据验证，边界声明清晰、克制。页面信息足以让潜在客户理解其合作方式和筛选逻辑，但缺乏案例、定价、公司背景等关键决策信息。建议有意向者先通过“提交业务问题”入口说明自身流程，以获取进一步判断；在进入初步诊断前，可主动询问过往项目经验、交付周期和费用结构。

不确定性提示： 本报告仅基于当前页面内容分析，未验证页面之外的公司资质、案例真实性或服务能力。

Page Analysis: Service Facts — Enterprise AI Workflow System Design and Delivery

Overview

This page presents a service offering for custom AI workflow automation systems targeting small and medium businesses. The core proposition is diagnosing existing workflows and delivering AI systems with explicit human oversight boundaries, rather than promising full automation. The page emphasizes process transparency, evidence-based delivery, and clear limitations.

Key Service Elements

Target customer profile: Small and medium businesses dealing with repetitive manual work, cross-system handoffs, or unstructured information.

Core offering: A custom AI automation system designed for day-to-day operation, delivered through a four-stage method: workflow diagnosis, human boundary design, real-sample validation, and delivery with ongoing improvement.

Human oversight model: The system handles organization, decisions, and cross-system handoffs where rules are stable and testable. Key decisions, high-risk actions, and exceptions retain explicit human confirmation and takeover points. High-risk actions such as external sending, payment, and publication require human authorization.

Delivery Contract

The page specifies six standard deliverables, though it notes exact scope varies by engagement:

1. Current workflow and system boundary mapping
2. Working system or reproducible deployment package (explicitly not just a presentation)
3. Operating guide with human review points
4. Exception and recovery paths (so interruptions don't require restarting)
5. Acceptance samples with expected and actual results
6. Handover and support scope (without implying a default service-level commitment)

Operating Boundaries

Data and access are treated as part of solution design, not post-delivery appendices. Key constraints include:

• Only necessary samples are requested, with sensitivity and permitted use agreed in advance
• Deployment location (local, private, or cloud) is chosen based on company systems and requirements
• Least-privilege access; company retains control of accounts, tokens, and final authorization where practical
• Retention, backups, and deletion defined per engagement; unspecified handling is not a default commitment
• Technical implementation is not presented as legal compliance or security certification without relevant agreement or review

Business Scenarios

Four concrete deliverable systems are presented:

1. AI search visibility diagnosis (GEO Console) — executive report and vendor task package for enterprise website visibility in AI search
2. Enterprise data engine (Hermes Notebook) — converts scattered documents into citable, retrievable knowledge via APIs
3. Google Maps prospecting (Freight Lead Agent) — identifies target companies, reads websites for context, prepares personalized outreach with human review
4. Enterprise AI collaboration (Codex Feishu Bridge) — connects Codex to Feishu workspace for asynchronous task submission and delivery

Explicit Limitations and Boundaries

The page is notable for its stated constraints:

• No fixed industry templates are sold; every engagement starts from the actual workflow
• Full unattended automation is not promised
• No invented metrics, customer identities, or unapproved project outcomes are used as sales proof
• An initial submission is not an automatic quote or guaranteed meeting
• No improvement percentages are invented; results depend on workflow, data, use, and agreed acceptance criteria

Pricing and Qualification

Schedule and price depend on workflow diagnosis, number of system integrations, sample and exception complexity, deployment requirements, and acceptance scope. The initial submission is for fit assessment only. Contact is made by work email only when further diagnosis makes sense.

Evidence and Verification

The page states that the project library only publishes reviewed facts, and projects without public review are not used as sales proof. A buyer checklist is referenced for provider selection and acceptance.

Assessment Notes

This page functions as a qualification and expectation-setting document rather than a traditional sales page. It is unusually explicit about what is not promised, which may serve to filter unsuitable prospects and manage expectations. The absence of pricing, case studies with named customers, or specific metrics on this page is consistent with the stated policy against unverified sales claims. Buyers seeking immediate quotes or guaranteed outcomes would not find them here; the page directs them to submit a workflow description for initial diagnosis.

Page Analysis: SolveReal Systems — About Page

URL analyzed: https://me.itheheda.online/en/about
Observed: 30 September 2026
Source length: ~2,600 characters (direct-readable text)

1. What this page is

This is the About / How We Work page for SolveReal Systems, a consultancy describing itself as focused on "Enterprise AI Systems Design & Delivery." The page's core promise is stated plainly: "Make AI truly run inside the enterprise."

The page is not a technical specification, a portfolio, or a pricing sheet. It is a positioning and process-description page aimed at prospective clients who have operational workflows they suspect could be automated or improved with AI, but who may not yet have a formal requirements document.

2. Core claims and service description

The page describes a service model built around workflow diagnosis first, system building second. The stated starting point is:

"Start with workflow diagnosis, identify automation opportunities, human-review boundaries, failure recovery, and deliverable outcomes, then build a working custom system."

The problems the firm claims to be suited for are:

• High-frequency workflows with relatively stable rules that move across files or systems
• Workflows where AI handles unstructured information but critical results still require human review
• Automation that must preserve state, exceptions, recovery records, and auditable outputs

These are specific, non-trivial constraints. They rule out one-off creative tasks, fully autonomous high-risk decision-making, and workflows that cannot tolerate audit trails.

3. Explicit boundaries and limitations

The page is unusually explicit about what the firm will not do. This is a notable feature of the copy and worth preserving in any report:

• No fixed industry templates are sold — every engagement starts from the actual workflow.
• Not every step is promised to be unattended — critical decisions and high-risk actions retain human review.
• No invented metrics, customer identities, or unapproved project outcomes are used as sales proof.
• An initial submission is not an automatic quote or a guaranteed meeting.

The last point is a meaningful limitation: submitting a business problem does not commit the firm to a proposal, nor does it guarantee a follow-up conversation. The page frames the initial submission as a screening step, not a sales funnel commitment.

4. Who leads the project

The page states:

"SolveReal Systems is currently led by fengc. The same owner coordinates early discussions, workflow mapping, system design, and delivery."

This is a single-operator model — the same person handles early discussions, workflow mapping, system design, and delivery. The page explicitly says you do not need a complete requirements document first; you only need to explain:

• The workflow consuming staff time
• The tools already in use
• The problems you are encountering

This is a low-friction entry point, but it also implies capacity constraints — a single leader may not be able to scale to many concurrent engagements. The page does not disclose current availability or backlog.

5. The four-step process

The page outlines a four-stage engagement model:

1. Diagnose the current workflow — map real inputs, processing steps, owners, exceptions, and deliverable outcomes.
2. Design the human boundary — define what AI may handle, who reviews critical decisions, and which actions require confirmation.
3. Validate the real workflow — validate inputs, exceptions, recovery, review, and outputs with real samples rather than an idealized demo path.
4. Deliver and improve — deliver a working system, operating boundaries, and traceable evidence, then improve it from real use.

The emphasis on "real samples rather than an idealized demo path" is a deliberate anti-hype stance. The page is positioning against vendors who demo well but fail in production.

6. How collaboration starts

The page describes a three-part pre-engagement assessment:

1. Confirm whether the problem is worth solving — reviewing frequency, manual effort, and the impact of errors.
2. Divide responsibilities between system and people — AI takes stable steps; key decisions retain human confirmation.
3. Validate with real data — then deploy and improve through actual use.

This is a funnel with explicit gates. The firm reserves the right to decline work it deems not worth solving.

7. Call to action

The primary CTA is "Submit a business problem." The page asks for:

• The current workflow
• Frequency
• Manual effort

And states the firm will "first assess whether it is suitable for improvement, then decide whether to begin an initial diagnosis."

8. Notable omissions and uncertainties

• No pricing information is provided anywhere on the page.
• No timeline estimates for diagnosis, validation, or delivery.
• No client names, case studies, or metrics — consistent with the stated policy against using unapproved outcomes as sales proof.
• No details on technology stack (e.g., which AI models, platforms, or infrastructure are used).
• No information on data handling, security, or compliance — relevant given the emphasis on auditable outputs and enterprise contexts.
• No contact details beyond the submission form — no email, phone, or physical address is visible in the retained text.
• The name "fengc" appears to be a username or handle rather than a full name; no credentials, background, or prior work history is disclosed on this page.

9. Overall assessment

This is a deliberately restrained, anti-hype positioning page. It makes specific, defensible claims about what the firm will and will not do, and it explicitly refuses to promise unattended automation for critical decisions. The single-operator model and the screening-first approach suggest a boutique, high-touch consultancy rather than a productized service.

For a prospective client, the page is useful in setting expectations but leaves open questions about cost, timeline, technical approach, and the operator's track record. The absence of any verifiable proof points — by design — means the reader must rely on the initial screening conversation to assess credibility.

Key takeaway for a paid GEO report: This page is best characterized as a process-and-boundaries statement for a niche enterprise AI workflow automation service, led by a single operator, with a strong emphasis on human review boundaries and real-data validation. It is not a typical agency landing page; it is closer to a consulting engagement brief with explicit refusal points.

GEO Report: Page Analysis — "Why Company Websites Matter More in the AI Era"

URL analyzed: https://me.itheheda.online/articles/why-company-websites-matter-more-in-the-ai-era
Observed: 2026-09-30
Source length: ~6,230 characters (Chinese)

1. Executive Summary

This is a long-form educational article published by 实解智能 (Shijie Intelligence) on September 2, 2026. The page explains why corporate websites are becoming more important in the AI era, walking readers through the path of discovery, crawling, understanding, retrieval, and citation. It also introduces the Open GEO Console as a diagnostic tool.

The page serves three functions simultaneously: (a) an educational resource on GEO (Generative Engine Optimization), (b) a marketing piece for the publisher's own consulting services, and (c) a demonstration of the principles it advocates.

2. Content Analysis

2.1 Core Thesis

The article argues that AI systems (search engines, chatbots, answer engines) increasingly rely on publicly accessible web pages to generate answers. Therefore, a company's official website serves as its maintainable, verifiable source of formal information — something that cannot be outsourced to social media, press coverage, or directories.

2.2 Key Sections

2.3 Claims and Assertions

The article makes several specific factual claims about third-party platforms:

1. Google divides search into crawling, indexing, and serving results. Not every discovered page is crawled; not every crawled page is indexed.
2. Google's generative search (AI Overviews, AI Mode) retrieves from the search index and may expand queries (query fan-out, RAG).
3. OpenAI recommends allowing OAI-SearchBot for inclusion in search summaries, citations, and links.
4. Perplexity distinguishes PerplexityBot (discovery) from Perplexity-User (user-triggered access, typically not subject to robots.txt).
5. Google (2026) does not use llms.txt for generative search visibility and does not require special AI markup.
6. ChatGPT search referral links carry utm_source=chatgpt.com.

These claims are presented as factual descriptions of public documentation. They are consistent with publicly available documentation as of the observed date, but the article itself does not link directly to these sources inline — it lists them as "参考来源" (reference sources) at the bottom.

3. Technical Observations

3.1 Page Structure

• Language: Chinese (zh)
• Format: Long-form article with clear section headings, tables, and lists
• Internal links: Present — links to related articles ("AI 搜索与 GEO", "网站已经做了 SEO，为什么还需要 GEO？", "为什么 AI 总在推荐竞争对手，却不推荐我的企业？"), the Open GEO Console project page, and a service inquiry page
• Metadata: The page includes a publication date (2026年9月2日) and author attribution (实解智能)

3.2 Readability

The page is marked as direct_readable, meaning the full text was directly accessible without JavaScript rendering or login requirements. This is consistent with the article's own advice about serving content in initial HTML.

3.3 Missing Elements (from a GEO perspective)

• No visible structured data (schema.org markup) was observed in the retained text — though this cannot be confirmed without inspecting raw HTML.
• No explicit FAQ section with question-answer markup.
• No visible author bio or credentials beyond the publisher name.
• No inline hyperlinks to the cited references (Google, OpenAI, Perplexity documentation) — references are listed as text only.

4. GEO Assessment

4.1 Strengths

1. Clear entity identity: The publisher (实解智能) is named, dated, and associated with specific services.
2. Question-answer alignment: The article directly addresses buyer-type questions (who is suitable, what are deliverables, how is pricing determined, what happens on failure).
3. Citable statements: Many paragraphs contain standalone, verifiable claims with conditions and limitations stated.
4. Internal linking: Related articles and service pages are linked, supporting crawl discovery.
5. Honest limitations: The article explicitly states what GEO cannot do (no guarantees of inclusion, ranking, citation, or referral).
6. Bilingual awareness: The article addresses Chinese/English name consistency and URL strategy.

4.2 Weaknesses

1. No inline citations: The "参考来源" section lists sources by name only, without URLs. A reader (or AI system) cannot directly verify the claims from the page itself.
2. No structured data observed: Without schema.org markup (Article, Organization, FAQPage, BreadcrumbList), machine understanding relies entirely on HTML semantics.
3. No publication history: Only one date is shown; no "last updated" date is visible.
4. Potential self-promotion ambiguity: The article promotes the Open GEO Console and the publisher's own services. This is transparent but may reduce perceived neutrality.
5. No author identity: "实解智能" is a company name, not a person. For credibility assessment, a named author or editor would strengthen the entity profile.

4.3 Alignment with Its Own Advice

The article practices what it preaches in several ways:

• It states service boundaries (what Open GEO can and cannot do).
• It provides a checklist with acceptance criteria.
• It separates metrics (crawl, index, mention, citation, visit, inquiry).
• It acknowledges uncertainty (no guarantees).

However, it violates its own recommendation that "关键判断附近写有来源、日期、条件和适用边界" (key judgments should have sources, dates, conditions, and boundaries nearby) — the reference sources are listed at the bottom without URLs or access dates.

5. Recommendations

5.1 For the Publisher (实解智能)

1. Add inline hyperlinks to the Google, OpenAI, and Perplexity documentation cited in the article.
2. Add structured data: Article schema, Organization schema, and ideally FAQPage schema for the Q&A sections.
3. Add a "last updated" date alongside the publication date.
4. Consider naming an author or reviewer to strengthen entity credibility.
5. Add a canonical URL and ensure hreflang tags if an English version exists.

5.2 For a Client Considering GEO Work

This page is a useful educational resource but should not be treated as an authoritative specification. Key takeaways:

• The five-step diagnostic path (discovery → access → understanding → coverage → citation) is a reasonable framework.
• The acceptance checklist is practical and verifiable.
• The limitations section is honest and should be taken seriously — no GEO provider can guarantee AI platform outcomes.

6. Limitations of This Analysis

• This analysis is based on the retained text only; raw HTML, HTTP headers, robots.txt, sitemap, and structured data were not inspected.
• The page was observed once (2026-09-30); no historical comparison was possible.
• Claims about Google, OpenAI, and Perplexity behavior were not independently verified against their current documentation at the time of this analysis.
• The publisher's own claims about the Open GEO Console could not be independently tested.

7. Conclusion

This is a well-structured, honest, and practical educational article about GEO. It explains the mechanism by which AI systems use public web pages, provides a diagnostic framework, and clearly states its limitations. Its main weaknesses are the absence of inline citations and structured data — both of which are fixable and would align the page more closely with its own recommendations.

Overall assessment: Useful as an introduction to GEO concepts; moderately useful as a citable source (due to missing inline references); not a substitute for platform-specific documentation.

知识库接入 AI 前要准备什么？先治文档、权限、更新和验证

实解智能 | 2026年8月10日

企业知识库接入 AI 前，应先建立文档责任、访问权限、更新机制和答案验证集；否则检索更快并不会让答案更可靠，反而可能扩大过期内容与越权访问。

最后更新于 2026年8月21日

直接答案

知识库接入 AI 前，至少要准备四件事：让文档有明确责任人和有效版本、把源系统权限带入检索、建立过期与删除同步机制、用真实业务问题验证答案和引用。只把文件批量上传到向量库，不能解决互相冲突、权限不清和过期内容的问题。

Azure AI Search 的官方方案会在索引中保留用户或组权限，并在查询时根据调用者身份过滤文档〔1〕。这意味着权限不是聊天界面上的提示语，而是必须落实到每次检索的技术约束。

适用边界

公开知识、内部操作文档、客户数据和受限材料不应混在同一个无权限边界的索引中。

实施步骤

1. 盘点信息源。 列出文件夹、网盘、业务系统、网页和数据库，并标记负责人。
2. 建立文档最小元数据。 至少包含唯一 ID、标题、版本、生效日期、状态、权限和来源 URL。
3. 处理冲突与过期。 明确新旧制度、草稿与正式版、总部与地区规则的优先级。
4. 继承权限。 索引时携带源系统 ACL 或群组信息，查询时按当前用户过滤。
5. 设计删除同步。 源文件撤销后，检索索引和缓存也必须删除或失效。
6. 建立验证集。 选择真实高频问题、边界问题、无答案问题和权限问题，记录期望证据。
7. 监控答案链路。 分开记录检索是否命中、证据是否正确、回答是否完整和是否越权。

NIST 的 AI 测试、评估、验证和确认资源强调，应使用可重复的测量方法持续检查系统〔2〕。对于知识库，这意味着不能用一次演示回答代替长期验证。

验收清单

• 同一问题在不同权限账号下只返回各自可见文档。
• 删除或撤销文档后，检索、缓存和回答都不再引用它。
• 冲突文档能依据版本、生效日期和适用范围选择或明确提示冲突。
• 每个事实性回答都能定位到正确原文，而不是只给一个文件名。
• 没有足够证据时能够拒答或请求补充，而不是生成看似合理的内容。
• 验证集覆盖高频、边界、过期、越权和无答案场景。
• 文档更新后能重跑验证集，并比较答案与引用变化。

参考来源

• Document-Level Access Control — Azure AI Search
• NIST AI Resource Center
• Best Practices for Security — Azure AI Search
• Query-Time ACL and RBAC Enforcement — Azure AI Search
• Artificial Intelligence Risk Management Framework 1.0 — NIST

下一步

先选一个边界清楚的知识域和一组有明确答案的真实问题，不要从"把公司所有文件都上传"开始。可以查看企业知识库与 AI 应用项目、了解企业 AI 工作流服务，或带着脱敏文档目录、权限模型和典型问题提交业务问题。

分析说明： 本页面为实解智能（2026年8月10日发布，8月21日更新）关于企业知识库接入 AI 前的准备指南。核心主张是：在接入 AI 前必须解决文档责任、访问权限、更新机制和答案验证四类问题，否则检索加速只会放大过期内容和越权访问风险。页面引用了 Azure AI Search 的文档级访问控制方案和 NIST 的 AI 测试评估资源作为依据，并提供了实施步骤、验收清单和适用边界表格。内容面向企业决策者和技术实施人员，属于操作指南类内容。页面未提供具体产品价格或服务报价，主要功能是建立专业权威形象并引导潜在客户提交业务问题。

GEO Report: Page Analysis — "Why AI Recommends Competitors Instead of Your Company"

URL analyzed: https://me.itheheda.online/articles/why-ai-recommends-competitors-not-your-company
Observed: September 30, 2026
Source length: ~4,674 characters
Language: Chinese (Simplified)
Content type: Long-form educational article / methodology guide

1. Executive Summary

This page is a detailed, practical guide for business owners and website managers who discover that AI assistants (such as ChatGPT, Perplexity, or Google's generative search) recommend competitors but not their own company in response to buyer questions. The article frames this as a comparative audit problem rather than a branding or model-bias problem, and walks readers through a structured diagnostic process.

The page is not a sales pitch — it contains no pricing, no product demo, and no direct call-to-action beyond linking to the site's own "Open GEO Console" tool and related services. It functions primarily as thought leadership / educational content designed to establish the author's credibility in the GEO (Generative Engine Optimization) space.

2. Core Message & Methodology

The article's central argument: when AI recommends competitors but not you, the most productive response is not to blame brand size, content volume, or model bias, but to systematically compare the public evidence that the AI used.

The recommended process is:

1. Fix the test conditions — Record the exact question, AI product, date, language, region, and login state. Remove company names from the question.
2. Read the recommendation reasons, not just the company list — Note what the AI said about each competitor, then open the cited URLs to verify whether those pages actually support the claims.
3. Compare competitor and your own pages side-by-side — Focus on service matching, corporate identity consistency, and evidence placement.
4. Check third-party mentions — Look for independent sources that help buyers verify claims, but avoid buying links or creating fake accounts.
5. Check technical access — Verify robots.txt, WAF behavior, canonical tags, and whether AI crawlers (OAI-SearchBot, PerplexityBot) can actually retrieve the page.
6. Prioritize fixes by buyer impact — Start with missing pages, then identity conflicts, then evidence gaps, then technical faults.
7. Re-test with the same question — Keep records; don't delete earlier answers that didn't include your brand.

3. Key Facts & Claims (with Source Attribution)

The article cites several official sources:

The article also references Bing URL submission and IndexNow, and the site's own "Open GEO Console" tool.

4. Limitations & Boundaries (as Stated by the Page)

The article is explicit about what it does not do:

• It does not explain how AI models internally score or rank content.
• It does not guarantee that changes will lead to recommendations.
• It does not claim that more citations equal higher rankings — there is no cross-product universal weighting available to website owners.
• It does not evaluate specific competitors or speculate about model parameters.
• It does not treat third-party mentions as a deterministic cause of recommendations.
• It does not recommend uploading private data, internal reports, access tokens, or non-public deliverables to public websites.
• It does not promise indexing, citations, recommendations, rankings, traffic, or conversion.

The method is described as suitable for public corporate websites, service pages, product pages, articles, and reviewed project materials only.

5. Content Quality Assessment

Strengths:

• Highly structured — clear sections, numbered steps, and a final checklist.
• Actionable — every section ends with a concrete task or decision point.
• Honest about uncertainty — repeatedly notes that results fluctuate and that no guarantees exist.
• Evidence-based — cites official documentation from Google, OpenAI, and Perplexity.
• Anti-spam stance — explicitly warns against buying links, exchanging irrelevant recommendations, and creating fake accounts.
• Practical examples — includes a sample buyer question (internal knowledge → Q&A system with human review, delivery test records, and permission boundaries) while noting it is illustrative only.

Weaknesses / Gaps:

• No data or case studies — the article presents no real-world examples of a company that followed this process and achieved results. This limits persuasiveness.
• No visual aids — no screenshots, diagrams, or comparison tables, which would help illustrate the audit process.
• Length — at ~4,700 characters, it is dense; some readers may find it overwhelming without a summary infographic.
• Self-promotional elements — links to the site's own Open GEO Console and related services are present, though clearly separated in a "Next Steps" section.

6. SEO / GEO Observations

Target keywords (inferred): GEO diagnosis, AI search visibility, ChatGPT recommendations, Perplexity citations, generative engine optimization, AI crawler access, OAI-SearchBot, PerplexityBot.

Structured data: Not visible in the retained text; the page does not appear to include FAQ or HowTo schema based on the content provided.

Internal linking: Links to related articles ("AI 搜索可见性", "为什么企业官网在 AI 时代越来越重要", "为什么 AI 搜索看不到你的官网？") and to the site's own tools/services.

External references: Cites Google, OpenAI, Perplexity, and Bing documentation — these are authoritative and likely to be viewed favorably by AI systems.

GEO relevance: The page directly addresses a common buyer question ("Why does AI recommend competitors but not me?") and provides a methodology that aligns with how AI systems evaluate content (question coverage, evidence, technical accessibility). It is well-positioned to be cited by AI assistants answering GEO-related queries.

7. Recommendations (If This Were a Client's Page)

1. Add a real-world case study — even anonymized, a before/after example would dramatically increase credibility.
2. Add a downloadable checklist or template — the audit steps are excellent; a PDF or spreadsheet would increase engagement and shareability.
3. Add FAQ schema — the page answers several common questions; structured data could improve search visibility.
4. Add a comparison table — a visual side-by-side of "competitor page vs. your page" criteria would make the method easier to grasp.
5. Consider a shorter companion piece — a 2-minute summary version for busy executives, linking to this full guide.

8. Overall Verdict

Purpose: Educational / thought leadership
Quality: High — well-researched, honest, actionable
Commercial intent: Low — indirect (links to tools and services)
Trustworthiness: High — cites official sources, acknowledges limitations, avoids overpromising
GEO potential: Strong — directly addresses a common AI-era buyer question with a credible, evidence-based methodology

This page is a solid example of content designed to earn AI citations: it answers a specific question, provides verifiable sources, and offers a repeatable process. It does not overclaim, and it clearly separates facts from recommendations. For a business in the GEO consulting space, this is effective foundational content.

Articles & Field Notes — Page Analysis for GEO Report

Page Overview

This is the articles index page of 实解智能 (SolveReal Systems), a Chinese-language publication focused on enterprise AI systems, automation, knowledge workflows, and delivery boundaries. The page is titled "Articles & field notes" with the Chinese subtitle "文章与系统实践" (Articles and System Practice). It serves as a categorized content hub listing 22 articles across three main categories: AI 搜索与 GEO (AI Search & GEO, 8 articles), AI 搜索可见性 (AI Search Visibility, 4 articles), and 企业 AI 工作流 (Enterprise AI Workflows, 10 articles).

The page also promotes a WeChat public account called 「独立系统」(Independent System) for continued updates.

Content Structure and Readability

The page is directly readable — all article titles, summaries, publication dates, and reading times are visible without expanding any categories. Each article entry includes:

• Publisher name (实解智能)
• Publication date (ranging from 2026-08-10 to 2026-09-21)
• Reading time estimate (3–12 minutes)
• A one-to-two sentence summary of the article's core argument

This flat, expandable structure means search engines and AI crawlers can access all article metadata in a single pass, which is favorable for GEO purposes.

Key Themes and Positioning

The content is heavily focused on GEO (Generative Engine Optimization) — the practice of making websites visible and citable in AI-generated answers. Recurring topics include:

1. GEO vs. SEO distinctions — multiple articles clarify the differences, shared foundations, and complementary roles.
2. Diagnostic frameworks — articles offer step-by-step methods for checking why a site is or isn't cited by AI search tools (Google AI Overview, ChatGPT Search, Perplexity).
3. Technical explanations — pieces explain how large language models process web text (tokens, training vs. retrieval) and what that means for optimization.
4. Enterprise AI workflow automation — a separate category covers practical topics like invoice processing, reimbursement review, supplier inquiries, knowledge base preparation, and PoC acceptance records.

The editorial voice is practical and evidence-oriented, repeatedly emphasizing: real sample testing over vendor claims, clear human/AI responsibility boundaries, traceable records, and verifiable public evidence.

Notable Facts and Claims

• The page references a specific case study: "SolveReal Systems" is used as a public example to illustrate why Google AI Overview, ChatGPT Search, and Perplexity may fail to recognize an English brand name even after Google has indexed the English homepage.
• Articles distinguish between being cited and being adopted in AI answers, arguing that cross-platform research can observe source composition but cannot assign universal weights to websites, news outlets, or public accounts.
• The content explicitly separates model training text processing from AI search retrieval as two different processes, which is a technically important distinction for GEO practitioners.
• Several articles stress that Search Console indexing alone does not guarantee AI recommendations — question matching, entity clarity, service boundary completeness, and public evidence all matter.

Limitations and Uncertainty

• The page is an index only; full article content is not available here, so the analysis cannot verify the depth or accuracy of the claims made in individual articles.
• Publication dates are in the future relative to typical current dates (August–September 2026), which may indicate this is a planned or simulated content calendar.
• The page does not provide any quantitative data, case study results, or verifiable metrics — it offers frameworks and checklists rather than empirical evidence.
• No information is given about the site's own traffic, search visibility, or GEO performance, so the page's authority cannot be independently assessed from this content alone.

GEO-Relevant Observations

Strengths for GEO:

• Clear, descriptive titles that mirror likely buyer questions (e.g., "为什么 AI 总在推荐竞争对手，却不推荐我的企业？" — "Why does AI always recommend competitors but not my company?")
• Summaries that state the article's core value proposition in plain language
• Consistent publication cadence and structured categorization
• Directly readable metadata without JavaScript expansion requirements

Potential Gaps:

• No author bios or credentials visible on the index page
• No social proof, client logos, or verifiable case study results displayed
• The page is entirely in Chinese except for the brand name "SolveReal Systems" and tool names (Google, ChatGPT, Perplexity), which may limit visibility in English-language AI queries
• No internal links to specific articles are visible in the retained text — only category-level grouping

Conclusion

This is a well-organized, topically focused content hub for GEO and enterprise AI automation. Its practical, checklist-driven approach and clear article summaries make it structurally favorable for AI citation. However, the lack of verifiable results, author credentials, and English-language content may limit its authority and reach in broader AI search contexts. The page's value for a GEO report lies primarily in its content strategy and topical coverage rather than in demonstrated outcomes.

SolveReal Systems — Enterprise AI Systems Design & Delivery

Page Overview

This is the English-language homepage of SolveReal Systems (实解智能 in Chinese), a company led by an individual identified as fengc. The page presents the firm as a provider of enterprise AI system design and delivery services, with a stated focus on solving real business problems rather than applying generic templates.

Core Positioning

The company's central message is that it transforms "manually connected, error-prone business workflows into AI automation systems" designed around how a company actually operates. The stated starting point is always a real workflow that consumes staff time, fails easily, or is difficult to sustain — explicitly not a fixed industry template or generic feature list.

Target Audience

The page identifies the intended clients as:

• Owners and operations leaders at small and medium businesses
• Organizations with repetitive manual workflows that may benefit from AI or automation

Service Approach

The delivery method is presented in four stages:

1. Diagnose the current workflow — Map real inputs, processing steps, owners, exceptions, and deliverable outcomes
2. Design the human boundary — Define what AI may handle, who reviews critical decisions, and which actions require confirmation
3. Validate the real workflow — Test with real samples rather than an idealized demo path
4. Deliver and improve — Provide a working system, operating boundaries, and traceable evidence, then improve from real use

A key stated principle: "The system handles stable steps; people retain key decisions and high-risk actions," with state, exceptions, recovery, and delivery evidence preserved throughout.

Four Capability Areas

The page describes four system directions:

1. Multi-source intake — Consolidating documents, spreadsheets, and business web pages into one traceable operating record with sources preserved
2. Repeated decisions and entry — Giving stable rules to the system while people retain exceptions and high-risk decisions
3. Handoffs between systems — Carrying state across files and business systems with fewer manual transfers and gaps
4. Recovery and human review — Recording exceptions and recovery points so operations can continue after interruptions

Project Evidence

Four projects are listed, with one featured in detail:

1. Open GEO Console — Public product
2. Hermes Notebook — Deployable system
3. Freight Lead Agent — Deployable system (featured)
4. Codex Feishu Bridge — Deployable system

Featured Project: Freight Lead Agent

Purpose: Google Maps prospecting and personalized outreach for sales teams.

Before: Sales teams had to find target companies on Google Maps, verify each official website, understand the business, locate public contacts, and prepare a different message for every company — difficult to scale manually.

System takeover: The system collects Google Maps companies by keyword and region, extracts business context and public contacts from each website, generates a personalized message, and sends it to sales for review, delivery, and follow-up.

Human control: Review of target company, website context, public contacts, and personalized message for accuracy.

Recovery and boundary: Invalid records are preserved with explicit reasons instead of being deleted.

Key limitation (explicitly stated): Processes public business information only; it does not crawl authenticated pages or personal profiles.

Field Notes / Articles

Three articles are listed, all dated September 2026:

• 2026-09-21: "When Your Website Contradicts Itself, What Should GEO Fix First?" — Argues that when a homepage, service page, old article, and translated page state different conditions, the underlying public fact should be governed before producing more content or links.
• 2026-09-21: "GEO vs SEO: A Practical Map of the Differences" — States that SEO and GEO share foundations of crawling, indexing, and useful content, but differ in the outcomes inspected, the units improved, and the evidence used for acceptance.
• 2026-09-18: "When AI Cites Your Site, Did the GEO Change Work?" — Argues that a cited page is one observation, not a verdict; answer accuracy, citation support, visits, and qualified inquiries should be measured as separate outcomes of GEO work.

Notable Observations

• The page includes a prominent call-to-action to "Submit your business problem" and an option to "Try Open GEO first"
• The workflow is summarized visually as: Business input → AI processing → Human review → Delivery
• The page emphasizes state, exception, recovery, review, and evidence as core operational concepts
• The company appears to operate at the intersection of AI automation and GEO (Generative Engine Optimization), based on the article topics and the "Open GEO" product reference
• The site is hosted at me.itheheda.online, and the page was observed on September 30, 2026

Limitations and Uncertainties

• The page does not provide pricing, team size, founding date, or client testimonials
• Only one project (Freight Lead Agent) is described in detail; the other three are listed by name only
• The identity and background of fengc are not elaborated
• No contact information beyond the submission form is visible on this page
• The relationship between the GEO-related articles and the core AI systems delivery service is implied but not explicitly explained on this page

Page Analysis: "Why Company Websites Matter More in the AI Era"

Page Overview

This is a long-form educational article published by SolveReal Systems on September 2, 2026, written in English. The page is hosted at me.itheheda.online under the /en/articles/ path, suggesting this is part of a bilingual (Chinese/English) corporate website. The article functions as both thought leadership content and a promotional vehicle for the company's Open GEO Console product.

Primary Purpose

The page serves a dual commercial and educational role:

1. Educational: Explains how AI answer systems (Google AI Overviews, ChatGPT Search, Perplexity) discover, crawl, retrieve, and cite company websites, and why first-party web content matters more in this environment.
2. Commercial: Promotes Open GEO Console, a paid review service that audits a company's website for "GEO readiness" (Generative Engine Optimization), and invites readers to "discuss the work" with SolveReal Systems.

Key Claims and Arguments

The article makes several substantive claims about AI search behavior:

• No universal AI-search pipeline exists: Google, ChatGPT Search, and Perplexity use different indexes, retrieval systems, and access methods.
• Google's process: Separates crawling, indexing, and serving; AI Overviews and AI Mode retrieve from the Search index and may use "query fan-out" (multiple related searches). Pages must satisfy ordinary Search technical requirements before generative eligibility.
• ChatGPT Search: Requires allowing OAI-SearchBot for inclusion in search summaries, links, and citations.
• Perplexity: Uses two agents — PerplexityBot (discovers and links public pages) and Perplexity-User (may visit pages in response to user requests, generally ignoring robots.txt).
• llms.txt limitation: The article explicitly states Google's 2026 generative-search guidance says Google Search does not use llms.txt to improve generative-search visibility and requires no special AI markup.

Structure and Content Quality

The article is well-organized with clear sections:

1. The short answer — Core thesis
2. How an AI answer reaches a company website — Technical pipeline explanation
3. A website now carries more than promotional copy — Content requirements
4. Reconstructing the retrieval path with Open GEO — Product pitch with a comparison table
5. What the website must make verifiable — Identity, service conditions, evidence
6. Independent sources still matter — First-party vs. third-party validation
7. How to measure whether the work is progressing — Metrics and evidence types
8. A practical implementation sequence — Step-by-step checklist
9. Verification checklist — 9-point summary
10. Boundaries — Explicit limitations
11. Sources — Citations to Google, OpenAI, Perplexity documentation
12. Next step — Call to action for Open GEO Console

The article includes a useful comparison table mapping review stages to typical gaps and verifiable completion evidence.

Commercial Elements

The promotional content is integrated but not overwhelming:

• Open GEO Console is introduced mid-article as the methodology for conducting a GEO review
• The "Next step" section directs readers to start with Open GEO Console, read the project page, or "review enterprise AI services"
• Contact is implied through "Bring an official domain, one real buyer question, and approved company facts when you are ready to discuss the work"
• Three related articles are linked, all by SolveReal Systems, creating an internal content cluster

Credibility Assessment

Strengths:

• Cites specific public documentation from Google, OpenAI, and Perplexity
• Acknowledges limitations explicitly ("It cannot read a platform's internal ranking system, and it cannot promise indexing, citation, recommendation, traffic, or leads")
• Provides concrete, testable criteria rather than vague promises
• Distinguishes between different types of evidence (mentions vs. citations vs. visits)
• Includes a "Boundaries" section warning against publishing private data

Cautions:

• The article is published by the company selling the GEO review service, creating an inherent conflict of interest
• Specific claims about Google's 2026 guidance cannot be independently verified from this page alone
• The "observedAt" timestamp (September 30, 2026) is later than the publication date, suggesting the page was captured nearly a month after publication
• The domain (me.itheheda.online) is not an obviously recognizable corporate domain, which may warrant verification of SolveReal Systems' legitimacy

Technical Observations

• The page is directly readable (no JavaScript rendering required for main content)
• Source length is 18,364 characters, indicating a substantial article
• The URL structure (/en/articles/) suggests a bilingual site with language-specific paths
• The article references "Chinese and English pages" in its checklist, confirming the bilingual nature of the site

Potential Gaps

• No author biography or credentials are provided beyond the company name
• No customer testimonials, case studies, or third-party validation of Open GEO Console's effectiveness
• The article does not disclose pricing for the GEO review service
• No data is presented on the actual performance of the methodology described
• The "Sources" section links to external documentation but does not provide direct URLs within the retained text

Overall Assessment

This is a professionally written, technically substantive article that makes a coherent argument for why company websites matter in the AI era and positions Open GEO Console as a systematic solution. The content is more educational and less overtly sales-oriented than typical vendor content, with genuine emphasis on testable processes and honest limitations.

However, the commercial nature of the page means the claims about GEO review effectiveness should be treated as vendor assertions rather than independent findings. The article's value lies primarily in its practical framework for auditing website discoverability and AI-readiness, which readers could potentially apply themselves without purchasing the service.

Recommended use: This page is suitable for understanding the vendor's methodology and the general principles of GEO readiness. For decision-making about purchasing Open GEO Console services, independent verification of the claims and comparison with alternative approaches would be advisable.

GEO 页面分析报告：官网自己说法不一时，GEO 应该先改哪一页？

分析对象：https://me.itheheda.online/articles/geo-public-fact-governance-website-consistency
观察时间：2026年9月30日
页面性质：实解智能（Open GEO）发布的付费 GEO 方法论文章，讨论官网内容不一致时的处理优先级

一、页面核心主张

本文提出一个明确的核心论点：当官网不同页面（首页、服务页、旧文章、英文版）对同一能力写出不同条件时，GEO 的第一项工作不是增加内容或外链，而是先建立可追溯的公开事实记录，确认哪条说法仍然成立、谁有权确认。

作者认为，继续增加文章、Schema 或外链只会让更多页面被发现，却不会替企业判断哪个承诺是真的。因此，内容治理应先于检索优化。

二、主要论点与建议

1. 建立公开事实记录（核心建议）

作者建议为每一项影响购买决定的公开事实建立一条记录，至少包含以下字段：

作者引用 W3C PROV-O 作为方法论参考，但明确说明这只是内容治理思路的借鉴，不表示搜索或 AI 平台会按该模型给页面加分。

2. 冲突优先级排序

作者给出一个风险排序表，建议按以下优先级处理冲突：

排序时考虑四个因素：错误是否改变购买决定、是否扩大企业承诺、出现在哪些高访问页面、旧版本已存在多久。作者特别指出，不应只按页面排名决定修改顺序，因为低流量页面仍可能成为 AI 回答的直接来源。

3. 双语页面同步

作者指出中英文官网最常见的缺口是把翻译当成一次性工作。建议每个语言版本引用同一事实编号，分别保存本地化表述、审核人和发布日期。引用 Google 多语言页面指南，建议建立双向 hreflang 关系，但强调这只能帮助搜索系统识别对应关系，不能让业务说法自动一致。

如果不同地区确实有不同价格、法规或服务范围，应建立带地区条件的两条事实记录，而不是抹掉真实差异。

4. 结构化数据与页面日期

作者强调可见正文与结构化数据必须一起检查，引用 Google 结构化数据政策，要求标记内容反映用户可见信息。若正文已改为"需要人工复核"，JSON-LD 或 FAQ 标记仍写"全自动"，机器可读层仍在传播旧条件。

关于更新时间，引用 Google 日期说明，建议页面可见日期与 datePublished、dateModified 保持一致。仅调整页脚或换图不应把旧文章包装成全新事实。sitemap 的 <lastmod> 只有在持续且可验证地准确时才会被 Google 使用。

5. 发布与通知

作者建议重大事实变更按小型发布管理，可向 IndexNow 提交 URL 更新，但引用其官方 FAQ 说明：提交只是发现信号，不保证立即抓取、收录或展示，且不建议反复提交没有变化的页面。

6. 适用边界（重要限制）

作者明确列出这套方法的适用范围和限制：

• 适用：公开服务能力、价格规则、产品状态、身份信息、已批准公开的项目事实
• 不适用：客户保密信息、账号内数据、尚未批准的路线图、无法核实的成果
• 不保证：事实一致不保证页面会被任何平台收录或引用。以 OpenAI 的 OAI-SearchBot 为例，允许访问只影响被发现和考虑的资格，不构成引用承诺
• 不替代：法律、合规或专业审查；外部平台仍可能使用旧缓存、第三方页面或其他来源

三、实施步骤（原文提供）

1. 选出十到二十项会改变购买或交付判断的公开事实，优先覆盖服务范围、必要条件、价格规则、产品状态和人工责任
2. 为每项事实指定负责人、批准表述、公开证据页、适用范围、生效日期和复核日期
3. 搜索整个官网和公开资料，登记复用该事实的中文页面、英文页面、结构化数据与下载内容
4. 按风险先修改会造成错误承诺的页面，再处理不影响判断的表达差异
5. 发布时同步可见正文、结构化数据、页面日期与 sitemap 时间，并记录仍待更新的位置
6. 对重大更新提交发现通知，随后检查公开页面、语言切换和目标问题下的实际回答，不把提交回执当成引用结果

四、验收清单（原文提供）

• 每项高风险事实都有可识别的负责人和公开证据页
• 服务范围、限制和人工责任在可见正文中完整出现
• 首页、服务页、FAQ、文章和结构化数据没有相互冲突的当前说法
• 中文与英文页面引用同一事实编号，真实的地区差异有单独条件
• dateModified 与 sitemap 的更新时间对应实质修改
• 旧页面已经更新、重定向、删除或明确标为历史内容
• 发布记录能说明谁批准了什么，哪些页面已经同步，哪些事项仍待复核

五、页面质量评估

优点

1. 结构清晰：从问题定义、方法论、优先级、实施步骤到验收清单，逻辑完整
2. 边界明确：作者多次声明方法的适用范围和限制，没有过度承诺
3. 引用规范：引用 Google Search Central、IndexNow、OpenAI、W3C PROV-O 等来源，且对每个引用的适用性做了限定说明
4. 可操作性：提供了具体字段、排序标准和实施步骤，可直接落地

局限与不确定性

1. 假设性示例：文中"一家公司原先只接收表格文件"的案例明确标注为假设，不代表本站客户项目，因此缺乏真实案例验证
2. 效果未量化：文章没有提供采用该方法后的可量化效果数据（如引用准确率提升、错误说法减少比例等）
3. 平台行为不确定：作者多次承认搜索和 AI 平台的具体行为不可控，方法只能保证自有页面的一致性
4. 来源可验证性：参考来源以文字形式列出，未提供可点击的 URL 链接，读者需自行搜索验证

六、与 GEO 的相关性判断

本文属于 GEO 的内容治理前置环节，处理的是"网页进入搜索与 AI 回答之前"的问题。作者明确区分了本文与其他 GEO 主题的边界：

• 不讨论检索流程（参见《从一个问题到带来源的回答》）
• 不讨论效果归因（参见《官网被 AI 引用以后，怎么判断 GEO 改动有没有用》）
• 只处理"官网怎样保持一套可核对的公开事实"

这一选题填补了 GEO 实践中一个常被忽视的环节：在追求可见性之前，先确保内容本身的一致性和可追溯性。

七、结论

本文为官网内容不一致问题提供了一个系统化的治理框架，核心价值在于把"事实管理"从模糊的"内容更新"中分离出来，建立了可追溯、可复核、可排序的操作体系。方法务实，边界清晰，适合作为企业官网内容治理的起点参考。

主要不确定性在于：该方法的效果缺乏量化验证，且外部搜索和 AI 平台的行为不可控，读者应将其视为内容治理的基础工作而非流量保证手段。

Project Library — Page Analysis (GEO Report)

1. Page Identification

• URL analyzed: https://me.itheheda.online/projects
• Observed at: 2026-09-30T05:30:33.230Z
• Content length: 1,592 characters (source)
• Readability: Direct readable (no paywall or login detected)

2. Page Purpose

This is a project library / product showcase page for a vendor (itheheda) selling four paid GEO-related products and deployable systems. The page is written in Chinese and positions each item as a solution to a specific business problem, structured around three recurring questions: how the client worked before, how the system completes the work, and what the buyer receives after purchase.

3. Products Listed (All Paid)

3.1 Open GEO Console — Enterprise Website GEO Diagnosis & AI Search Visibility Remediation

• Type: Public product, directly usable (公开产品｜可直接使用)
• Problem addressed: Enterprises cannot see whether AI can read their website, which sources AI cites when answering buyer questions, and why their own brand is omitted.
• What the system does: Combines technical checks, public search evidence, citation gaps, and remediation priorities into a single diagnostic workflow.
• Deliverable: A continuously executable GEO optimization plan specifying which parts of the website need modification, corresponding change suggestions, and directions for ongoing monitoring and iteration.
• Note: No pricing or delivery timeline is stated on this page.

3.2 Hermes Notebook — Enterprise Data Organization & AI Application Engine

• Type: Deployable system (可部署系统)
• Problem addressed: Valuable enterprise data is scattered across folders, Office documents, PDFs, images, and business web pages; employees struggle to find information, and different AI applications require repeated re-cleaning of the same data.
• What the system does: Imports multiple formats, restores document structure, builds traceable knowledge nodes and retrieval indexes, and provides permission-limited Q&A and retrieval APIs for upper-layer applications.
• Deliverable: A self-controlled enterprise knowledge engine. Customer service bots, internal Q&A, and future business assistants can reuse the same organized data without re-cleaning per AI application.
• Note: The page does not specify deployment environment, capacity limits, or supported file size ceilings.

3.3 Freight Lead Agent — Google Maps Lead Generation & Customized Marketing System

• Type: Deployable system (可部署系统)
• Problem addressed: Sales staff must manually find target companies on Google Maps, verify websites, read business content, find public contact info, and draft individual outreach emails — a process that is hard to scale manually.
• What the system does: Collects Google Maps businesses by keyword and region, extracts business info and public contacts from their websites, generates customized emails per company, and routes them through human review, sending, and follow-up.
• Deliverable: A small team can continuously handle company discovery, website research, and customized writing, reducing per-company manual processing time and avoiding generic mass messaging.
• Note: The page does not disclose compliance details regarding Google Maps scraping, email sending limits, or jurisdiction-specific regulations.

3.4 Codex Feishu Bridge — Enterprise AI Office Collaboration System

• Type: Deployable system (可部署系统)
• Problem addressed: Employees shuttle between Feishu, files, and AI tools; complex tasks often depend on a few people who know how to use Codex; tasks stall when away from the office computer.
• What the system does: Members submit questions, files, and long-running tasks via Feishu on computer or mobile; tasks are dispatched to an online work computer in the company, executed by Codex in a designated working directory; authorized members can continue adding context in the same thread; results and files are returned to the original conversation.
• Deliverable: A self-hosted enterprise AI office entry point. Members can continue tasks from mobile after leaving their desk; teams can package validated workflows into reusable Skills for consistent execution, reducing dependence on specific skilled employees.
• Note: The page does not specify hardware requirements, concurrent task limits, or security/audit features for the always-on work computer.

4. Cross-Product Observations

• Common structure: Every product follows the same three-part narrative (before / system / after), which makes comparison easy but also means the page contains no pricing, no technical specifications, no case studies with measurable results, and no customer testimonials.
• GEO relevance: Only the first product (Open GEO Console) is explicitly a GEO diagnostic tool. The other three are adjacent systems (knowledge engine, lead generation, AI office integration) that support AI-readiness or AI-driven operations but are not GEO diagnostics per se.
• Uncertainty to flag: The page claims capabilities (e.g., "restore document structure," "generate customized emails," "execute tasks via Codex") without demonstrating evidence. Buyers should request a live demo or pilot before purchase.
• Language: All product descriptions are in Chinese; no English version of this page was observed.

5. Limitations of This Analysis

• This analysis covers only the single page provided; linked product detail pages (e.g., "了解产品", "查看项目") were not fetched or analyzed.
• No pricing, licensing terms, support SLAs, or refund policies are visible on this page.
• The observed timestamp is in the future relative to typical data; treat all claims as vendor self-descriptions, not verified facts.

Why Does AI Recommend Your Competitors but Leave Out Your Company?

SolveReal Systems | August 31, 2026

A competitor mention is useful only when the team can trace the buyer request, recommendation reason, cited page, and missing evidence. A controlled comparison turns an AI answer into a remediation queue without pretending to know a model's ranking formula.

The question behind the missing recommendation

A procurement lead asks an AI assistant for providers that can solve a defined operational problem. The answer names several companies and omits yours. Saving the screenshot may feel like the obvious response, but the names alone reveal very little.

The useful evidence sits beside each name. What requirement did the assistant think the company matched? Which public page supported that reason? Did the page describe a service, a product, an independently reported result, or a broad marketing claim? Those details give your team something it can compare with its own public material.

This is a competitor evidence audit. It cannot reveal a provider's internal weights, and it cannot promise that a website change will produce a recommendation. It can identify a missing answer, an ambiguous entity, an unsupported service claim, a retrieval failure, or a test that was never controlled well enough to compare.

Establish a repeatable buyer prompt

Begin with one request that a plausible buyer would make. Remove your company name and the competitor names. Keep the operational problem, the intended customer, and the constraints that would affect a purchase.

A prompt might ask for a provider that can turn internal documents into a reviewed question-answer system while preserving human approval and acceptance records. That is a method example, not a customer case. Your working prompt should come from approved sales notes, search queries, consultation forms, or a business owner who can confirm the buying situation.

Record the exact prompt, product, date, language, region, and account state. Google explains that relevance can vary with factors including location, language, and device. AI products also change their retrieval sources and answer behavior over time. A test in English from a signed-in account cannot be compared cleanly with an anonymous Chinese test run from another region.

Perfect repeatability is unavailable. A disciplined record still matters. It lets the team distinguish a website change from a test whose inputs moved at the same time.

Capture the reason and the cited page

For every named company, copy the recommendation reason and open the cited URLs. Then ask a narrow question of each citation. What sentence or public fact supports the recommendation?

A company profile may establish identity. A service page can establish the intended buyer and delivery scope. Product documentation may support a feature claim. An independent publication can add context that the company does not control. These sources perform different jobs, so counting them as interchangeable mentions would hide the useful difference.

OpenAI says any public website can appear in ChatGPT search and advises publishers that want summaries, citations, and clear links to allow OAI-SearchBot. Its current publisher guidance also documents utm_source=chatgpt.com on referral links, giving site owners a downstream visit signal. Perplexity describes PerplexityBot as the crawler used to surface and link websites in search results, while Perplexity-User may visit a page in response to a user's request.

Those documents explain access and linking behavior. They do not publish a universal recommendation score. The audit should report the reasons and sources visible in the observed answer, not an invented formula for why one company won.

Build an evidence coverage comparison

Place the buyer requirements down the left side of a working sheet. For each requirement, record the competitor page, your page, the type of evidence, and the next action. Empty cells should remain empty. They are more useful than a guessed claim.

Start with service fit. Can a reader identify the operational problem, intended customer, required inputs, human review point, deliverable, and acceptance method? If the competitor answers the buyer's constraint and your homepage offers only a broad category label, the first remediation is a reviewed service explanation. Publishing ten pages that repeat the category will leave the decision gap in place.

Check entity clarity next. Company name, product name, official domain, service scope, and contact route should agree across the homepage, product pages, articles, language variants, and structured data. Metadata can express a clean relationship. It cannot repair visible pages that describe the product as a service in one place and a completed customer result in another.

Then examine claim support. A public method, named deliverable, acceptance record template, limitation, or documented demo boundary gives a buyer a way to verify what the company offers. When no approved customer outcome is public, publish the method and its actual limits. An invented percentage would make the page less defensible, even if it made the comparison sheet look full.

Treat independent references as evidence, not inventory

An industry article, partner document, maintained professional directory, or technical citation can give a buyer context beyond the company's own copy. Review its origin, date, and link to the underlying fact. A lightly edited press release remains company-supplied material even when another domain hosts it.

Reference volume has no stable meaning across AI products. Purchased profiles, reciprocal mentions, and unrelated backlinks create pages without clarifying the service. They also leave a poor evidence trail for a human procurement team.

The durable starting point is first-party material worth citing. Publish a stable explanation of the service method, a truthful project boundary, original research, or a reproducible public artifact. When a partner or journalist genuinely uses that material, help them retain the correct company name and source URL. Independent references should follow real work and real relationships.

Verify the retrieval path separately

Strong public evidence still has to be reachable by the relevant system. Test robots rules, HTTP status, canonical URL, server-rendered content, and edge-security behavior on the production page.

Google's current guidance for generative search says a page must be indexed and eligible for a Search snippet before it can appear in those features. Google also states that meeting requirements does not guarantee crawling, indexing, or serving. OpenAI recommends allowing OAI-SearchBot for ChatGPT search discovery and citation. Perplexity recommends allowing PerplexityBot and, where a WAF is involved, checking both its published IP ranges and user agent.

A browser receiving HTTP 200 proves only the browser path. A WAF can return a challenge or 403 response to identified automated traffic. Each intended access path needs its own production response check.

Retrieval is one row in the comparison. Once access works, move back to the buyer request. Technical eligibility does not supply a missing delivery method or acceptance boundary.

Turn the audit into a short remediation queue

Prioritize the gap that prevents a buyer from making a decision.

If no public page answers the target request, create or revise that page with an authorized fact owner. If the page exists but the company and product relationship conflicts across the site, resolve the public identity. When service fit is clear but the important claims lack deliverables, acceptance evidence, or limitations, add reviewed support. Repair a crawl, canonical, or WAF problem when a production test proves it exists.

Write each action with a target URL, fact owner, closing evidence, and review date. "Improve GEO" cannot be verified. "Add the human approval point and acceptance record to the service page, then obtain business-owner review" can be closed and retested.

Open GEO Console can organize the public pages, buyer question, answer citations, evidence gaps, and remediation priorities in one review. It does not approve company facts and cannot control the answer produced by any search or AI platform.

Retest without erasing the earlier result

After the reviewed change is public and available for discovery, run the recorded prompt under the same conditions. Preserve the earlier answer. Compare recommendation reasons and citations before reducing the result to a brand present-or-absent flag.

Keep the measurement chain separate. Indexing shows that a search engine processed a page under its reporting rules. An AI mention is an observed answer event. Search impressions and referral visits come from their own reporting systems. An inquiry belongs in the form or CRM record. Movement at one stage does not prove movement at the next.

One retest is still one observation. A useful trend needs dated, comparable runs. The team should also record when the AI product, prompt, language, or region changed, because those changes limit the comparison.

Boundaries

This method applies to public company sites, service and product pages, articles, and approved project evidence. Customer-confidential data, private reports, tokens, account-only pages, and unapproved deliverables should remain private.

The audit does not infer model parameters, publish negative claims about a named competitor, or treat a third-party mention as proven causation. AI products use different and changing retrieval systems. A recommendation-readiness review can improve the clarity, accessibility, and support behind public claims, but it does not guarantee indexing, citations, recommendations, rankings, traffic, or inquiries.

Open GEO Console evaluates public technical foundations, buyer-question coverage, answer material, citation evidence, and remediation priorities. Its output is a set of reviewable website actions, not a promise that a platform will select the company.

A practical review sequence

Select one commercial request — Use an approved buying situation and remove all company names. Preserve the intended buyer, operational problem, and meaningful constraints.

Save one complete answer — Record the prompt, product, time, language, region, named companies, recommendation reasons, and cited URLs. Open every citation and note the claim it supports.

Compare the supporting pages — Map buyer requirements to a small set of relevant pages for each company. Review service fit, entity consistency, evidence type, independent context, and production access.

Close one material gap — Choose the gap with the clearest effect on the buyer's decision. Assign the target page and fact owner, then define the evidence required to close it.

Repeat the recorded test — Retest after the public change can be discovered. Keep both answers and report each stage of the measurement chain separately.

Sources

• Google guide to how Search works
• Google guide to generative AI features on Search
• OpenAI publishers and developers FAQ
• Perplexity crawler documentation
• Bing URL submission and IndexNow guidance
• Open GEO Console in English

Next step

If an observed AI answer already names several competitors, bring one real buyer request and the relevant public URLs to Open GEO Console. The free check begins with the homepage, robots.txt, sitemap.xml, and public content, then organizes the first question-coverage, citation-evidence, or technical-access gaps.

Review the Open GEO Console project boundary for its public scope and simulated-demo limits. You can also inspect the enterprise AI service and other public projects. When the production URLs, buyer request, and expected deliverables are clear, submit the problem for a scoped discussion.

Related articles

AI search visibility — Google Indexed Your Website. Why Won't AI Recommend Your Company? — Google indexing makes a page eligible for search. It does not make a company defensible in an AI-generated supplier answer. Recommendation readiness depends on buyer-question fit, entity clarity, service boundaries, traceable evidence, and accessible public pages. — SolveReal Systems / Aug 30, 2026 / 9 min read

AI search and GEO — Why GEO Still Matters When Your Website Already Has SEO — SEO gives a website the foundations for discovery and indexing. GEO checks whether its brand, services, and evidence can also be understood and used in generated answers. — SolveReal Systems / Sep 8, 2026 / 7 min read

AI search and GEO — How Language Models Read Web Pages and What It Means for GEO — An accessible explanation of tokens, context, and semantic retrieval, with practical examples showing what language-model text processing means for GEO. — SolveReal Systems / Sep 4, 2026 / 8 min read

GEO vs. SEO: Practical Differences — Page Analysis

Executive Summary

This page, published by 实解智能 (Shijie Intelligence) on September 21, 2026, provides a practical, task-oriented comparison between Search Engine Optimization (SEO) and Generative Engine Optimization (GEO). The article argues that while SEO and GEO share technical foundations (crawling, indexing, quality content), they differ in what they optimize, how results are measured, and how work is validated. The page is well-structured, uses a comparison table, provides a hypothetical case study, and includes actionable implementation steps. It is written in Chinese and targets enterprise website owners and marketing teams.

1. Page Classification & Intent

2. Key Claims & Definitions

2.1 Core Definition Offered

The article proposes a working definition:

• SEO manages how pages perform in search discovery, display, and visits.
• GEO manages whether public facts appear in generative answers, whether they are stated correctly, and whether they receive appropriate source support.

Important caveat stated by the author: The boundary between SEO and GEO does not imply two separate algorithms inside search systems. The same page and the same facts can participate in both types of results simultaneously. The difference lies in which "task ticket" the team opens, what observations are recorded, and who decides the work is complete.

2.2 Google's Position (as cited)

The article cites Google's official guidance:

Google explicitly states that AI Overviews and AI Mode are built on its core search ranking and quality systems; existing SEO practices remain applicable, and there is no mandatory set of additional GEO-specific technical requirements.

The article interprets this as: GEO should be managed as content, evidence, and measurement work for generative answers — not as a replacement ranking system for SEO.

2.3 Source References Used

The page cites the following sources (verified as of September 21, 2026):

1. Google SEO Starter Guide
2. Google AI Features & Websites
3. Google Generative AI Search Optimization Guide
4. Google Generative AI Performance Report
5. Bing AI Performance documentation
6. OpenAI Publisher & Developer documentation
7. The original GEO paper (by Aggarwal et al.)

Note: The article explicitly warns that platform documentation and report fields may change, and execution should follow current official pages.

3. The SEO vs. GEO Comparison Table (Reproduced in Summary)

Author's caveat: This table represents a division of labor, not a technical boundary. SEO also handles entities, passages, and authority; generative answers also resolve to pages and URLs. The difference is where validation attention is placed.

4. Why the Optimization Unit Changes

The article explains that classic search results present page links, so teams work URL-by-URL. In Google's generative features:

• The system retrieves materials first, then organizes an answer.
• Google publicly states these features use retrieval-augmented generation (RAG) and may use query fan-out (splitting one question into multiple related searches).
• One answer may draw capabilities from a service page, implementation conditions from a documentation page, and evidence from a case study or policy page.

Key implication for content review: Page topic clarity remains important, but key facts must also stand alone when removed from context. Information such as: who provides the service, who it suits, what inputs are supported, which actions require human confirmation, how results are validated, and what limitations exist — if scattered across vague paragraphs, the system may retrieve the page but fail to combine facts accurately.

Important limitation stated: The original GEO paper studies content visibility in generative answers and proposes a testing framework. It provides a research starting point but does not establish a unified industry ranking standard. Experimental results on specific benchmarks cannot be directly converted into citation growth for any commercial website.

5. Hypothetical Case Study (发票识别 / Invoice Recognition)

The article presents a hypothetical example (explicitly stated as not representing a client or delivered project):

Scenario: An enterprise software company offers invoice recognition services. Its product page is indexed and appears for queries like "invoice OCR." The team opens two task tickets for the same page.

The buyer's specific question:

Which invoice recognition service supports Chinese scanned invoices, allows finance staff to review before writing to ERP, and can be deployed in the enterprise's own environment?

The original page only said: "high precision, automated, flexible deployment" — directional but not answering the three purchase-deciding questions.

Two validation tickets for the same page:

Key insight: These two tickets can drive the same page modification but cannot share a single completion condition. If product and delivery owners have not confirmed the on-premise deployment scope, the SEO task can continue observing page performance, but the GEO task cannot expand "flexible deployment" into a definite commitment.

Author's warning: Page ranking rising does not prove answer accuracy. An answer citing the page does not prove the user clicked or submitted an inquiry. These results are related but not interchangeable.

6. Metrics: The Most Confusing Difference

6.1 SEO Metrics (Mature)

• Google Search Console: clicks, impressions, CTR, average position
• Site analytics: post-visit reading, inquiries, other behaviors

6.2 GEO Metrics (Immature)

The article states clearly: There is currently no cross-platform universal "AI ranking." Different products expose different data.

Platform-specific data as of the article's writing:

6.3 Recommended Four-Layer GEO Recording

The article recommends recording GEO observations in four separate layers:

1. Appearance — Did the brand or page enter the answer and source list?
2. Expression — Did the answer correctly describe the service audience, capabilities, and limitations?
3. Support — Does the cited page actually support the corresponding claim in the answer?
4. Follow-up behavior — Did the user visit the official site, read key pages, or initiate an inquiry?

Warning: Recording only "mentioned by AI" conflates all four layers. Accurate mentions may lack citations; citations may contain incorrect descriptions; visits may not come from the target question.

7. Symptom-Based Prioritization Guide

Author's note: If you don't even know whether pages can be accessed or indexed, start with an AI search visibility audit. If foundations are stable, proceed to answer accuracy and source support.

8. Shared Foundations (Record Once)

The article lists shared foundations that serve both SEO and Google's generative features:

• Stable URLs
• Readable body text
• Internal links
• Accurate titles
• True facts
• Consistent structured data

Google's stated requirement (as cited): Pages need to be indexed and eligible for snippet display to appear as support links in Google's AI features.

Recommendation: Teams should not build two separate technical checklists for the two types of work. Shared issues go into one website backlog; SEO and GEO each keep their own validation records. GEO additionally maintains buyer questions, full answers, actual sources, and fact-check results.

9. Team Division of Labor

Key governance point: One person can hold multiple roles, but approval authority must be explicit. Content teams can discover that "whether on-premise deployment is supported" has no answer — but they must not invent the answer themselves. GEO brings facts into answers, which brings business commitments into public environments. High-impact claims require confirmation from someone who truly understands delivery conditions.

Review process: A joint question checklist can start a review, but final records should remain separate. SEO owners record page/query performance; GEO owners record answers and sources; business owners judge description accuracy.

10. Three Common Misconceptions

10.1 Treating GEO as Mass Q&A Writing

Creating a page for every long-tail question can produce duplicate and low-value content. Page count ≠ fact coverage. First confirm what decision information buyers need, then choose the best home (service page, documentation, case study, or article).

10.2 Treating Citation Counts as Final Business Results

Citations prove a page became a source in some answer — not that the answer was accurate, that users visited, or that inquiries occurred. Each layer needs its own evidence.

10.3 Combining Everything into One "GEO Score"

Rankings, citations, fact accuracy, visits, and inquiries have different meanings. Compressing them into one score hides where the problem occurred. Dashboards are acceptable, but raw fields and evidence should be preserved.

11. Four-Week Implementation Plan

Outcome after four weeks: A problem list, page list, fact gaps, and a re-checkable baseline — more useful for planning next work than an undefined "GEO score."

12. Applicability Boundaries (Explicitly Stated)

• The article discusses visibility of public websites in search results and generative answers.
• Different platforms use different indexes, crawl policies, retrieval methods, and answer interfaces.
• Google's rules cannot substitute for Bing, OpenAI, or other platforms' actual documentation.
• Completing SEO or GEO changes does not guarantee indexing, ranking, citations, traffic, or inquiries.
• Platform reports only prove what their fields cover. Business value assessment requires combining website behavior, lead quality, and actual deal processes.

13. Validation Checklist (Provided by the Article)

1. SEO tasks specify target queries, URLs, and search performance evidence.
2. GEO tasks specify target questions, platforms, factual claims, and answer evidence.
3. Shared technical issues get one backlog item; two validation records remain independent.
4. Rankings, citations, accurate descriptions, visits, and inquiries are not combined into one vague score.
5. High-impact claims have a fact owner; content staff do not expand service promises on their own.
6. Chinese and English pages are checked separately; retesting preserves full answers and actual sources.
7. Reports clearly distinguish observations, inferences, and unverified results.

14. Sources Cited (as listed on page)

1. Google SEO Starter Guide
2. Google AI Features & Websites
3. Google Generative AI Search Optimization Guide
4. Google Generative AI Performance Report
5. Bing AI Performance documentation
6. OpenAI Publisher & Developer documentation
7. GEO original paper

Verification note: Materials checked September 21, 2026. Platform documentation and report fields may update; actual execution should follow current official pages.

15. Internal Links & Related Content

The page links to related articles on the same site:

1. 网站已经做了 SEO，为什么还需要 GEO (September 8, 2026, 7 min read) — Why GEO is needed after SEO is done
2. 官网被 AI 引用以后，怎么判断 GEO 改动有没有用 (September 18, 2026, 6 min read) — How to judge whether GEO changes worked after AI citations
3. 从一个问题到带来源的回答：AI 搜索怎样使用网页 (September 16, 2026, 6 min read) — How AI search uses web pages, from question to sourced answer

Conversion paths: Open GEO Console, Open GEO project description, enterprise AI systems services, contact page.

16. Quality Assessment

Strengths

• Clear operational framework: The SEO/GEO comparison table is practical and actionable.
• Appropriate caveats: The article consistently warns against over-interpretation (e.g., citations ≠ business results, no universal AI ranking exists).
• Platform-specific accuracy: Correctly distinguishes Google, Bing, and OpenAI data availability.
• Governance emphasis: Strong on who approves facts and why that matters.
• Honest about limitations: Explicitly states that the GEO paper is research, not an industry standard.
• Structured for scanning: Tables, symptom guides, and checklists make it usable.

Limitations / Uncertainties

• Dated information: All platform references are as of September 2026; the article itself acknowledges these may change.
• Hypothetical case study: The invoice recognition example is explicitly not a real client project.
• No original research: The article synthesizes existing public documentation rather than presenting new data.
• Chinese-language focus: The brand consistency example (Chinese vs. English answers) may not apply to all markets.
• No measurement of its own claims: The article does not provide evidence that its recommended framework produces results.

Potential Conflicts of Interest

• The article promotes the author's own tools (Open GEO Console) and services (enterprise AI systems) at the end. This is disclosed through context but is a commercial motivation to consider.

17. Overall Assessment

This is a high-quality, practical reference article for teams trying to operationalize GEO alongside existing SEO work. Its primary value lies in:

1. Providing a clear division of labor between SEO and GEO
2. Establishing a four-layer measurement framework
3. Offering a symptom-based prioritization guide
4. Emphasizing governance and fact-approval processes

The article is not a research paper, does not present original experimental data, and its platform-specific claims are time-bound. It is best used as a working framework rather than an authoritative specification.

Recommended use: Teams with an existing SEO foundation who want to extend into GEO should use this article's checklist and four-week plan as a starting point, while verifying all platform-specific claims against current official documentation.

Analysis based on page content observed September 30, 2026. All platform references are as cited by the article and dated September 21, 2026.

Freight Lead Agent — 页面分析报告

页面概述

该页面介绍了一个名为 Freight Lead Agent 的可部署系统，定位为“Google 地图获客与定制营销系统”。页面以中文撰写，面向的受众是销售人员或小团队，核心卖点是将“从 Google 地图找企业 → 读官网 → 找联系方式 → 写定制开发信 → 发送跟进”这一完整流程自动化，同时保留人工审核环节。

核心功能拆解

页面将系统能力分为四个层次：

1. 系统接管的部分：按关键词和地区从 Google 地图采集目标企业，保留企业来源与官网；逐家访问官网，提取业务介绍、服务信息与公开联系方式；结合对方官网内容和发件方资料生成定制开发邮件，进入人工审核与发送计划。
2. 人工保留的部分：逐条确认目标企业、官网内容、公开联系方式和定制邮件是否准确；人工决定是否修改、发送、安排节奏或停止联系，并处理后续回复。
3. 最终交付物：包含企业来源、官网、公开联系方式和处理状态的目标客户库；依据每家企业官网内容生成、可逐条审核的定制开发邮件；带发送节奏、发件账号、投递状态和回复跟进的发送计划。
4. 技术形态：项目已形成可独立部署的 Docker 系统，覆盖从企业发现到回复监控的完整链路。

使用范围与限制

页面明确列出两条边界：

• 只处理公开企业信息，不抓取登录后页面或个人资料。
• 外联内容必须人工审核，系统不提供无差别批量自动发送。

这两条限制值得注意：前者界定了数据来源的合法性边界，后者则明确该系统并非全自动群发工具，而是“半自动”的辅助系统。

页面末尾的特别声明

页面最后一段写道：“你的业务不需要复制这个项目；我们会从实际流程重新诊断。”并附有“讨论你的流程”的引导。这表明该页面不仅是产品介绍，还承担着销售线索转化的功能——它暗示读者无需自行部署这套系统，而是可以直接与卖方讨论定制化方案。

页面中未明确的信息

• 定价：页面未提及任何价格信息。
• 部署要求：虽然提到 Docker，但未说明服务器配置、运行环境或维护成本。
• 数据合规细节：虽然声明“只处理公开企业信息”，但未涉及 GDPR、中国《个人信息保护法》或其他司法辖区的具体合规说明。
• 效果数据：未提供任何案例数据、成功率或客户见证。
• 技术细节：未说明 Google 地图数据采集的具体方式、频率限制或反爬应对策略。
• 交付周期：未说明从购买到部署完成的时间预期。

页面语气与可信度评估

页面语气务实，没有夸大承诺。它明确承认“这套工作很难靠人工持续放大”，并强调人工审核环节不可省略，这在一定程度上增加了可信度。但页面也缺少可验证的第三方证据（如客户案例、技术文档链接或可运行的演示环境），因此其实际效果仍需通过进一步沟通验证。

总结

这是一个功能描述清晰、边界明确的 B2B 工具介绍页。它解决的是销售开发流程中“找企业、读官网、写定制信”这一重复性劳动密集环节，同时保留了人工审核的最终控制权。页面最大的未解问题是定价、实际效果和合规细节，这些都需要通过页面底部的“讨论你的流程”入口进一步获取。

How to Measure GEO AI Citations: A Practical Framework

Publisher: 实解智能 (Shijie Intelligence) | Date: September 18, 2026

Core Argument

A single AI citation is only a starting point for observation. To determine whether GEO (Generative Engine Optimization) changes to your website are working, you must separately track four layers of evidence: whether the answer is accurate, whether the citation supports the conclusion, whether visits occurred, and whether inquiries were valid. A citation link alone only shows that your website appeared in a particular answer or platform statistic; it cannot prove the answer described your business correctly, that anyone clicked through, or that a later inquiry resulted from that specific change.

What a Screenshot Can and Cannot Prove

The article uses a hypothetical example (explicitly labeled as fictional, not a project result) of a company that completes the applicable conditions on its service page. The next day, an employee asks a connected assistant a question and sees the new page cited. The screenshot proves what that account, at that time, with that phrasing, displayed. It cannot show whether the page was occasionally cited before the change, how other buyers might phrase their questions, or rule out changes from search indexing, user demand, or model updates.

A more easily overlooked issue is the answer text itself. If the link points to your service page but the answer still describes conditional capabilities as unconditional promises, the citation actually masks the problem. The article cites the ALCE citation evaluation research, which separates answer correctness from citation support, as the reason for this distinction.

Four Layers of Evidence

Platform Data: Read According to Its Own Definitions

Bing Webmaster Tools: The AI Performance documentation lists visible citations, cited pages, and topic phrases used for retrieval as observation items. These phrases are not the user's original queries, and citation counts do not represent ranking, authority, or clicks. Coverage is limited to Microsoft Copilot, Bing's AI summaries, and some partner experiences—not a total count across all AI platforms. Bing's official release notes also position this data as aggregated citation visibility.

Google: Official documentation states that traffic from AI Overviews and AI Mode is included in the overall Search Console "Pages" search type data. Therefore, you cannot read the total table as showing how many visits each individual AI answer generated.

OpenAI: The publisher FAQ states that links from ChatGPT to websites carry utm_source=chatgpt.com, allowing sites to view recordable referral visits. Visit records still differ from answer impression counts and cannot replace demand assessment in sales conversations.

How to Run a Comparable Before/After Observation

1. Select real buyer questions encountered in sales or delivery—not just phrasings most likely to surface your brand name. For each, write down the facts and conditions you want the answer to state clearly, and which page on your site provides the evidence. For example, "Can you handle scanned documents?" should also record "requires text recognition first" and "results are human-reviewed." Counting only brand mentions will miss these critical errors.

2. Before changes: Fix the exact question wording, product entry point, whether search is enabled, language and region, and test date. Save the full answer and citation URL, and record the website page version.

3. After changes: Re-ask using the same recording criteria. Keep records of non-citations and incorrect answers as well.

4. Sample size: Different dates, phrasings, and products can yield different answers, so evaluate a set of questions over a period—do not cherry-pick the single best result. There is no universal "test ten times for significance" threshold; sample size should be determined by decision risk and question volatility.

5. Connect visits and inquiries, but maintain evidence boundaries: Use your own analytics for traffic, and forms, email, or CRM records to verify whether inquiries match target needs. Without client consent or a reliable source record, write only "observed concurrent inquiries"—do not attribute them to a specific AI answer.

6. Allow time for discovery and re-crawling: Google states that content changes may take time to be reflected in search systems.

The Website's Role in This Framework

The website is both a potential citation source and the place buyers verify promises after clicking. Service descriptions should state current capabilities and limitations; project records should show only publicly confirmed facts. Writing facts clearly, making target pages accessible, and ensuring the same question can find corresponding evidence are tasks within your control. Whether a platform chooses to cite you and how it displays you remains the platform's decision.

Open GEO Console: Where It Fits

The Open GEO Console is positioned as a website diagnosis and remediation entry point: it checks public pages against target buyer questions, organizes gaps in accessibility, content expression, question coverage, and citation evidence, and provides pages needing modification with priorities. It is not a real-time monitoring system for cross-platform citations or sales conversion, and cannot prove causal improvement from a single test. Product scope is described on the Open GEO project page.

Applicability Boundaries

This method suits judging "whether problems decreased, whether answers became more accurate, and whether follow-up business signals appeared"—it is not a randomized controlled experiment. Manual questioning is sampled observation; platform reports have their own coverage limits; website analytics can be affected by browser settings and collection methods. Bing's official documentation explicitly states that citation trend changes cannot be attributed to a single content modification. Without sufficiently stable before/after samples, write "observed changes," not "the modification caused growth."

Implementation Steps

1. Select real buyer questions; first record the current website page and the original AI answer.
2. Mark key facts, missing conditions, and cited text in each answer.
3. After modifying the website, keep the page version and re-ask using the same recording criteria.
4. Finally, cross-reference your own visit and inquiry records, and summarize "answer improvement," "citation appearance," "visit occurrence," and "valid demand" separately—do not merge them into one seemingly precise GEO score.

Acceptance Checklist

A reviewable record should include: the original question and questioning environment, date, full answer, citation URL, the website version checked, item-by-item fact judgments, and the source of visit or inquiry data. Negative samples must also be kept. If evidence reaches only "the website was cited," stop the conclusion at that layer; if subsequent visits and inquiries can be verified, continue downward, but do not skip attribution limitations.

Next Steps

If you already know what buyers will ask but are unsure which pages state the answers clearly, you can start with Open GEO Console for website detection. To discuss specific pages, business questions, and reviewable acceptance criteria, use the contact page. Test results and final public statements remain the responsibility of the website owner for verification.

Reference Sources

• Bing Webmaster Tools: AI Performance metrics documentation
• Bing official blog: AI Performance public beta release notes
• Google Search Central: AI features and websites
• OpenAI: Publisher and developer FAQ
• ALCE: Citation-based long-form answer evaluation research

Related Articles (from this site)

• GEO vs. SEO: What's the Difference? An Executable Work Map (实解智能, 2026/9/21, 12 min read)—Both share crawling, indexing, and quality content foundations, but differ in observed outcomes, optimization units, and acceptance methods.
• From a Question to a Sourced Answer: How AI Search Uses Web Pages (实解智能, 2026/9/16, 6 min read)—Explains how connected AI rewrites questions, finds candidate pages, selects material, and generates cited answers along a verifiable process.
• After a Page Update, When Will a Large Model Read It? From Training Text to GEO Retrieval (实解智能, 2026/9/16, 6 min read)—Distinguishes how models process text during training from how AI search retrieves pages when answering, clarifying what GEO can and cannot change.

Note: This analysis is based solely on the supplied page content. The page is a Chinese-language article from 实解智能 (Shijie Intelligence) published September 18, 2026. All quotations and factual claims above are drawn directly from the retained text. The article itself emphasizes that its example is fictional, that platform data must be read according to each platform's own definitions, and that causal claims require stable before/after samples that manual observation may not provide.

Project Library — Page Analysis for GEO Report

Page Overview

This page, titled "Project library," serves as a catalog of four enterprise AI-related products and systems offered by the site. The page's stated purpose is to explain "what each project solves and why it is worth buying," with each case structured to show the customer's current workflow, which steps the system takes over, and what the customer gains after purchase.

The page presents four distinct offerings, each labeled with a number and a category tag. Two are described as "deployable systems" and one as a "public product," with the fourth also marked as a deployable system. All descriptions follow a consistent three-part format: current workflow, how the system works, and what the customer gains.

Project Summaries

1. AI Search Visibility Diagnosis and Remediation (Public product, available now)

This is the only offering explicitly marked as currently available. It diagnoses an enterprise website's visibility in AI search by examining technical foundations, buyer questions, public answers, and citation evidence. The output is described as an executive report and a vendor task package.

The stated problem: companies cannot tell whether AI systems understand their site, which sources buyer answers cite, or why their brand is omitted. The system consolidates technical checks, public-search evidence, citation gaps, and remediation priorities into one diagnostic workflow. The claimed benefit is an actionable, ongoing GEO optimization plan covering what to change, how to change it, and what to monitor.

A link labeled "Open GEO Console" and an "Explore the product" button are provided.

2. Hermes Notebook — Enterprise Data Engine for AI Applications

This deployable system converts enterprise information scattered across PDFs, Word, Excel, PowerPoint, images, local folders, and business web pages into citable, retrievable knowledge. The knowledge is intended to support customer-service bots, internal Q&A, and other business assistants through APIs.

The stated problem: valuable company information is fragmented across multiple formats, employees struggle to find it, and separate AI applications repeatedly prepare the same material. The system imports multiple formats, restores document structure, builds traceable knowledge nodes and retrieval indexes, and supplies evidence to downstream applications through permission-scoped query and retrieval APIs.

The claimed benefit: customers receive an enterprise knowledge engine they control, allowing multiple AI applications to reuse the same prepared data instead of cleaning and importing it repeatedly.

3. Freight Lead Agent — Google Maps Prospecting and Personalized Outreach

This deployable system finds target companies through Google Maps, reads each official website for business context and public contacts, prepares personalized outreach messages, and carries them through review, sending, and reply follow-up.

The stated problem: sales teams must manually find companies on Google Maps, verify websites, understand each business, locate contacts, and write different messages per company — a process that is difficult to scale. The system automates company collection by keyword and region, extracts context and contacts, generates personalized messages, and routes them to sales for review, delivery, and follow-up.

The claimed benefit: a small team can sustain discovery, research, and personalized writing while reducing the effort of starting from zero for each prospect and avoiding generic mass messaging.

4. Codex Feishu Bridge — Enterprise AI Collaboration System

This deployable system connects Codex to an enterprise Feishu workspace. Members can submit questions, files, and long-running tasks from a computer or phone, then add context, check status, and receive deliverables remotely while an online company workstation continues execution.

The stated problem: employees repeatedly move requests, files, and results between Feishu and separate AI tools; complex work is often delegated to a few Codex users; and tasks become difficult to continue once people leave the workstation. The system hands work to Codex on an online company workstation, accepts follow-up input from authorized members in the same topic, and returns results and files there.

The claimed benefit: an enterprise AI work entry point deployed on the customer's own workstation, with mobile access for continued task management and the ability to turn proven workflows into reusable "Skills" so later work follows the same method without depending on a few experienced employees.

Observations and Limitations

• Pricing and availability: No pricing information appears anywhere on this page. Only the first project is explicitly marked as "available now"; the other three are labeled "deployable system" without stated availability dates or deployment timelines.
• Evidence and validation: The page provides no customer testimonials, case-study metrics, technical specifications, security certifications, or independent verification of the claimed capabilities. All benefits are stated as assertions rather than demonstrated outcomes.
• Geographic and language context: The URL contains "me.itheheda.online," and the page is written in English. No information is provided about supported languages, geographic availability, or data-residency options.
• Technical depth: The descriptions are conceptual. No API documentation, architecture diagrams, integration details, or system requirements are provided on this page.
• Contact and next steps: Each project has a "View project" or "Explore the product" link, but no contact form, email address, or sales inquiry mechanism is visible on this page itself.
• Page freshness: The page was observed on September 30, 2026. No publication or last-updated date is shown on the page.

Conclusion

This page functions as a high-level catalog of four AI-related enterprise offerings. It clearly explains each product's intended problem, mechanism, and claimed benefit in a consistent format. However, it lacks pricing, availability details for three of four products, technical specifications, customer evidence, and direct contact mechanisms. Buyers would need to follow the individual project links to obtain deeper information before making a purchasing decision.

Page Analysis: Articles & Field Notes (me.itheheda.online)

Overview

This page serves as the articles and field notes hub for SolveReal Systems, presenting a categorized collection of research and practical guidance on enterprise AI systems, automation, AI search visibility, and delivery boundaries. The page is structured as a content index with three categories: AI search and GEO (8 articles), Enterprise AI delivery (1 article), and AI search visibility (5 articles), totaling 14 articles.

Content Assessment

Category 1: AI search and GEO (8 articles)

The largest category focuses on Generative Engine Optimization (GEO) and its relationship to traditional SEO. Articles cover:

• Content consistency governance — "When Your Website Contradicts Itself, What Should GEO Fix First?" (2026-09-21, 10 min read) addresses prioritizing underlying public facts over content production when pages conflict.
• GEO vs. SEO distinctions — "GEO vs SEO: A Practical Map of the Differences" (2026-09-21, 13 min read) outlines shared foundations (crawling, indexing, useful content) while differentiating outcomes, improvement units, and acceptance evidence.
• Measurement methodology — "When AI Cites Your Site, Did the GEO Change Work?" (2026-09-18, 8 min read) argues a citation is "one observation, not a verdict" and recommends measuring answer accuracy, citation support, visits, and qualified inquiries separately.
• AI source usage mechanics — Articles dated 2026-09-16 cover how AI products move from buyer questions to cited answers, the relative weight of websites/news/social posts (noting "citation counts alone do not reveal their influence"), and the difference between model training and retrieval processes.
• Foundational GEO concepts — "Why GEO Still Matters When Your Website Already Has SEO" (2026-09-08) and "How Language Models Read Web Pages and What It Means for GEO" (2026-09-04) provide accessible explanations of tokens, context, and semantic retrieval.

Category 2: Enterprise AI delivery (1 article)

A single article, "How to Choose an AI Automation Provider: Delivery and Acceptance Checklist" (2026-09-04, 9 min read), covers provider evaluation criteria including SaaS/RPA/custom fit assessment, real samples, human-review boundaries, recovery paths, data permissions, and acceptance evidence.

Category 3: AI search visibility (5 articles)

This category addresses practical visibility problems:

• Website importance in the AI era — "Why Company Websites Matter More in the AI Era" (2026-09-02, 14 min read) traces the path from discovery to business action.
• Brand name misreading — "Google Indexed the Site. Why Does AI Search Still Misread the English Brand Name?" (2026-09-01, 12 min read) presents a public review separating Google AI Overviews, ChatGPT Search, and Perplexity, with a bilingual repair checklist.
• Competitor exclusion diagnosis — "Why Does AI Recommend Your Competitors but Leave Out Your Company?" (2026-08-31) proposes controlled comparisons to build remediation queues "without pretending to know a model's ranking formula."
• Recommendation readiness — "Google Indexed Your Website. Why Won't AI Recommend Your Company?" (2026-08-30) identifies buyer-question fit, entity clarity, service boundaries, traceable evidence, and accessible public pages as readiness factors.
• Visibility audit framework — "Why AI Search Cannot See Your Website: A Practical GEO Visibility Audit" (2026-08-29) converts discovery, crawling, indexing, coverage, and citation gaps into "owned, verifiable remediation task[s]."

Publication Patterns

• Recency: All articles were published between 2026-08-29 and 2026-09-21, indicating an active, ongoing content program.
• Consistency: All content is attributed to SolveReal Systems, suggesting a single-author or single-brand operation.
• Format: Each listing includes title, summary, date, and read time; the page notes that "Titles, summaries, and full-article links remain readable without opening the category."

Positioning and Authority Signals

The content demonstrates a methodical, evidence-based approach to GEO. Notable positioning elements:

• Measured claims: Articles repeatedly caution against over-attributing outcomes (e.g., "A cited page is one observation, not a verdict").
• Practical frameworks: Content consistently converts abstract AI behavior into testable, owner-verifiable tasks.
• Transparency: One article is explicitly framed as a "public SolveReal Systems review," suggesting willingness to share real client work.
• Educational tone: The page positions SolveReal Systems as a thought leader explaining complex AI-search mechanics to a business audience.

Potential Gaps and Observations

• Category imbalance: The Enterprise AI delivery category contains only one article, which may appear thin relative to the other categories.
• No author attribution: Individual authors are not named; all content is attributed to the brand.
• No engagement signals visible: The page does not display comments, social shares, or other reader interaction metrics.
• Limited supplementary content: No featured articles, newsletters, downloadable resources, or calls-to-action are visible on this index page.

Overall Assessment

This page functions as a well-organized, topically coherent content hub for a specialized consultancy. The article collection demonstrates deep subject-matter expertise in GEO and AI search visibility, with a consistent editorial voice emphasizing practical, verifiable methodologies over speculation. The publication cadence (14 articles in approximately three weeks) indicates active content production. For a paid GEO report, this page suggests the site owner is investing in educational content that addresses common buyer questions about AI search visibility — content that itself may serve as citable material for AI answer engines.

Page Analysis: "From Prompt to Answer: How Text LLMs Use Web Pages"

URL: https://me.itheheda.online/articles/from-prompt-to-answer-how-text-llms-work
Observed: September 30, 2026
Source length: ~3,038 characters (Chinese)

1. What This Page Is

This is a Chinese-language explainer article published by 实解智能 (Shijie Intelligence) on September 16, 2026. It walks readers through how connected AI search systems take a user question, rewrite it, retrieve candidate web pages, select material, and generate a cited answer. The article is explicitly framed as a GEO (Generative Engine Optimization) educational piece — its stated purpose is to explain what a corporate website can and cannot do in the context of AI search visibility.

The page is directly readable (no paywall, no login required) and appears to be a standard content-marketing article with internal links to the company's service page, Open GEO project page, and contact page.

2. Core Claims and Key Facts

The article makes several specific, verifiable claims about how AI search systems work:

The article is careful to state its limits: it does not claim to know internal ranking formulas, query scoring, or per-platform retrieval details. It explicitly says the described flow is a "comprehensible path" assembled from public documentation and the RAG paper, not a disclosure of any commercial product's internal implementation.

3. Practical Guidance Offered

The article provides a concrete methodology for GEO auditing:

Implementation steps:

1. Start with a real buyer question; record the original wording, query time, and whether the target AI product had search enabled.
2. Save the full answer and cited URLs.
3. Go sentence-by-sentence back to the original website and media sources; mark which sentence is supported by which material, and where dates, limitations, or human-confirmation conditions were dropped.
4. Check whether the corresponding website page is publicly accessible and whether the body text puts capabilities, prerequisites, and evidence together.
5. After making changes, re-test with the same question set and keep before/after records.

Acceptance checklist (four separate items):

• Whether the page can be retrieved by the target system
• Whether the answer mentions the company
• Whether key facts are accurate
• Whether displayed citations actually support those facts

The article stresses that a page returning HTTP 200 only proves accessibility, and a single citation in one answer does not prove long-term stable adoption.

4. Internal Links and Site Context

The page links to several internal destinations:

• 服务页 (Services page) — describes the company's AI service approach and delivery boundaries
• Open GEO 项目页 (Open GEO project page) — describes the public facts of the website-visibility diagnostic product
• 其他项目页 (Other project pages) — provide project evidence
• Open GEO Console — a tool for checking public pages, buyer-question coverage, and verifiable citation evidence
• 联系页 (Contact page) — for submitting specific URLs and questions

It also links to two related articles:

• "AI 搜索可见性" (AI Search Visibility) — September 2, 2026
• "GEO 和 SEO 到底有什么区别？" (GEO vs. SEO) — September 21, 2026
• "官网被 AI 引用以后，怎么判断 GEO 改动有没有用？" (How to judge whether GEO changes worked after citation) — September 18, 2026

5. External References Cited

The article lists five external sources:

1. OpenAI: ChatGPT 搜索说明 (ChatGPT search documentation)
2. Google: AI 搜索功能与网站 (AI search features and websites)
3. Google: 搜索系统如何工作 (How Google Search works)
4. Lewis et al.: 检索增强生成原始论文 (original RAG paper)
5. Gao et al.: ALCE 引用评估研究 (ALCE citation evaluation study)

Note: The article does not provide direct URLs for these references — only names. A reader cannot click through to verify the sources from this page.

6. Quality Assessment

Strengths:

• Clearly distinguishes between what is publicly documented and what is not (e.g., ranking formulas, query scoring)
• Explicitly states that the described flow applies to search-enabled answers, not all text-model behavior
• Uses a fictional buyer question and explicitly labels it as fictional, avoiding false claims about client work
• Separates the concepts of "page is accessible," "page is indexed," "page is retrieved," and "page is cited" — a useful distinction often blurred in GEO marketing
• Provides a repeatable testing methodology rather than vague advice
• Acknowledges uncertainty: "任何'被引用率提升多少'的结论，都需要指定产品、问题、时间与可复测的观察" (Any conclusion about citation-rate improvement requires specifying the product, question, time, and repeatable observation)

Weaknesses / limitations:

• External references are named but not linked, reducing verifiability
• The article is self-published by a company that sells GEO services; it is not an independent source
• No data, case studies, or measurements are presented — the article is purely explanatory
• The "related articles" section promotes the company's own content, which is expected but worth noting
• The page does not state its own publication date's relevance window; AI search behavior may change after September 2026

7. Commercial Intent

The page is clearly a lead-generation / trust-building article for 实解智能's GEO consulting and Open GEO diagnostic products. It educates potential clients on how AI search works, then positions the company's tools as the way to audit and improve a website's visibility. The tone is measured and non-hyped, which strengthens credibility, but the commercial purpose is unambiguous.

8. Overall Verdict

This is a well-structured, honest, and technically careful explainer about how AI search systems use web pages, written for a business audience. It avoids overclaiming, clearly separates documented facts from inference, and provides a practical audit methodology. Its main weaknesses are the absence of clickable external references and the lack of any empirical data. For a paid GEO report, this page is a useful educational reference but should not be treated as an authoritative or independent source — it is vendor content with a commercial agenda.

Suggested citation for this page:
实解智能 (2026, September 16). 从一个问题到带来源的回答：AI 搜索怎样使用网页. Retrieved September 30, 2026, from https://me.itheheda.online/articles/from-prompt-to-answer-how-text-llms-work

Page Analysis: Codex Feishu Bridge (Enterprise AI Office Collaboration System)

Page Overview

This page presents Codex Feishu Bridge, an enterprise AI office collaboration system that integrates OpenAI's Codex into Feishu (Lark), a Chinese enterprise communication platform. The system is described as a deployable solution that allows team members to submit questions, files, and long-term tasks via Feishu from computers or mobile devices, with tasks executed by work computers that remain online within the company.

Key Claims and Capabilities

Problem Statement

The page identifies three core workflow problems:

• Employees must repeatedly transfer requirements and results between Feishu, files, and AI tools
• Complex tasks are typically delegated to a few employees who know how to use Codex
• Tasks cannot proceed when members leave their company computers

Proposed Solution

• Members submit requirements and attachments via Feishu on computer or mobile
• The system assigns tasks to online work computers within the company
• Codex executes tasks in a designated working directory
• Authorized members of the same topic can continue adding information
• Results and files are returned to the original topic thread

System Functions

The system claims to handle:

• Permission verification for group chats and members
• Distinguishing between direct answers and formal trackable tasks
• Recording task states: accepted, queued, executing, waiting for input, and delivered
• Returning text, attachments, or Feishu cloud space links to the original topic
• Converting validated workflows into reusable "Skills" for future tasks

Human Oversight

The page emphasizes that humans retain control over:

• Providing clarification when Codex lacks critical input
• Reviewing final content and attachments for completeness
• Authorizing deployment, real external sending, and paid generation actions
• Confirming new Skills before they are used in subsequent tasks

Technical Requirements and Limitations

Deployment Requirements

• Windows 10/11 x64 deployable version is available
• Client workstations must have Codex installed and logged in
• Both the work computer and bridge service must remain online for task execution

Explicit Limitations

• Mobile Feishu serves only as a remote task entry point; actual processing occurs on company work computers
• Tasks cannot proceed if the work computer is off or the bridge service is offline
• The system does not automatically learn employee behavior without confirmation
• The system does not replace human confirmation of final content, files, or external actions
• Access is restricted to authorized group chats and members only

Deliverables

The page states the system provides:

• Direct answers in the original Feishu topic, or asynchronous task status with task numbers
• Text, images, Office files, videos, and other artifacts generated or modified by Codex
• Queryable records of acceptance, execution, waiting, completion, failure, or uncertain status
• Reusable Skills derived from repeated effective workflows

Commercial Positioning

The page concludes with a notable statement: "Your business does not need to replicate this project; we will re-diagnose from your actual processes" — indicating this is a custom service offering rather than a standardized product. A call-to-action button reads "Discuss your process."

Assessment Notes

This is a project showcase page rather than a technical documentation page. It provides a high-level overview of capabilities and limitations but lacks specific technical details such as architecture diagrams, API documentation, security specifications, or performance benchmarks. The page is written in Chinese and appears targeted at enterprise decision-makers evaluating AI workflow integration solutions.

Uncertainty factors: The page does not specify pricing, implementation timelines, scalability limits, or detailed security/compliance certifications. The actual performance and reliability of the system would require direct verification through a demonstration or pilot deployment.

Page Analysis: "When Your Website Contradicts Itself, What Should GEO Fix First?"

Overview

This is a paid GEO (Generative Engine Optimization) report analyzing the page at https://me.itheheda.online/en/articles/geo-public-fact-governance-website-consistency, published by SolveReal Systems on September 21, 2026. The page addresses content governance for websites with contradictory public claims, positioning itself as a practical guide for establishing a single source of truth before pursuing retrieval and measurement strategies.

Key Facts and Claims

Core Thesis: When a homepage, service page, old article, and translated page state different conditions for the same capability, the first GEO task is to establish which claim is still true and who can approve it. The page argues that more articles, Schema markup, and backlinks cannot decide which business promise is valid.

Proposed Framework: The article recommends creating one governed record for every public fact that can change a buying decision. Each record should include:

• Fact ID and approved claim
• Fact owner (named role or person with decision authority)
• Public evidence page
• Scope (customer, region, version, or input qualifications)
• Evidence status (live, planned, experimental, or third-party)
• Time status (effective date and next review date)
• Dependencies (pages and machine-readable surfaces)
• Language status

Priority Matrix: The page provides a conflict-priority table identifying high-risk contradictions:

• Eligibility and service scope
• Price, region, and delivery time
• Human review, safety, and privacy conditions
• Product status and version requirements
• Brand and legal identity

Stylistic variation is classified as low priority.

Technical References: The article cites Google Search Central guidance on publication dates, sitemaps, structured-data policies, localized versions, and canonical tags, plus IndexNow FAQ, OpenAI crawler documentation, and W3C PROV-O. It notes that canonical tags have no role in deciding which commercial claim is true, and that structured data must represent visible content.

Limitations and Boundaries

The page explicitly states several limitations:

• Consistent facts do not guarantee indexing or citation by any platform
• Allowing OAI-SearchBot makes a page eligible for consideration but does not promise placement or citation
• The record does not replace legal, compliance, or professional review
• External systems can still use old caches, third-party pages, or other sources
• The hypothetical company example does not describe a SolveReal Systems customer or project

Practical Recommendations

The article suggests selecting ten to twenty public facts that materially affect buying decisions, assigning owners and evidence pages, auditing all language versions and structured data, repairing false commitments before stylistic variations, and treating fact changes as content releases with synchronized visible copy, structured data, and dates.

Verification Checklist

The page includes a seven-point verification checklist covering fact ownership, visible copy accuracy, cross-page consistency, language version alignment, date accuracy, historical page handling, and release records.

Next Steps Offered

The page promotes "Open GEO Console" as a tool that examines public pages against buyer questions, but clarifies it cannot decide which business claim is true, approve changes, or publish content. It directs readers to review the current service description, browse public projects, or use the contact page with a public URL and buyer questions.

Source Quality

The article cites official documentation from Google Search Central, IndexNow, OpenAI, and W3C. The page is directly readable with a source length of 12,677 characters. The content is structured with clear headings, tables, and a practical checklist format.

Overall Assessment

This page provides a structured, actionable framework for content governance in GEO contexts, with clear boundaries on what it does and does not address. It is written for organizations managing bilingual or multi-page websites with conflicting public claims. The recommendations are specific and verifiable, and the limitations are stated transparently.

GEO 来源权重分析：官网、新闻与公众号的实际占比

页面分析报告 | 2026年9月30日

页面概述

本文由“实解智能”发布于2026年9月16日，标题为《GEO 里官网、新闻和公众号各占多少？先分清被引用与被采纳》。文章旨在澄清生成式引擎优化（GEO）中“来源权重”这一概念的常见误解，并基于2026年的两项预印本研究，说明官网、新闻媒体和公众号在不同AI产品中的可见引用占比及其局限性。

核心论点

文章的核心主张是：不存在适用于所有AI产品的通用来源权重比例表。作者区分了三个容易混淆的概念：

1. 进入候选结果的机会（某类页面被检索系统选中的概率）
2. 被显示为引用的频率（链接出现在最终回答中的次数）
3. 事实被采纳的程度（页面内容是否真正支撑了回答中的具体判断）

文章强调，这三者的统计分母不同，不能混为一谈。

关键数据与研究发现

跨平台引用占比（2026年预印本研究）

文章引用了一项分析602个受控问题、21,143条有效搜索层引用的跨平台研究，给出了以下可见引用中的来源类型占比：

重要限制条件（原文明确说明）：

• 这些数字统计的是特定问题下已经显示出来的引用，并非产品设定的内部权重
• 研究未取得整个网络的候选池
• 来源以美国和英文网页为主
• 分类存在未知与噪声
• “官方类”涵盖多种机构页面，不能直接等同于企业官网
• 被引用的页面未必支撑回答中的关键判断

中文场景研究

另一项2026年中文生成式搜索预印本研究，对豆包、DeepSeek、腾讯元宝和通义千问的网页及App入口，围绕614个问题各采样三次，发现：

• 同一产品的网页端与App端，引用的来源集合存在系统性差别
• 研究集中在几个消费领域，仅观察已出现在引用列表中的网页
• 不能据此为所有行业的公众号或官网定价

方法论与证据评估

研究局限（原文承认）

1. 两项研究均为预印本，未经同行评审
2. 研究使用从答案和页面特征构造的代理分数，无法读取模型内部注意力分配
3. 论文未证明把某种格式加进网页就能提高采纳率
4. 相关性与代理分数不能直接证明因果关系
5. 特定问题和时间窗口的样本，不能外推至所有场景

引用评估研究

文章引用ALCE引用评估研究指出：在其ELI5测试集上，即使表现最好的系统，也有约一半的回答未获得完整的引用支持。但作者明确说明，该结果属于当时的研究数据集，不能换算成今天某款产品的错误率。

实践建议

来源的证明任务区分

文章建议根据问题类型选择合适来源：

• “这家公司现在提供什么、边界在哪里” → 官网服务说明是一手来源
• “它是否得到独立报道” → 应回到新闻原文，官网自述不能代替第三方报道
• “用户遇到了什么具体问题” → 有作者和时间的公开经验文章可能有价值，但需核实是否亲历、是否引用原始记录

实施步骤

1. 从真实买家问题开始，保留原问法
2. 记录各来源能证明的事实（原始网址、作者、发布日期、是否转载、矛盾之处）
3. 在目标AI产品的具体入口复问，保存时间、回答全文和引用页面
4. 逐句标出哪些主张由官网支持、哪些来自第三方、哪些无证据
5. 隔一段时间重复同一组问题，观察答案是否采纳官网更新

验收清单

• 官网被引用但未提供所需能力/限制/证据 → 补齐页面事实
• 旧报道压过当前说明 → 核对发布日期和冲突原文
• 多篇文章同稿转载 → 合并为一条证据链
• 区分“出现链接”“事实被正确采用”“引用支持了具体句子”三个结果

商业推广内容

文章末尾推广了“Open GEO Console”工具，声称可对官网和目标买家问题做GEO检测，检查公开页面、问题覆盖与引用证据，输出问题清单、修改建议和整改优先级。明确声明：该工具不读取AI平台的内部权重，也不保证下一次回答一定引用官网。

总体评估

优点

• 清晰区分了“被引用”与“被采纳”两个关键概念
• 提供了具体的研究数据，同时明确标注了所有限制条件
• 对研究方法的局限性有诚实披露
• 实践建议具体可操作

局限

• 所有数据来自2026年预印本研究，未经同行评审
• 样本以美国英文网页为主，中文场景研究仅覆盖消费领域
• 文章本身由GEO服务商发布，存在推广自家工具的潜在利益冲突
• 未提供任何可复现的原始数据或完整研究方法细节

适用边界（原文明确）

• 论文未公开各AI产品的内部选源规则
• 公众号、封闭社区或登录后内容是否进入某次回答，需按具体产品与网址核对
• 不能从一次截图推导平台长期的渠道占比

报告日期：2026年9月30日
分析对象：https://me.itheheda.online/articles/geo-source-weight-official-website

Hermes Notebook — 企业数据整理与 AI 应用引擎（页面分析）

页面概述

该页面介绍 Hermes Notebook，一个面向企业的数据整理与 AI 应用引擎。产品定位为“可部署系统”，核心价值是让企业资料可被 AI 检索和引用。页面以中文撰写，采用“原来的流程 / 系统接管什么 / 人工保留什么 / 最终交付什么”的结构化对比方式说明产品价值，并明确列出使用范围和限制。

核心功能

根据页面描述，Hermes Notebook 当前具备以下能力：

• 多格式导入：支持文件、文件夹和 URL 导入，涵盖 PDF、Word、Excel、PPT、图片及业务网页
• 文档结构恢复：按文件格式提取正文、标题、表格和图片文字，并保留每段内容与原始资料位置的关系
• 可追溯检索：建立可检索的知识节点、来源记录和完整度状态
• 带引用问答：基于证据生成带引用答案
• 受权限限制的 API：受 Token 和知识库范围限制的检索与问答接口，供上层应用调用
• 部署方式：支持 Docker Compose 本地部署

解决的问题

页面描述的目标场景是：企业有价值的数据散落在文件夹、Office 文档、PDF、图片和业务网页中，员工查找困难，且不同 AI 应用需要重复整理同一批资料。Hermes 的定位是让客服机器人、内部知识问答和业务助手复用同一份整理好的企业数据，避免为每个 AI 应用重新清洗和导入。

人工保留的职责

页面明确说明系统并非全自动，以下环节仍需人工介入：

• 管理员决定哪些文件夹和资料进入哪个知识库，检查导入失败、重复文件和资料完整度
• 业务负责人确认检索结果和引用是否足以支持回答，决定哪些知识库可提供给具体应用
• 客服话术、对外回复和业务动作继续由对应应用和人员审核

使用范围与限制

页面明确列出以下边界：

• Hermes 仅提供企业知识整理和检索内核；客服机器人、企业微信或飞书接入、界面和业务动作需由上层应用完成
• 原始资料缺失、过期或扫描质量过低时，系统无法凭空补齐事实，需企业先修正资料来源
• 企业群聊和部门身份不能只依赖模型判断，必须由接入层明确映射到允许访问的知识库

交付物

最终交付包括：

1. 按业务主题整理的企业知识库（含来源、原文位置、完整度和处理记录）
2. 供应用直接调用的证据检索结果，以及带来源引用的知识问答
3. 面向客服机器人、内部知识问答和其他业务助手的检索与问答 API

页面其他信息

页面底部注明“你的业务不需要复制这个项目；我们会从实际流程重新诊断”，并附有“讨论你的流程”的行动号召，表明该项目可能由特定服务商提供定制化实施。页面顶部有“返回项目库”导航链接，说明该页面属于某个项目集合的一部分。

分析说明：本页面为产品介绍页，未提供定价、技术架构细节、性能指标或客户案例。页面中“项目目前支持”的表述暗示功能仍在迭代中。所有功能描述均以页面原文为准，未做超出页面信息的推断。

GEO vs SEO: A Practical Map of the Differences — Page Analysis

Executive Summary

This page, published by SolveReal Systems on September 21, 2026, provides an operational comparison between Generative Engine Optimization (GEO) and traditional Search Engine Optimization (SEO). It argues that while both share technical foundations, they differ in what they optimize, how success is measured, and who can approve completed work. The page is structured as a practical guide with comparison tables, a hypothetical case study, measurement frameworks, prioritization guidance, and a four-week implementation sequence.

Key Claims and Positions

The page makes several specific claims about the current state of GEO and SEO:

On Google's generative features: The article states that Google's AI Overviews and AI Mode are built on core ranking and quality systems, with no separate technical requirements for sites. It references Google's guide to generative AI search and the Generative AI performance report introduced in 2026.

On Bing: It notes that Bing Webmaster Tools introduced an AI Performance public preview reporting citation counts, cited pages, and grounding queries. The page includes Microsoft's warning that these aggregates do not represent rankings or authority.

On OpenAI: The page states that allowing OAI-SearchBot supports inclusion in ChatGPT search, and that referral links include utm_source=chatgpt.com.

On the GEO research paper: It describes the original GEO paper as "a research starting point, not an industry-wide ranking standard" and cautions against converting benchmark results into citation promises for commercial websites.

Content Structure and Quality

The page is well-organized with clear sections: a direct answer, operational differences table, unit-of-optimization discussion, a hypothetical enterprise software example, measurement layers, prioritization matrix, team responsibility division, common mistakes, a four-week sequence, boundaries, and an acceptance checklist.

The writing is clear and practical. The hypothetical example of an invoice-recognition software company effectively illustrates how SEO and GEO acceptance records differ for the same page. The four-layer measurement model (appearance, representation, support, downstream action) is a useful framework.

Notable Limitations and Uncertainties

The page acknowledges several important boundaries:

• Platform differences: "Google's rules cannot stand in for Bing, OpenAI, or another provider's current documentation."
• No guarantees: "SEO and GEO changes do not guarantee indexing, rankings, citations, traffic, or inquiries."
• Report limitations: "Platform reports prove only the fields they cover."
• Sample size caution: "When the sample is small, preserve the observations without presenting one movement as a stable trend."

The sources were checked on September 21, 2026, and the page notes that platform guidance can change.

Potential Concerns

Commercial context: The page is published by SolveReal Systems, which offers GEO-related services. It includes calls to action for an "Open GEO Console," enterprise AI system services, and a contact page. The content, while informative, serves as lead generation for these services.

Source verification: The page references several official sources (Google SEO Starter Guide, Google's AI features documentation, Bing AI Performance, OpenAI's publisher FAQ, the GEO paper) but does not provide direct URLs. Readers cannot verify these references without additional searching.

Date sensitivity: The page is dated September 21, 2026, and references platform features described as current for 2026. Given the rapid evolution of AI search products, some claims may become outdated quickly.

Hypothetical example: The enterprise software case is explicitly stated as not describing a SolveReal Systems customer, which is appropriate transparency.

Overall Assessment

This is a substantive, well-structured guide that offers practical distinctions between SEO and GEO work. It is notably measured in its claims, explicitly warning against over-interpreting citation counts and undefined "GEO scores." The emphasis on separate acceptance records and the need for business-owner approval of public claims is a responsible approach.

The page is best understood as both an educational resource and a marketing piece for SolveReal Systems' services. The educational content appears technically sound and appropriately cautious, but readers should verify platform-specific claims against current official documentation before acting on them.

Rating: 4 out of 5 — Useful practical guidance with appropriate caveats, though the commercial context and lack of direct source links warrant reader caution.

Page Analysis: Codex Feishu Bridge — Enterprise AI Collaboration System

Overview

This page presents Codex Feishu Bridge, a deployable enterprise system that connects OpenAI's Codex to a Feishu (Lark) workspace. The core value proposition is enabling team members to submit AI tasks, files, and follow-up requests from Feishu on desktop or mobile, while Codex executes on a company workstation that remains online.

Key Claims and Capabilities

The system is described as providing:

• Task submission from Feishu — Members can submit questions, attachments, and long-running tasks from computer or phone.
• Asynchronous execution — Work is handed to Codex on an online company workstation; members can leave their desks and continue tasks remotely.
• Trackable task states — The system tracks acceptance, queueing, execution, clarification, and delivery states, returning results to the original Feishu topic.
• Reusable Skills — Repeatedly proven workflows can be packaged into "Skills" for other team members to invoke, reducing dependence on a few experienced Codex operators.
• Authorization controls — Group and member authorization is validated before requests are processed.

Technical Requirements and Limitations

The page states several important constraints:

• Deployment: Windows 10/11 x64 release; the customer workstation must have Codex installed and signed in.
• Dependency on uptime: Tasks cannot continue when the workstation is powered off or the bridge service is offline.
• Feishu permissions: The system depends on the customer's own Feishu application permissions and signed-in Codex environment.
• Not open access: Unauthorized groups or members cannot use the system.

Human Oversight and Boundaries

The page emphasizes that certain functions remain human-controlled:

• Codex asks clarifying questions in the original topic and waits for unambiguous answers from authorized members.
• Members must verify final responses and attachments for completeness and usability.
• Deployment, real external sending, and paid generation are separately authorized actions and are not triggered automatically by ordinary chat.
• New Skills require team confirmation of conditions, execution steps, and acceptance criteria before use.
• The system does not learn from employee behavior without confirmation.

Delivered Outputs

According to the page, outputs include:

• Direct answers in the original Feishu topic, or trackable asynchronous tasks with identifiers.
• Text, images, Office files, videos, and other artifacts generated or revised by Codex.
• Queryable records for accepted, running, waiting-for-input, completed, failed, or uncertain task states.
• Reusable Skills derived from approved, effective workflows.

Context and Positioning

The page frames the problem as: employees repeatedly move requests, files, and results between Feishu and separate AI tools; complex work is often delegated to a few Codex users; and tasks become difficult to continue once people leave the company workstation. The system positions itself as an "enterprise AI work entry point" deployed on the customer's own workstation.

Notable Caveats

• The page does not provide pricing, deployment timeframes, or technical architecture details.
• No customer testimonials, case studies, or performance metrics are presented.
• The claim that "your business does not need to copy this project; diagnosis starts from the actual workflow" suggests a consulting-oriented approach rather than a pure product sale.
• The page does not specify which Feishu plan or API tier is required, nor does it detail security certifications or data handling policies beyond the stated authorization controls.

Overall Assessment

This is a capability and positioning page rather than a detailed technical specification. It clearly communicates the system's purpose, operational boundaries, and human oversight requirements. The main uncertainties for a potential buyer would be: actual performance in production, setup complexity, Feishu API limitations, and the extent of customization needed for specific workflows. The page's emphasis on human confirmation and separate authorization for external actions suggests a cautious, compliance-aware design, but independent verification of these claims would be advisable before procurement.

Page Analysis: "When AI Cites Your Site, Did the GEO Change Work?"

URL: https://me.itheheda.online/en/articles/how-to-measure-geo-ai-citations
Publisher: SolveReal Systems
Date: September 18, 2026
Content type: Educational article on measuring Generative Engine Optimization (GEO) outcomes

Overview

This article provides a framework for measuring whether changes made to a website under a GEO program actually produced meaningful business results. The author argues that a single AI citation is insufficient evidence of success and proposes four separate outcome measurements: answer accuracy, citation support, site visits, and qualified inquiries.

Key Claims and Recommendations

Core Thesis

The central argument is that "a cited page is one observation, not a verdict." The article explicitly warns against merging multiple outcomes into a single "AI visibility" number, arguing that each outcome has a different owner and verification method.

Four Measurement Categories

1. Answer accuracy — Whether the AI's response correctly describes the offer, including necessary conditions and qualifications. The article notes that "a brand mention without those qualifications can be commercially worse than no mention."

2. Citation support — Whether the cited URL actually supports the adjacent claim. The article distinguishes between presence, support, and currency as separate fields, citing the ALCE research (Gao et al.) which evaluates correctness separately from citation support.

3. Site visits — Observable analytics events with referral context. The article cautions against converting visits into "AI-generated revenue" without evidence.

4. Qualified inquiries — Business-defined leads meeting an accepted qualification rule (e.g., named organization, relevant problem, reachable contact, plausible budget). The article states that "a web analytics tool cannot infer it reliably."

Practical Methodology

The article recommends:

• Using a fixed buyer question (with recorded wording, language, product, search mode, date, answer, and citations) for comparability
• Maintaining a claim ledger marking each material statement as accurate, inaccurate, incomplete, or not answered
• Preserving baseline records before changes and retesting without overwriting
• Comparing defined periods before/after changes while accounting for context (campaigns, seasonality, site changes)

Platform-Specific Notes

• Google: AI-feature traffic appears in Search Console's overall Web reporting; there is no separate AI-features report. Web performance is a "useful supporting signal" but cannot isolate AI Overview or AI Mode.
• OpenAI: OAI-SearchBot must be allowed for ChatGPT Search consideration; GPTBot serves a separate training purpose. ChatGPT Search appends utm_source=chatgpt.com to referral URLs, which is "useful referral context, not proof" of commercial causation.
• Bing: AI Performance reports aggregated citation activity and grounding-query phrases across Microsoft AI surfaces. These are aggregated signals, not original prompts, and the product does not report referral traffic or organic ranking.

Limitations and Boundaries Acknowledged

The article is explicit about what its method does not do:

• Does not reveal internal ranking or source weights of AI platforms
• Does not guarantee citation, traffic, or commercial demand
• Does not make live systems deterministic
• Does not provide clean causal attribution (buyers may use AI answers and later return via other channels)
• Ten citations to syndicated copies of one press release are not ten independent confirmations

Product Promotion

The article concludes by promoting GEO Console, described as a tool that "audits official pages against buyer questions and organizes access, content, answer, and citation gaps into prioritized website work." It explicitly states that GEO Console does not provide live citation tracking or conversion tracking and cannot promise platform citations.

Source Citations Used

1. Google: AI features and your website
2. OpenAI: Publishers and Developers FAQ
3. Bing: AI Performance
4. Bing Webmaster Blog: AI Performance public preview
5. Gao et al., "Enabling Large Language Models to Generate Text with Citations"

Assessment Notes

Strengths: The article provides a disciplined, evidence-based approach that resists oversimplified metrics. The separation of outcomes into distinct measurement categories is methodologically sound. The explicit acknowledgment of platform limitations (especially Google's lack of separate AI reporting) is honest and useful.

Caveats: The article is published by a vendor (SolveReal Systems) that sells GEO services, so the framework serves partly as thought leadership for its product. The hypothetical example question about "scanned quotations" is clearly labeled as illustrative. The method requires significant manual effort (claim ledgers, saved answers, source-by-source review) that may be impractical for large-scale operations.

Uncertainty: The article does not provide empirical evidence that its framework produces better business outcomes than simpler approaches. It also does not address how frequently measurements should be taken beyond "an agreed interval."

This analysis is based solely on the supplied page content. External sources referenced within the article were not independently verified.

GEO Report: Page Analysis — "网站已经做了 SEO，为什么还需要 GEO？"

URL analyzed: https://me.itheheda.online/articles/why-geo-after-seo
Observed: September 30, 2026
Source length: ~3,540 characters (Chinese)

1. Executive Summary

This page is a well-structured Chinese-language explainer article published by 实解智能 (Shijie Intelligence) on September 8, 2026. It argues that websites which have already completed traditional SEO should still pursue GEO (Generative Engine Optimization), positioning SEO as the necessary foundation and GEO as the layer that extends visibility into AI-generated answers. The article is practical, cautious in its claims, and clearly written for enterprise website owners who have done SEO but are not yet appearing as sources in AI answers.

Overall assessment: The page is strong on clarity, structure, and practical guidance. It avoids overpromising, cites official sources appropriately, and provides a concrete starting methodology. Its main weaknesses are the absence of original research data, limited internal linking, and a lack of author credentials or company background information.

2. Content Quality Analysis

2.1 Core Argument and Structure

The article opens with a clear thesis: SEO gives a website the foundation to be crawled and indexed; GEO extends that work to ensure brands, services, and evidence can enter generative answers. It then walks through:

1. Why SEO alone is insufficient — users increasingly ask full questions to AI systems, which synthesize answers before users decide whether to click through.
2. What SEO already provides — citing Google's official documentation on AI Overviews and AI Mode, the article correctly notes that technical SEO (robots.txt, stable HTML, internal links, structured data) remains necessary.
3. What GEO adds — three categories of work: clarifying brand and service descriptions, attaching conditions/sources/scope to key claims, and verifying what target platforms actually retrieve and how they use it.
4. Where pages fail — a four-step breakdown: bot access, content retrieval, answerability judgment, and final answer presentation.
5. Who should do GEO now — five signals indicating GEO checks are worthwhile.
6. A minimal starting method — ten real buyer questions, checked against the website, with structured observation records.
7. Limitations and implementation steps — honest about what GEO cannot guarantee.

This structure is logical and easy to follow. The article practices what it preaches: it states conditions, sources, and boundaries around its own claims.

2.2 Source Citations

The article cites four external sources:

• Google AI 搜索功能说明 (Google AI search features documentation)
• Google 生成式搜索优化指南 (Google's generative search optimization guide)
• OpenAI 爬虫说明 (OpenAI crawler documentation)
• GEO 原始论文 (the original 2024 KDD GEO paper)

These are appropriate and credible references. The article correctly notes that the KDD paper "provides a research starting point" and "cannot guarantee citation results on any commercial platform" — an important and accurate caveat.

One limitation: the citations are text links without visible URLs or access dates beyond a general "资料于 2026 年 9 月 8 日核对" (materials verified September 8, 2026). For a page that emphasizes verifiability, inline URLs or a more detailed reference list would strengthen trust.

2.3 Accuracy and Currency

The technical claims are accurate as of the stated verification date:

• Google's AI Overviews and AI Mode do rely on the search index and traditional SEO fundamentals.
• OpenAI does distinguish between OAI-SearchBot, GPTBot, and ChatGPT-User for different purposes.
• The description of the four-step failure path (access → retrieval → answerability → presentation) aligns with how AI answer systems generally operate.

The article is appropriately cautious: it repeatedly notes that different platforms use different indexes and mechanisms, and that Google's rules cannot be directly applied to OpenAI or other systems.

2.4 Originality and Depth

The article is original in its framing — the "four-step drop-off path" and the "three categories of GEO work" are useful conceptual contributions. However, it contains no original research, case studies, or proprietary data. It is a synthesis of existing knowledge presented clearly, not new findings.

3. SEO and Technical Analysis

3.1 On-Page SEO Elements

3.2 Content Readability

The Chinese text is clear and accessible. Sentences are moderately short, technical terms are explained in context, and the article avoids unnecessary jargon. The "直接答案" (direct answer) summary at the top is a good practice for both human readers and AI systems that may extract content.

3.3 Technical Accessibility

The page appears to be a standard HTML article page. The extraction succeeded with "direct_readable" status, suggesting the content is present in the initial HTML rather than requiring JavaScript execution — which aligns with the article's own advice about ensuring content is retrievable.

4. GEO-Specific Analysis

4.1 Answerability

The article is highly answerable. It directly addresses a specific question ("Why GEO after SEO?") and provides:

• A clear direct answer in the opening section
• Structured sub-arguments with headers
• Concrete examples (the procurement manager query about customer service AI)
• Practical checklists

An AI system could easily extract the core argument and key points from this page.

4.2 Entity Clarity

The brand (实解智能) is mentioned, but the page does not clearly state:

• What 实解智能 does (the "企业 AI 系统服务" link suggests they offer enterprise AI system services, but this is not explained in the article body)
• Who the authors are
• Company background or credentials

This is a missed opportunity for entity clarification — one of the three categories of GEO work the article itself recommends.

4.3 Evidence and Conditions

The article does well here:

• Claims are attributed to sources (Google, OpenAI, KDD paper)
• Limitations are stated explicitly ("完成 GEO 整改不保证收录、引用、排名、流量或咨询转化")
• The verification date is provided
• The caveat about platform differences is clear

4.4 Citation Likelihood

The page is a strong candidate for citation in AI answers about:

• The relationship between SEO and GEO
• Whether GEO is necessary after SEO
• How to start GEO work
• Common failure points in AI answer visibility

Its clear structure, direct answers, and practical checklists make it easy for AI systems to extract quotable content.

5. Strengths and Weaknesses

5.1 Strengths

1. Clear, direct answers — the "直接答案" section is exemplary for GEO purposes
2. Honest limitations — explicitly states what GEO cannot guarantee
3. Practical methodology — the "ten questions" approach is actionable
4. Credible sources — cites official documentation from Google and OpenAI
5. Good structure — logical flow from problem to solution to implementation
6. Verification-friendly — encourages saving platform, time, question, answer, and sources for comparison

5.2 Weaknesses

1. No original data — no case studies, metrics, or proprietary research
2. Limited internal linking — only three internal links; related articles are listed but not linked inline
3. Missing entity information — does not explain who 实解智能 is or why readers should trust them
4. No author byline — the article is attributed to the company, not an individual
5. Reference URLs not visible — links are present but the extraction does not show the actual URLs
6. No visual elements — no diagrams, charts, or images to illustrate the four-step path or other concepts

6. Recommendations

6.1 High Priority

1. Add an author or reviewer byline with credentials to build trust and entity clarity.
2. Explain what 实解智能 does within the article body, not just in a link at the bottom.
3. Add visible URLs to the reference section so readers can verify sources without clicking through.

6.2 Medium Priority

4. Add a simple diagram illustrating the four-step drop-off path (access → retrieval → answerability → presentation).
5. Include one concrete example of a before/after GEO adjustment, even if anonymized.
6. Link to the related articles inline where relevant, rather than only listing them at the bottom.

6.3 Lower Priority

7. Add FAQ schema or a structured FAQ section to increase answerability.
8. Consider adding an English version of the article, given the international nature of GEO discussions.
9. Add social sharing or citation metadata to make the article easier to reference.

7. Conclusion

This is a well-crafted, honest, and practical article that effectively answers its title question. It is likely to be cited by AI systems for its clear explanations of the SEO-GEO relationship and its practical starting methodology. The main opportunities for improvement are in entity clarification (who is 实解智能?), visible source URLs, and the addition of original data or examples to increase uniqueness and citation value.

Estimated citation potential: Moderate to high for queries about SEO vs. GEO, GEO necessity, and GEO implementation steps. The article's clear structure and direct answers make it easy for AI systems to extract and cite.

Overall rating: 7.5/10 — strong content, good structure, honest limitations; held back by missing entity information, no original data, and limited internal linking.

GEO Analysis Report: Hermes Notebook

URL analyzed: https://me.itheheda.online/en/projects/hermes-notebook
Observed: September 30, 2026

1. Page Summary

This is a project showcase page for Hermes Notebook, described as an "Enterprise data engine for AI applications." The page presents a system that converts enterprise information scattered across PDFs, Word, Excel, PowerPoint, images, local folders, and business web pages into "citable, retrievable knowledge" that can support customer-service bots, internal Q&A, and other business assistants through APIs.

The page is structured as a product description with sections covering the previous workflow, what the system automates, what remains human-controlled, delivered outputs, and usage scope. It closes with a call to action: "Your business does not need to copy this project; diagnosis starts from the actual workflow. Discuss your workflow."

2. Entity and Concept Extraction

Primary entity: Hermes Notebook (enterprise knowledge engine)

Key concepts and features mentioned:

• Multi-format ingestion (PDF, Word, Excel, PowerPoint, images, local folders, URLs)
• Document structure restoration
• Traceable knowledge nodes and retrieval indexes
• Permission-scoped query and retrieval APIs
• Citation-backed Q&A
• Token-restricted API access
• Knowledge-base scope restrictions
• Local Docker Compose deployment
• Evidence retrieval with source citations
• Completeness states and processing records

Human oversight elements:

• Administrator control over knowledge base content
• Business owner review of evidence and citations
• Manual mapping of enterprise group/department identity to knowledge bases
• Review of customer-service wording and business actions

Limitations disclosed:

• Cannot reconstruct facts from missing, outdated, or poorly scanned sources
• Identity mapping must be explicit, not model-judgment-based
• Customer-service bots, WeCom/Feishu integration, interfaces, and business actions are outside the system's scope

3. Claims and Fact-Checking Assessment

The page makes several verifiable claims about system capabilities. Based on the information provided on this page alone, I cannot independently verify these claims through external sources. The claims are presented as product descriptions rather than as independently audited facts.

Claims presented:

1. The system imports multiple formats and restores document structure — plausible for a document-processing engine, but unverified here.
2. It builds traceable knowledge nodes and retrieval indexes — consistent with the described architecture.
3. It supplies evidence through permission-scoped APIs — a design claim, not externally verifiable.
4. It supports local Docker Compose deployment — a specific technical claim that could be tested but is not demonstrated on this page.
5. Citation-backed Q&A is supported — presented as a feature, not demonstrated with examples.

Uncertainty note: No screenshots, technical documentation links, case studies, or customer testimonials appear on this page. The claims rest entirely on the page's own descriptive text.

4. Source Credibility and Authority Signals

Domain: me.itheheda.online — a subdomain of itheheda.online. The domain does not carry obvious signals of a well-known enterprise software vendor. No company name, legal entity, founding date, team information, or contact details beyond the workflow discussion prompt appear on this page.

Authority signals present:

• Clear technical description of system architecture
• Honest disclosure of limitations and scope boundaries
• Explicit statement that identity mapping cannot rely on model judgment

Authority signals absent:

• No customer logos, testimonials, or case studies
• No technical whitepaper or documentation links
• No version numbers, release dates, or maintenance information
• No security certifications or compliance mentions
• No company information, team credentials, or founding context

Assessment: The page reads as a competent technical product description, but the absence of verifiable third-party validation, company context, or demonstrable deployments limits its authority for a paid decision.

5. Content Gaps and Missing Information

For a prospective buyer evaluating this system, the following information is absent:

• Pricing model — no cost information or licensing terms
• Deployment requirements — hardware, infrastructure, or scaling considerations beyond "Docker Compose"
• Performance benchmarks — no data on ingestion speed, retrieval latency, or accuracy
• Security and compliance — no mention of encryption, data residency, GDPR, SOC 2, or similar
• Integration specifics — no API documentation, SDKs, or example code
• Support and maintenance — no service-level agreements or support channels
• Reference customers — no named deployments or industry examples
• Comparison to alternatives — no positioning against other enterprise knowledge engines
• Implementation timeline — no indication of how long deployment takes
• Team and company background — no information about who built or maintains this system

6. SEO and Content Optimization Observations

Strengths:

• Clear, descriptive headings that map to user intent
• Plain-language explanations of technical concepts
• Honest limitation disclosures that may build trust
• A clear call to action at the end

Weaknesses:

• No structured data markup visible in the retained text
• No FAQ section targeting long-tail queries
• No internal links to related projects or supporting documentation
• No external references or citations to industry standards
• The page appears to be a single project page without surrounding content hub structure

7. Overall Assessment and Recommendation

Verdict: This page provides a coherent, technically plausible description of an enterprise knowledge engine, with honest scope limitations. However, it lacks the third-party validation, company context, pricing, security details, and demonstrable evidence that would typically justify a paid enterprise software decision.

Recommendation for the buyer: Treat this page as an initial introduction, not as sufficient evidence for procurement. Before any paid engagement, request:

1. A live demo or proof-of-concept deployment
2. Technical documentation and API specifications
3. Security and compliance documentation
4. Reference customers or case studies
5. Clear pricing and support terms
6. Company background and team credentials

The page's own closing line — "Your business does not need to copy this project; diagnosis starts from the actual workflow" — suggests a consulting-led approach. That may be appropriate, but the consulting entity itself remains unidentified on this page, which is a material gap for a paid engagement decision.

Page Analysis: "From Search Question to Cited Answer: How AI Uses Web Sources"

Overview

This page, published by SolveReal Systems on September 16, 2026, is an educational article explaining how web-enabled AI products move from a buyer's question through retrieval, synthesis, and citation. It is written for an audience interested in Generative Engine Optimization (GEO) and AI search behavior.

Key Content Summary

The article walks through a hypothetical scenario: a procurement manager asking whether a software provider can process scanned quotations and who reviews the result. It then explains the pipeline:

1. Query fan-out – One question may become multiple searches. The article cites OpenAI's ChatGPT Search rewriting capabilities and Google's query fan-out across subtopics.

2. Page eligibility – Pages must be discovered, crawled, and indexed before they can appear. OpenAI requires publishers to allow OAI-SearchBot for inclusion in summaries; Google requires indexed pages eligible for normal snippets. The article stresses that eligibility does not guarantee placement.

3. Passage selection and synthesis – The article introduces Retrieval-Augmented Generation (RAG) as an architectural model, citing the original Lewis et al. paper. It warns that if only one of several relevant passages is selected, the answer may overstate capabilities.

4. Citation verification – The article distinguishes between visible citations and complete source logs, referencing Gao et al.'s ALCE framework, which separates citation correctness from citation completeness.

5. GEO role of official sites – The official company website is described as the "accountable reference" for identity, scope, and conditions. The article notes that repetition does not make a repost independent evidence.

Practical Guidance Provided

The page includes a verification checklist with six questions, covering:

• Whether web search was enabled for a recorded answer
• Whether the target page is eligible for the candidate set
• Whether selected material contains capability and conditions together
• Whether each citation supports its adjacent claim
• Whether first-party facts and independent evidence are recorded separately
• Whether a person reviewed claims affecting public or commercial decisions

It also offers practical steps: preserve the buyer question's wording, time, product mode, answer, and cited URLs; read passages against claims; and retest after reviewed changes.

Scope and Limitations Acknowledged

The article explicitly states it cannot reveal internal weights, predict citation rates, or describe exact queries or ranking scores. It notes that products differ in search, indexes, passage selection, and citations. It also cautions that fluent prose is not proof that a condition remains current.

Sources Cited

The article references six sources:

• OpenAI: Searching the web with ChatGPT
• OpenAI: Publishers and Developers FAQ
• Google: AI features and your website
• Google: How Google Search works
• Lewis et al., Retrieval-Augmented Generation for Knowledge-Intensive NLP Tasks
• Gao et al., Enabling Large Language Models to Generate Text with Citations

Related Articles

Three related articles are listed, all dated September 2026:

• "When AI Cites Your Site, Did the GEO Change Work?" (Sep 18)
• "In GEO, How Much Do Websites, News, and Social Posts Count?" (Sep 16)
• "GEO vs SEO: A Practical Map of the Differences" (Sep 21)

Assessment Notes

Strengths:

• Clear, structured explanation of a complex pipeline
• Explicit separation of distinct concepts (eligibility, selection, citation, claim)
• Honest acknowledgment of uncertainty and undocumented internal mechanisms
• Practical, actionable checklist for verification
• Well-sourced with primary references

Limitations:

• The article is educational rather than empirical; it does not present original research or data
• It relies on public documentation from OpenAI and Google, which may change
• The hypothetical scenario is illustrative, not a real observed case
• No information about the author's credentials or the publication's editorial standards is provided on this page

Potential concerns:

• The page promotes SolveReal Systems' Open GEO Console product, which may introduce commercial bias
• The "Sources" section lists titles but does not include full URLs or access dates on this page
• The article's date (September 16, 2026) is in the future relative to typical knowledge, so the cited OpenAI and Google documentation may have evolved

Recommendation for Use

This page is useful as a conceptual framework for understanding AI search and GEO verification. It is not a substitute for checking current official documentation from OpenAI or Google, nor does it provide empirical evidence about citation behavior. For material decisions, readers should follow the article's own advice: inspect original records, dates, and authorities directly.

How Language Models Read Web Pages: What It Means for GEO Optimization

Publisher: 实解智能 (Shijie Intelligence) | Date: September 4, 2026

This article explains, with accessible examples, how large language models process web page text—from tokenization and context understanding to semantic retrieval—and what these principles mean in practice for GEO (Generative Engine Optimization) content work.

Direct Answer

A company website may describe its products thoroughly, yet AI answers still fail to mention it. The cause could be that the page was never retrieved, or that the content retrieved was insufficient to support an answer. To decide what to fix, it helps to first understand how a language model processes text.

The article draws on Transformer and retrieval-augmented generation papers, as well as public documentation from Google and OpenAI. It discusses only what is relevant to website content, does not expand formulas, and does not treat model principles as search ranking rules.

GEO stands for Generative Engine Optimization. For a company website, the goal can be understood as making public information easier for AI search systems to discover and use accurately in relevant answers. Two connected processes are involved: how the model understands input, and how the search system decides which content to hand to the model. A website must address both.

What Happens When Text Enters a Model

When a model processes text, it typically first passes through a tokenizer, which converts text into small units called tokens and maps them to numeric IDs. A token may correspond to a word or only part of one; the exact segmentation depends on the tokenizer. It is not correct to assume a Chinese character or an English word always equals one token. Hugging Face tokenizer documentation

The model then converts these IDs into numerical representations and processes context through multiple layers. The Transformer's attention mechanism combines information from different positions according to computed weights, allowing one position's representation to incorporate content from other positions. This is a technical foundation for understanding word relationships. Transformer original paper

A teaching example illustrates this relationship:

The document assistant supports uploading PDFs, but scanned files require OCR first.

To understand this sentence accurately, the relationship between "scanned files" and "OCR first" must be preserved. If only "supports uploading PDFs" remains, the applicable condition is lost.

For website editors, the practical task is to place limitations next to the relevant feature descriptions. Product names should be explicit, and pronouns should make clear what they refer to. These are editorial suggestions for reducing ambiguity; they cannot be used to claim that a particular sentence structure will receive higher model attention or reliably increase citations.

Tokens are also not the same as web page paragraphs. The former is a model input unit; paragraphs are how authors organize content; and the retrieval snippets used by search systems are yet another level. There is no fixed one-to-one correspondence among them.

What an AI Search Engine Actually Receives May Be Only Part of a Page

Beyond input processing, one step further matters: what does the answering model actually receive?

Retrieval-augmented generation, usually abbreviated as RAG, combines information retrieved from external sources with a generative model, allowing answers to use material from current inputs. Early RAG papers studied this combination and used retrieved text passages to assist generation. RAG original paper

The following simplified flow is for understanding; it describes a common approach and does not mean all commercial platforms run the same architecture:

Web page → acquisition and processing → retrievable material User question + retrievable material → retrieve relevant content → hand to model → generate answer

Google publicly states that AI Overviews and AI Mode may expand a question into multiple related queries and then look for web pages that support an answer. Material for a single question may come from multiple pages. Google AI search feature documentation

This means writing should consider partial reading. If a feature description is extracted on its own, does the reader still know which product it refers to? Does a result only hold under specific test conditions? Is there nearby evidence to verify it?

A simple editorial check: copy a key paragraph into a blank document, ignore the preceding brand introduction, and read whether it can still be accurately restated. If you must search back and forth for what "it," "this solution," or "the above capability" refers to, add the necessary names and conditions back in.

There is no need to chop the entire article into many short blocks. Google's generative search guidance explicitly states there is no fixed ideal length and no requirement to fragment content for AI. Similarly, llms.txt and dedicated Schema are not required conditions for Google AI search. These conclusions apply to Google and cannot be generalized to every system. Google generative search optimization guide

Similar Meaning Matters More Than Repeating the Same Word

Retrieval systems can compare texts using semantic vectors. This sounds complex but can be understood as converting a passage into a set of numbers used to calculate similarity with other texts. Sentence-BERT demonstrates sentence-vector-based semantic similarity comparison. Sentence-BERT paper

This is a technical capability; it does not mean we know the retrieval implementation of every AI search platform. Commercial systems may combine other search and ranking methods.

For website writing, a more useful check follows from this: when users ask in their own words, does your page contain information that can answer the question?

The following two passages are fictional product copy written for illustration; they do not correspond to 实解智能 or any real client project.

This tool has powerful document processing capabilities, suitable for many business scenarios.

A more specific version could be written as:

The example document assistant is for procurement staff who need to organize supplier quotes. After users upload a PDF with a text layer, it can extract product names, quantities, and prices, and export them to a table. Scanned files require OCR first; exported results still need procurement staff to verify, and the tool will not send purchase orders on the user's behalf.

The second passage helps readers judge whether the tool suits quote processing, how to prepare scanned files, and which work remains human. It adds information usable for decisions. Whether the rewrite is more easily retrieved and cited by a particular platform requires actual observation; it cannot be inferred from semantic vector principles alone.

Longer Content Does Not Replace Clearer Evidence

Models have an upper limit on input length per processing run, usually called the context window. Allowing more text does not mean every piece of information will be used equally reliably.

The 2023 paper Lost in the Middle found, in the models and tasks tested at the time, that the position of relevant information in the input affected performance, with information in the middle of long contexts being particularly poorly utilized. It provides a specific research result; it cannot be used to assert that all models today behave identically, nor to conclude that "putting the brand at the beginning will get it recommended." Long-context research paper

For editorial work, the safer approach is to place conclusions and applicable conditions together. When presenting test results, state the sample and test time nearby, and provide sources that support the result. If a limitation would change a buyer's decision, do not leave it in a distant footnote.

Citations also need verification. A link's presence only shows a source was provided; the reader must still be able to find, in that source, content supporting the statement. Describing a model's general capability cannot prove a specific product achieves the same result.

This step can be done manually. Read the most important factual sentences one by one, open the nearby source, and judge how much the source actually supports. If the scope is overstated, pull it back; if evidence is missing, add it or delete the claim.

Scope and Limitations

This article is aimed at content editing and technical checks for public websites. The following steps are for identifying problems; they cannot guarantee indexing, citation, or business conversion.

Implementation Steps

First make the text obtainable, then observe how answers use it.

The editorial work above has a precondition: the target system can access the body text. Google's JavaScript documentation notes that some pages only show actual content after scripts execute, and not all crawlers can run JavaScript. Having key explanations readable in the initial HTML helps reduce this access dependency. Google JavaScript documentation

In practice, compare the page response with what appears in a browser. Beyond titles and navigation, check whether key limitations in the body text and tables are preserved. A browser displaying the full page alone does not prove a given target crawler obtained the same information.

Access settings should also be distinguished by purpose. OpenAI separates OAI-SearchBot, used for search, from GPTBot, which may be used for training data collection. Allowing search access does not require allowing training collection; and allowing access does not guarantee citation. OpenAI crawler documentation

Acceptance Checklist

Use a set of real buyer questions to observe the effect of changes. Before testing, fix the question list; record the platform, language, and whether web search is enabled; and save the test time, the original answer text, and the URLs actually cited. Check Chinese and English separately; do not substitute Chinese answers for English acceptance testing.

Record three outcomes separately: whether the website appears in sources, whether the answer describes it accurately, and whether users continue to visit the website. A first mention after changes counts as one observation; repeated checks under similar conditions are needed before discussing whether the change is stable.

References

Technical papers and official documentation are linked inline at the relevant sections. Materials were verified on September 4, 2026. Platform rules may change; follow the current instructions at the linked URLs.

Next Steps

If you are starting from an existing website, you can use Open GEO Console to organize issues related to access, public content, and citation evidence, then let the website owner determine the order of fixes. Starting with one key service page and filling in the facts it lacks to answer buyer questions will make results easier to see than changing the entire site at once.

You can also review the Open GEO project description, learn about enterprise AI system services, or use the contact page to describe the website and buyer questions you need checked.

Related Articles

AI Search and GEO — Why Do You Need GEO If You Already Have SEO? SEO gives a website the foundation to be found and indexed; GEO goes further to check whether brand, services, and evidence can enter generative answers. This article explains how the two work together and where a company website should start. 实解智能 / September 8, 2026 / 7-minute read

AI Search Visibility — Why Your Company Website Matters More in the AI Era AI answers often encounter a company's public web pages first, then decide whether to understand, cite, or drive visits. Following the path of discovery, crawling, understanding, retrieval, and citation, this article explains why the company website is becoming a maintainable, verifiable formal information source. 实解智能 / September 2, 2026 / 12-minute read

AI Search Visibility — Why Does AI Always Recommend Competitors, Not My Company? When the same procurement question lists competitors but not your company, fix the test conditions first, then compare answer reasoning, cited pages, service evidence, and crawl status to find the one place worth changing. 实解智能 / August 31, 2026 / 9-minute read

Freight Lead Agent — Project Page Analysis

Overview

This page presents Freight Lead Agent, a deployable system for B2B sales prospecting that combines Google Maps company discovery with website analysis and personalized outreach. The project is positioned as a complete pipeline covering everything from initial company discovery through reply follow-up.

What the System Does

According to the page, the system performs four main functions:

1. Company discovery — Collects target companies from Google Maps by keyword and region, preserving source and official-site information
2. Website analysis — Visits each official website to extract business descriptions, service information, and public contacts
3. Personalized message generation — Uses the prospect's website context and sender profile to create a tailored outreach message
4. Delivery management — Provides send plans with pacing, sender account, delivery status, and reply follow-up

Human Involvement

The page emphasizes that human review remains central to the workflow. Specifically:

• Humans review the target company, website context, public contacts, and personalized message for accuracy
• A human decides whether to edit, send, schedule, or stop contact
• Humans handle subsequent replies

Delivered Outputs

Three outputs are listed:

• A prospect library containing source, official website, public contacts, and processing status
• A reviewable personalized outreach message generated from each company's website context
• Send plans with pacing, sender account, delivery status, and reply follow-up

Usage Scope and Limitations

The page includes two explicit limitations:

• The system processes public business information only; it does not crawl authenticated pages or personal profiles
• Outreach content requires human review; the system does not provide indiscriminate automated bulk sending

Technical Implementation

The project is described as a self-contained Docker system covering the full pipeline: Google Maps company discovery, official-site analysis, public-contact extraction, personalized message generation, human review, send plans, and reply monitoring.

Problem Statement

The page frames the problem as a scaling issue: sales teams must find target companies on Google Maps, verify each official website, understand the business, locate public contacts, and prepare a different message for every company — which is difficult to do manually at scale.

Claimed Benefit

The stated benefit is that customers can use a small team to sustain company discovery, website research, and personalized writing while spending less time starting from zero for every prospect and avoiding sending one generic message to everyone.

Important Caveat

The page includes a notable disclaimer at the end: "Your business does not need to copy this project; diagnosis starts from the actual workflow." This suggests the page is not merely a product pitch but positions the project as a reference example, with the actual service being workflow diagnosis rather than replication of this specific system.

Assessment Notes

• The page is concise and clearly structured, with a logical flow from problem to solution to outputs
• The human-in-the-loop emphasis is a meaningful differentiator, particularly given the stated limitation against automated bulk sending
• The scope limitation to public business information is clearly stated
• The final disclaimer is unusual and worth noting — it reframes the project as illustrative rather than prescriptive
• No pricing, timeline, or technical requirements beyond Docker are mentioned
• No customer testimonials, case studies, or performance metrics are provided

This page functions as a project overview rather than a detailed technical specification. For a paid GEO report, the key facts are the system's capabilities, its human-review requirement, its public-data-only scope, and the explicit statement that the project need not be copied — diagnosis starts from the actual workflow.

In GEO, How Much Do Websites, News, and Social Posts Count?

SolveReal Systems | September 16, 2026

This page addresses a common question in Generative Engine Optimization (GEO): how much "weight" different source types—company websites, news outlets, and social posts—carry in AI-generated answers. The article argues that there is no fixed media mix and that source influence depends on the specific claim, the question asked, and the AI product used.

Key Points

The Short Answer

• Public research does not establish fixed percentages for how much AI answers weigh different source types.
• A source can occupy at least three positions: retrieved as candidate material, used while forming an answer, or shown as a citation. These events overlap but are not interchangeable.
• The original retrieval-augmented generation paper (Lewis et al.) distinguishes generating text while conditioning on retrieved documents versus relying only on training parameters—but this is a conceptual model, not a blueprint for every consumer AI product.

Three Layers That Are Easy to Confuse

The article uses a buyer's question as an example: "Which parts of an AI workflow can this provider handle, and who approves external sends?"

• Company website: Normally the right source for claims about what a company currently offers and stands behind. A dated, named service page with a clear owner is more useful than a broad promise.
• Independent report: Can corroborate that a product launch happened or describe an external event, but cannot replace the provider's current service terms. A four-year-old press story may be excellent evidence about history and poor evidence about today's service boundary.
• Practitioner or creator posts: Can add operational texture but are weak evidence for formal commitments unless the author has direct authority and the statement is attributable. Some posts may be inaccessible to certain products or regions.

The article stresses that channel labels ("news," "social," "official site") describe where a page lives, not whether the claim is current, primary, attributable, independent, or entailed by the words on the page.

Retrieval, Use, and Citation Are Different Observations

• Availability only tells you a page could be considered; a retrieved passage may never affect the wording.
• A citation tells you which link was displayed, not the source's internal weight.
• The strongest check: open the cited page and see whether it supports the particular sentence beside the link.
• The ALCE benchmark separates correctness from completeness—a response can cite something accurately while leaving other important statements unsupported.

What Recent Studies Do and Do Not Show

The article cites three studies:

1. The original GEO paper (Aggarwal et al.): Made answer visibility an empirical research problem but did not disclose platform-internal channel percentages.

2. A 2026 preprint (Zhang et al.): Analyzed 602 controlled prompts and 21,143 valid search-layer citations. Displayed citation compositions were:

   • ChatGPT: Official 34.22%, news 31.17%, vertical 22.13%
   • Google AI Overview: Official 46.35%, news 18.99%, vertical 22.00%
   • Perplexity: Official 44.07%, news 16.07%, vertical 18.99%

   The article emphasizes these are sample compositions, not internal source weights. "Official" includes several kinds of institutional pages, not only company websites. The study's "absorption score" is a proxy, not a causal trace of model attention. The sample is largely US and English oriented.

3. Another 2026 preprint (Zhen et al.): Asked 614 Chinese-language questions three times across web and app interfaces of Doubao, DeepSeek, Tencent Yuanbao, and Qwen. Source sets differed even between interfaces of the same product. The observed outputs cannot set a rule for every enterprise query or disclose a system's ranking function.

Conclusion from studies: Source mix varies with question, language, interface, and product. The research does not support putting a permanent "40% official site" label on GEO.

The Official Site's Role: Accountability

• The company website can hold the organization accountable for a current statement, even though no AI system is obliged to prefer it.
• A good official page can name the service, define what it will and will not do, identify human decision points, and link to governing evidence.
• The site serves as the anchor for first-party facts, giving outside coverage something precise to confirm, dispute, or contextualize.
• However, a vague page can be less helpful than a careful independent explanation, and an official site cannot establish independent praise, market adoption, or customer outcomes by assertion alone.

Repetition Is Not Corroboration

One press release can be republished by ten outlets; creators and public accounts may quote it further. A citation sampler could show many domains while the underlying evidence comes from one organization. For each important claim, trace back to origin:

• Who first made it?
• Who has authority to update or correct it?
• Does the later page add independently reported facts, or only repeat earlier text?
• Is the cited passage actually sufficient for the sentence it is attached to?

A Useful GEO Review Starts with One Answer

The article recommends a practical method:

1. Choose one buyer question with a real decision behind it.
2. Write down what a careful answer would need before asking an AI product.
3. Ask the target AI system the exact question; preserve the date, answer, citations, and product interface.
4. Read each cited passage; mark which claims were answered, supported, merely asserted, and which sources were repeated.
5. Repeat the same questions after a website update to see whether the answer reflects the current statement.

The article notes OpenAI's search help warns that cited material can be incomplete, outdated, or incorrect.

Where Open GEO Fits

Open GEO Console is described as a tool that helps turn a GEO review into website decisions. It checks a site's technical foundation, public content, question coverage, and citation evidence, organizing findings into an issue list, page recommendations, and remediation priorities.

Limitations acknowledged: It does not measure hidden source percentages, prove that an uncited page was used, or promise that an AI product will cite the site.

Sources Cited

• Lewis et al., Retrieval-Augmented Generation for Knowledge-Intensive NLP Tasks
• Aggarwal et al., GEO: Generative Engine Optimization
• Zhang et al., From Citation Selection to Citation Absorption
• Zhen et al., What Do Chinese-Language Generative Search Engines Cite and Surface?
• Gao et al., Enabling Large Language Models to Generate Text with Citations
• OpenAI, Searching the web with ChatGPT
• Google, AI features and your website

Related Articles

• "When AI Cites Your Site, Did the GEO Change Work?" (Sep 18, 2026)
• "From Search Question to Cited Answer: How AI Uses Web Sources" (Sep 16, 2026)
• "GEO vs SEO: A Practical Map of the Differences" (Sep 21, 2026)

Assessment: This is a well-structured, evidence-based article that directly addresses a common GEO misconception. It provides specific study data while carefully noting limitations. The practical methodology for conducting a GEO review is actionable. The article is transparent about what public research can and cannot establish, and it does not overstate the capabilities of the author's own tool (Open GEO Console). The main limitation is that the cited studies' data points are presented as sample compositions, and the article itself emphasizes they should not be generalized into universal rules.

Page Analysis: "Google 已收录，AI 搜索为什么仍认不出英文品牌名？双语官网检查与修复清单"

Page Overview

This is a Chinese-language article published by 实解智能 (SolveReal Systems) on September 1, 2026, addressing a specific GEO (Generative Engine Optimization) problem: a bilingual corporate website whose English brand name has been indexed by Google but is still not reliably recognized as a brand entity by AI-powered search engines. The article uses the company's own public case as a worked example.

Core Subject Matter

The page documents a real diagnostic situation:

• The company: 实解智能, with English brand name SolveReal Systems, providing enterprise AI system design and delivery services.
• The problem: Both the Chinese homepage (/) and English homepage (/en) show "URL is indexed on Google" in Search Console as of September 1, 2026. However, when searching the English brand name, AI-generated answers (Google AI Overview, ChatGPT Search, Perplexity) may interpret "SolveReal Systems" as ordinary words related to mathematical problem-solving rather than as a brand entity.
• The stated cause: Indexing alone does not mean the search system has connected the English name, Chinese name, official domain, and enterprise services into one public entity. AI systems, lacking clear relational signals, default to literal word-level interpretation.

Key Technical Claims and Recommendations

The article makes several specific, verifiable claims about how the three AI search products operate:

Google

• AI Overview and AI Mode operate within Google's existing search infrastructure.
• Pages must first be indexed and qualify for ordinary search snippets before appearing in generative features.
• No special AI markup is needed; mechanically repeating long-tail keyword variants does not increase relevance.
• Google's official documentation is cited for: search workings, Organization structured data, site name guidelines, multilingual site guidelines, generative AI search explanations, and re-crawl request procedures.

ChatGPT Search (OpenAI)

• OpenAI advises publishers not to block OAI-SearchBot if they want content to appear in ChatGPT Search summaries, citations, and links.
• Website hosts or CDNs must allow OAI-SearchBot's public IP ranges.
• OAI-SearchBot (search discovery) and GPTBot (potential training use) are distinct and should be controlled separately.
• ChatGPT Search referral links carry utm_source=chatgpt.com, which can be isolated in analytics.

Perplexity

• Official documentation distinguishes PerplexityBot (discovery and linking of public pages; manageable via robots.txt) from Perplexity-User (visits pages during user queries; generally not subject to robots.txt restrictions).
• When using a WAF, Perplexity recommends verifying both User-Agent and official IP ranges; matching only a User-Agent string can admit disguised requests, and IP-only configurations become stale after official range updates.

Recommended Fixes (as presented)

1. Write both names in the same visible paragraph on each homepage — e.g., Chinese page states "实解智能，英文品牌名 SolveReal Systems"; English page states "SolveReal Systems, also known as 实解智能 in Chinese." Both pages should use identical official domain, service scope, and contact paths.
2. Unify public identity configuration — article authors, page creators, Organization structured data, and site name should not each use different name variants. Google's Organization structured data documentation lists name, alternateName, and url as relevant properties; Google also advises site name consistency with homepage visible usage. Structured data expresses existing facts; it cannot resolve contradictions in body text.
3. Use separate stable URLs per language — Chinese at /, English at /en, with hreflang annotations and visible language signals. English article detail pages should only contain reviewed full English text; wrapping Chinese content in an English summary on an English URL reintroduces ambiguity.
4. Request re-crawl via Search Console — only for live, checked URLs; Google's official waiting range is days to weeks; repeated submissions do not speed crawling. The article notes its own requests returned "Requested indexing" receipts, which prove acceptance but not that brand relationships have updated or AI answers have changed.

Evidence Status and Honest Limitations

The article is notably disciplined about what it does and does not claim:

• No Before/After success screenshot is published. Re-crawl requests were accepted; same-condition AI answer re-testing had not been completed at publication time.
• A status table separates: (a) initial check (indexed, brand query interpreted literally), (b) page modifications live, (c) re-crawl requests accepted, (d) follow-up re-test pending — with explicit statements of what each stage does not prove.
• The article explicitly states the method cannot control model answers or guarantee knowledge panels, AI citations, recommendations, rankings, traffic, or inquiries.
• Legal entity names, if different from brand names, should be separately identified and not stuffed into alternateName for keyword purposes.

Measurement Guidance

The article distinguishes five separate events that should not be conflated:

1. Indexing — Search Console.
2. Brand recognition — whether brand queries point to the official site.
3. AI citation — whether answers reference the site.
4. Visits — Search Console for Google; utm_source=chatgpt.com referrals for ChatGPT Search; separately saved Perplexity referrals.
5. Inquiries — forms or CRM.

A citation without a visit may mean the answer satisfied the user; a visit without an inquiry does not retroactively invalidate the article's value. Teams must decide which step a given article is meant to serve, then examine the corresponding data.

Verification Checklist (as given)

• Chinese and English homepages both directly accessible from official URLs.
• Both homepages' visible text states the relationship between Chinese and English names.
• Page titles, authors, site name, and Organization data use reviewed names.
• Chinese and English pages have separate canonicals and mutual language annotations.
• Search Console records indexing status and re-crawl receipts for both URLs.
• Re-test preserves original query, product, date, language, region, answer, and cited URLs.
• Crawler access for Google AI Overview, ChatGPT Search, and Perplexity checked separately.
• Indexing, brand recognition, AI citation, visits, and inquiries are not merged into one metric.
• Legal entity, brand, product, and project names are not substituted for one another as keywords.
• The report promises no knowledge panel, recommendation, ranking, traffic, or inquiry outcomes.

Applicability Boundaries

The method is stated to suit: corporate websites with Chinese and English names, brand renames, abbreviation launches, and coexisting regional names. Public names must be confirmed by the enterprise. Visible text, structured data, and multilingual URLs can reduce ambiguity and improve the chance of consistent facts reaching search systems — but cannot control model responses.

Sources Referenced

The article cites official documentation from: Google (search workings, Organization structured data, site name, multilingual sites, generative AI search, re-crawl requests), OpenAI (publisher/developer guidance), and Perplexity (crawler documentation). Specific URLs are not listed inline; the page presents them as named references.

Promotional Elements

The page promotes Open GEO Console, described as a free check covering public identity, robots.txt, sitemap.xml, visible body text, and buyer-question coverage. It explicitly states the free check gives verifiable gaps and does not promise platform adoption of site content. Links are also provided to the company's Open GEO Console project description, enterprise AI services, and a contact/submission path.

Related Articles (internal links)

Three related Chinese-language articles are listed, all under the "AI 搜索与 GEO" category:

1. "官网自己说法不一时，GEO 应该先改哪一页？" (September 21, 2026, 9 min) — about establishing traceable public fact records when homepages, service pages, old articles, and English versions state different conditions for the same capability.
2. "GEO 和 SEO 到底有什么区别？一张可执行的工作地图" (September 21, 2026, 12 min) — a task/indicator/case comparison of SEO and GEO.
3. "官网被 AI 引用以后，怎么判断 GEO 改动有没有用？" (September 18, 2026, 6 min) — about separating answer accuracy, citation support, visits, and inquiries when evaluating GEO changes.

Overall Assessment

This page functions as a transparent, methodical case study rather than a promotional success story. Its principal value lies in:

• Clearly separating indexing from entity recognition and from AI answer behavior.
• Providing product-specific crawler guidance (OAI-SearchBot vs. GPTBot; PerplexityBot vs. Perplexity-User) with appropriate WAF caveats.
• Insisting on visible-text name relationships rather than relying on structured data alone.
• Explicitly refusing to claim results before same-condition re-testing is complete.

The article's main limitation is that it presents a diagnosis and a remediation plan without completed outcome data — a limitation the page itself acknowledges and frames as methodologically necessary. For a reader evaluating whether to follow this approach, the page offers a concrete, auditable procedure with clearly stated boundaries on what can and cannot be guaranteed.

Open GEO Console — 页面分析报告

页面概况

本页为 Open GEO Console 的产品介绍与落地页，定位为“企业官网 GEO 诊断与 AI 搜索可见性整改”的公开产品入口。页面同时承担产品说明、使用流程展示、报告样例预览和转化引导四项功能。

核心内容与价值主张

页面围绕一个核心痛点展开：企业难以判断 AI 搜索引擎能否读懂其官网、在回答买家问题时引用了谁，以及为何自己的品牌被遗漏。产品将技术检查、公共搜索证据、引用缺口和整改优先级整合为同一条诊断流程，最终交付“一套可以直接进入整改的诊断结果”。

交付物明确列为四项：证据问题清单、页面修改建议、整改优先级、持续监测方向。

使用流程（四步）

1. 输入网站 — 提供企业官网、重点页面及希望覆盖的买家问题
2. 执行检查 — 核对技术基础、公开内容、问题覆盖与引用证据
3. 标出问题 — 将访问、语义、答案和引用缺口汇总为同一份证据清单
4. 形成整改 — 按影响与实施顺序整理修改建议和后续监测方向

报告样例与结构说明

页面内嵌一份真实产物样例（中文深度报告预览），检测对象为“实解智能官网”，生成日期为 2026-09-04。报告结构包含 6 个完整章节：

• 00 结果摘要
• 01 逐题回答与诊断
• 02 网站与页面分析
• 03 行动计划
• 04 内容示例
• 05 方法与完整证据

样例中展示了“网站判断”的节选内容，对实解智能官网给出了具体诊断结论：网站内容清晰、直接回答了买家关键问题、案例与文章丰富，但缺少结构化数据等技术层面的信息，可能影响 AI 爬虫的深度理解。

重要限制与声明

页面明确标注了以下边界条件：

• 报告保留生成当日的公开页面证据，用于展示当前报告结构与交付深度；它不是对网站或 AI 平台状态的实时监测
• 人工判断仍然保留：AI 平台、公开页面和查询结果会持续变化，诊断仅用于确定问题与整改优先级，关键结论、发布修改和后续投入仍由网站负责人复核决定

转化引导

页面提供两个主要行动入口：“进入 Open GEO 正式产品”和“查看完整中文深度报告”，另有“阅读 AI 可见性审计方法”作为辅助链接。页面底部明确提示“Open GEO Console 已开放使用，可直接进入网站开始体验”。

页面可信度评估

从付费 GEO 报告的角度看，本页具备以下特征：

优势：交付物定义清晰、流程透明、报告结构完整，且主动披露了报告的时间局限性和人工复核必要性，降低了过度承诺的风险。

需注意：样例报告仅展示节选内容，完整报告的实际深度和质量需通过“查看完整中文深度报告”进一步验证；页面未披露诊断所覆盖的具体 AI 平台范围、检查的技术指标清单，以及报告更新的频率机制。这些信息在付费决策前值得向供应商进一步确认。

总体判断：该页面适合作为 GEO 诊断服务的初步了解入口，但付费前建议索取完整报告样例，并明确询问技术检查的具体项目、AI 平台覆盖范围及报告时效性保障。

Analysis: "When Does an AI Model Read an Updated Web Page? Training, Retrieval, and GEO"

Publisher: SolveReal Systems
Date: September 16, 2026
URL: https://me.itheheda.online/en/articles/llm-training-text-chunks-geo-retrieval

Summary

This article explains the difference between how AI models encounter web content during training versus how AI search systems retrieve pages when answering questions. It clarifies that updating a page today does not guarantee an AI assistant will know about it tomorrow, and distinguishes between three separate concepts often conflated as "chunks": tokens, training sequences, and retrieved passages.

Key Points

Training vs. Retrieval Distinction

• A model may learn patterns from training data; an answer system may also retrieve a page while answering a question. These are separate events.
• Page updates are generally outside the site owner's control regarding training runs. GEO work is more concrete when following the public page through access, retrieval, and answer use.

Tokenization and Training

• Tokenizers turn text into token identifiers; a token may be a word or part of one. Chinese characters do not reliably equal one token.
• The GPT-3 paper describes autoregressive next-token prediction and notes short documents were packed into training sequences for efficiency. The paper describes one model and one setup; its sequence length is not a rule for every model today.
• Even if a page appeared in training data, it would not become a live, individually editable record. Outsiders generally cannot verify page-by-page what a commercial model's training set contained.

Three Different "Chunks"

• Token: input unit for the model
• Training sequence: span of tokens processed under a particular training setup
• Retrieved passage: material selected from an external source to help with a question
• There is no universal word count converting one into another. The original RAG paper combines a pretrained model with retrieved external passages without retraining on the page first.

How a Page Reaches an Answer

• Google says AI Overviews and AI Mode may issue several related searches and find supporting pages. Eligibility requires the page to be indexed and eligible for snippet display in ordinary Search—eligibility does not promise display.
• After an edit, the relevant system must fetch the page again and update searchable material. Google says recrawling can take days to months.
• Some pages need JavaScript execution before main text appears; other bots may not have the same rendering ability.

Boundaries

• The article covers text models and public web pages. GPT-3 and RAG papers are documented technical examples; Google's guidance describes its own features. None reveals another platform's passage splitting or citation rules, or promises indexing, citations, or inquiries.

Practical Steps

1. Pick one important service page; inspect its initial response for main explanation, audience, and important limits.
2. Check crawl and index status.
3. Ask a real buyer question in the AI product you care about.
4. State one capability with its conditions and supporting evidence nearby; use headings and source links.
5. Google says its AI Search features require no special AI file or dedicated schema markup.

Checking Results

• Save the question, time, answer, and cited URLs. Record access, answer use, and visible citation separately.
• If an answer overstates the service, check the service page's wording. For project-specific claims, look for public evidence on project pages. If no URL appears, record that observation without inferring what was in training data.

Sources Cited

• GPT-3 paper
• Hugging Face tokenizer documentation
• Original retrieval-augmented generation paper
• Google AI Search guidance
• Google JavaScript Search guide
• Google recrawling guidance
• Google site-owner guidance

Related Articles (from same publisher)

• "GEO vs SEO: A Practical Map of the Differences" (Sep 21, 2026)
• "When AI Cites Your Site, Did the GEO Change Work?" (Sep 18, 2026)
• "From Search Question to Cited Answer: How AI Uses Web Sources" (Sep 16, 2026)

Assessment Notes

Strengths: The article provides a clear, well-sourced explanation of the distinction between training and retrieval. It appropriately hedges claims, cites specific papers and Google documentation, and avoids overpromising. The practical steps are concrete and actionable.

Limitations: The article does not provide platform-specific passage-splitting or citation rules beyond Google. It explicitly acknowledges this limitation. The "short answer" section is somewhat indirect; a more direct statement of the core finding would improve scannability.

Uncertainty: The article appropriately notes that recrawling timelines vary (days to months), that eligibility does not guarantee display, and that commercial training data cannot be verified externally. These uncertainties are preserved accurately.

Overall: This is a technically sound, well-referenced guide suitable for website owners seeking to understand GEO mechanics. It avoids hype and clearly separates editorial judgment from citation-ranking claims.

GEO 页面分析报告：官网已被 Google 收录，为什么 AI 还是不推荐你？

分析对象：https://me.itheheda.online/articles/google-indexed-ai-does-not-recommend-website
页面标题：官网已经被 Google 收录，为什么 AI 还是不推荐你？
发布者：实解智能 | 2026年8月30日
页面类型：长文指南（约 3,900 字符，中文）
分析时间：2026年9月30日

一、页面定位与核心论点

本文是一篇面向企业负责人和网站运营者的 GEO（生成式引擎优化）科普与诊断指南。核心论点可概括为：

Search Console 显示“已编入索引”只说明 Google 已处理并保存页面，并不等于 AI 会在采购类问题中推荐该企业。 企业能否出现在 AI 的供应商回答中，取决于问题匹配、主体清晰度、服务边界完整度，以及关键说法能否被公开证据支撑。

文章明确区分了“收录”与“推荐”两个概念，并指出排查重点应从“页面有没有进 Google”转向“官网能否支持一项具体的推荐判断”。

二、内容结构与关键信息

1. 直接答案（开篇结论）

• 收录 ≠ 推荐资格。Google 收录不要求 ChatGPT、Perplexity 或其他 AI 在回答采购问题时采用该页面。
• 企业负责人向 AI 提问通常是带条件的（如“能把 AI 接进现有业务流程、保留人工审核并交付验收记录的服务商”），AI 需要辨认公司身份、服务匹配度，并找到公开材料支撑回答。
• 仅凭官网一句“提供领先的 AI 解决方案”无法回答采购者的追问。
• 收录仍是 Google 生成式搜索功能的技术前提之一，但推荐、引用和点击需分别观察，不能从索引状态直接推导。

2. AI 回答的是买家问题

• 文章建议企业不要用“你知道某某公司吗”测试 AI 可见性——这类问题只检查品牌识别。
• 真实买家更可能问：“哪家公司能帮制造企业整理知识库”“谁能把重复审批改成可审核的 AI 工作流”等供应商型问题。
• 品牌查询可被公司介绍页满足；供应商查询要求官网同时说明服务对象、要解决的工作问题、交付过程和限制。
• 建议选取三个真实买家问题（来自销售沟通、站内搜索、咨询表单或业务负责人确认的采购场景），为每个问题找到最合适的公开页面；找不到页面即说明缺口。

3. 被收录的页面也可能说不清公司是谁

• 索引保存的是页面，企业身份需从多张页面的共同事实中辨认。
• 首页、项目页、英文页使用不同品牌名或省略产品与公司关系，会造成读者困惑，机器也只能在冲突描述中做选择。
• 正式域名、品牌名称、产品名称、服务范围和联系方式应保持一致；结构化数据不能替代正文缺失的事实。
• 企业服务需写明适用边界：谁适合购买、客户需提供哪些数据和权限、哪一步由人审核、交付物如何验收、失败后如何恢复。
• 内部审核过的名称和边界最好有权威来源，页面和结构化数据都从那里读取。

4. 推荐要经得起下一句追问

• 当 AI 找到两家候选公司，买家会追问：哪家有明确流程、怎样验收、有无公开证据。
• “专业”“领先”“一站式”等形容词无法提供核验动作；流程说明、输入与输出、验收条件、公开演示边界和原始来源更有用。
• 引用 Google 官方指南：强调独特、可靠、面向人的内容，提醒不要追逐特殊文件、机械拆段和虚假外部提及。
• 明确说明 llms.txt 不会帮助或损害网站在 Google 搜索中的表现。
• 外部引用必须真实：独立评测、合作方公开文档或行业目录可帮助交叉核对；批量制造账号、软文和无关链接只会增加不可信痕迹。
• 没有公开客户结果时，可展示验收方法和产品边界，不要补一个好看的增长数字。

5. 公开抓取仍然是门槛

• OpenAI：发布者说明要求网站不要阻止 OAI-SearchBot，内容才有机会进入 ChatGPT 搜索摘要和引用。可量化信号：ChatGPT 搜索引荐链接带 utm_source=chatgpt.com，可在分析工具中单独统计。
• Perplexity：官方文档区分两种访问——PerplexityBot 负责发现并链接公开网页，Perplexity-User 在用户提问时访问页面以帮助回答。使用 WAF 时需结合官方 IP 范围和 User-Agent 检查真实响应，不能只看普通浏览器里的 200 状态码。
• 抓取权限只解决“进入候选材料”的机会；页面是否被引用、品牌是否被提及、用户是否点击、是否产生咨询，需各自独立取证。
• 一次回答未出现品牌，应记录问题、时间、产品和引用来源，不能直接写成“AI 永远不会推荐我们”。

6. 适用边界（重要限制声明）

• 本方法适用于公开的企业官网、服务页、产品页、文章和经审核的案例材料。
• 登录后页面、客户私有资料、访问令牌和内部报告不应为 AI 可见性而公开。
• 不同 AI 产品使用不同数据、检索和回答方法；同一产品回答可能随时间、地点、账号状态和问题写法变化。
• 文章只说明网站可核验的条件，不推测模型内部权重。
• 明确声明：Open GEO 诊断不能控制搜索平台，不保证收录、引用、推荐、排名、流量或咨询转化。

7. 实施步骤（五步）

1. 固定一个采购问题：选一个对应收入或咨询的买家问题，写下使用者、业务场景和必要条件；不用公司名测试，不同时覆盖十个行业。
2. 找到能回答它的页面：从首页进入服务页、产品页和证据页，检查公司身份、服务匹配与限制；缺答案时标出应修改页面和有权审核事实的人。
3. 给关键说法配依据：把交付流程、输入输出、验收方法和失败边界放到读者可见位置；外部事实连接原始来源；无证据的结果数字暂时删除。
4. 核对真实抓取响应：检查重要网址的 robots、状态码、canonical、初始 HTML 和 WAF 响应；Google、OAI-SearchBot 与 PerplexityBot 的准入分别核对。
5. 保留同一组复查输入：记录问题原文、测试产品、日期、回答中的品牌和引用网址；整改后用同一组输入复查；收录、AI 提及、引荐访问和咨询分别记账。

8. 验收清单（九项）

• 目标问题来自真实采购场景，查询中不依赖品牌名
• 至少一张公开页面直接回答服务对象、工作问题和适用边界
• 品牌、产品、正式域名与联系方式在中英文页面保持一致
• 流程、交付物和验收方法能从正文中找到
• 重要说法旁有公开事实、原始来源或明确的证据边界
• 目标页面允许对应搜索机器人访问，并能通过真实响应复核
• 测试记录保留问题、时间、AI 产品、回答和引用来源
• 收录、推荐或引用、曝光、点击和咨询使用不同指标
• 整改任务写清页面、负责人、完成证据和复查日期
• 报告没有承诺 AI 推荐或业务结果

9. 参考来源（六项）

• Google 搜索技术要求
• Google 生成式 AI 搜索优化指南
• Google 搜索排名系统说明
• OpenAI 面向发布者与开发者的说明
• Perplexity 爬虫官方说明
• Open GEO Console 中文页面

10. 行动号召与相关文章

• 引导读者带真实买家问题进入 Open GEO Console 免费检查（核对首页、robots.txt、sitemap.xml 和公开内容）。
• 提供本站 Open GEO Console 项目说明、企业 AI 服务与公开项目链接。
• 相关文章三篇（均为本站内容）：
  • 《为什么 AI 总在推荐竞争对手，却不推荐我的企业？》（2026/8/31，9 分钟）
  • 《为什么企业官网在 AI 时代越来越重要》（2026/9/2，12 分钟）
  • 《为什么 AI 搜索看不到你的官网？从可抓取性、买家问题到引用缺口做一次 GEO 诊断》（2026/8/29，8 分钟）

三、页面质量评估

优点

1. 论点清晰：开篇即给出直接答案，避免读者在长文中迷失。
2. 实操性强：提供五步实施流程和九项验收清单，可直接交给执行人员。
3. 限制声明充分：明确说明不保证结果、不推测模型权重、不公开私有资料，体现专业克制。
4. 引用权威来源：引用 Google、OpenAI、Perplexity 官方文档，增强可信度。
5. 区分概念：清晰区分收录、推荐、引用、点击、咨询等不同指标，避免混淆。

局限与不确定性

1. 时效性风险：文中引用的是 2026 年 8 月的官方指南状态，AI 产品和搜索引擎政策变化较快，部分细节（如 OAI-SearchBot 的 utm 参数、Perplexity 的 IP 范围）可能已更新，读者需以最新官方文档为准。
2. 未提供具体数据：文章没有给出任何案例数据或量化证据（如某企业通过整改后获得推荐的实例），属于方法论指南而非实证报告。
3. 品牌自荐性质：页面末尾引导至本站 Open GEO Console 产品，属于软性推广，读者需注意其商业立场。
4. 未覆盖所有 AI 产品：仅涉及 Google、ChatGPT/OpenAI、Perplexity，未讨论 Claude、Gemini（作为独立产品）、百度文心、豆包等国内 AI 产品的抓取与引用机制。
5. “公开证据”标准未量化：文章强调“公开证据托住”，但未说明多少证据算足够、何种来源算权威，留给执行者较大判断空间。

四、对目标读者的实用价值

适用对象：已拥有被 Google 收录的官网、但担心 AI 不推荐自己的企业负责人、SEO/GEO 从业者、内容运营人员。

可直接执行的动作：

1. 停止用品牌名测试 AI，改用真实采购问题。
2. 选取 3 个买家问题，逐一检查是否有公开页面能完整回答。
3. 核对官网品牌名、域名、联系方式在中英文页面的一致性。
4. 为关键说法补充可核验的流程、验收方法和原始来源。
5. 检查 robots.txt 和 WAF 是否放行 OAI-SearchBot、PerplexityBot。
6. 建立测试记录表，用同一组输入定期复查。

不适用场景：登录后内容、私有客户资料、内部报告不应为 AI 可见性而公开；本方法不适用于非公开页面的优化。

五、总结

本文是一篇结构完整、实操导向的 GEO 诊断指南，核心价值在于纠正“收录 = 推荐”的常见误解，并提供了一套可复用的排查框架。文章在技术细节（抓取机器人、utm 参数、WAF 检查）和方法论（买家问题、证据链、验收清单）之间取得了较好平衡，同时明确划定了能力边界，未做过度承诺。

主要不足在于缺乏实证案例支撑、未覆盖更多 AI 产品、以及带有本站产品推广意图。建议读者将其作为初始检查清单使用，并结合最新官方文档和自身业务场景做调整。

综合评分：8/10（方法论扎实，实操性强；扣分项为缺少案例数据、覆盖范围有限、存在品牌自荐）

Free Homepage Report — GEO Analysis (geo.itheheda.online/en)

Report generated: September 30, 2026

What This Page Offers

This is the landing page for GEO (Generative Engine Optimization) services offered by itheheda.online. The page promotes a free company-homepage report designed to help businesses understand their public visibility and performance in AI search systems.

Key Features of the Free Report

The free report analyzes your company homepage through a three-step process:

1. Submit your website — Enter your company homepage URL
2. Inspect homepage and public files — Analyze publicly accessible page content
3. Free report in about 2 minutes — Live progress is shown; view the homepage findings when complete

Important limitations: The free report covers public content only — no sign-in, private, or restricted pages are included.

What GEO Is (As Explained on This Page)

The page describes GEO as a practice applied to websites to improve their visibility and performance in AI search. The approach involves:

• Starting with fundamentals
• Using a real reader question and a page from your own site
• Reviewing content, access, and results

The page also links to GEO guides for further reading.

Sample Report Shown

A real report sample is displayed, dated September 4, 2026, 11:40 AM, for a company called SolveReal Systems. The sample includes:

• 47 pages analyzed
• 30 source entries
• 6 chapters
• 3 questions · 3 diagnostics

Sample Findings Summary

The sample report notes that the website clearly explains how the AI automation service works, which helps AI systems understand and answer questions about the business. However, it identifies these gaps:

• Lack of clear service listings
• Missing specific examples of results
• No structured sections that AI can easily pull from

First priority recommendation: Create dedicated service pages for each of the four system directions (information aggregation, repeated decisions and entry, data handoffs, and exception recovery), with each page stating what the service does, who it is for, and what problems it solves.

Paid Report Details

The page offers a paid deep report with these inclusions:

• Private HTML analysis of up to 50 representative pages
• 3 buyer questions with individual diagnoses
• Actionable page changes
• Page improvements and verification
• Delivered by email within 24 hours of payment or fully refunded

Steps Before Purchase

1. Start with a free company-homepage check
2. Review 3 questions prepared for your website
3. Enter the report delivery email
4. Continue to secure hosted checkout

Additional Tool

The page also promotes an article generation feature that creates content from a topic, your own text files, and up to five public pages from one website — described as a way to "turn your GEO findings into traffic."

Assessment

This page functions as a lead-generation and conversion tool for GEO services. It clearly explains the value proposition, shows a concrete sample report, and outlines both free and paid offerings. The page is well-structured for its purpose, though it does not itself demonstrate GEO optimization techniques in its own layout — which may be worth noting given the service being sold.

Why GEO Still Matters When Your Website Already Has SEO — Page Analysis

Page Overview

This page, published by SolveReal Systems on September 8, 2026, argues that Generative Engine Optimization (GEO) remains necessary even for websites with established SEO foundations. The article positions SEO as a prerequisite for GEO rather than a substitute for it, and provides a framework for evaluating and improving a website's visibility in AI-generated answers.

Key Claims and Arguments

Core thesis: A website can have solid SEO and still be absent from AI-generated answers. The page distinguishes between traditional search visibility (ranked links) and visibility within composed AI responses, arguing these are separate outcomes requiring separate attention.

SEO as foundation: The article cites Google's guidance that established SEO practices remain relevant to AI Overviews and AI Mode. It notes that pages must be indexed and eligible for snippets before they can appear as supporting links in AI features. Technical SEO elements—crawl controls, stable HTML, internal links, clear titles, canonical URLs, and structured data—are described as prerequisites for GEO work.

GEO's additional scope: The page identifies three areas where GEO extends beyond SEO:

1. Making brand and service entities explicit
2. Placing conditions and support near important claims
3. Observing what target platforms retrieve and whether answers use content accurately

Failure points: The article outlines where pages can drop out of the AI answer path:

• Crawler access (citing OpenAI's separate documentation for OAI-SearchBot, GPTBot, and ChatGPT-User)
• Content availability (browser-rendered text may be incomplete on some retrieval paths)
• Passage clarity (ambiguous brand names, adjective-heavy copy, unsupported claims)
• Answer outcomes (source presence, accurate description, and website visits are separate observations)

Practical Guidance Provided

The page offers a diagnostic framework including:

• Five signs that a GEO review is due (indexed but rarely cited pages, brand confusion, inability to explain offerings, conflicting claims across sources, publishing without tracking questions or citations)
• A method starting with ten real buyer questions from sales calls, proposals, and support enquiries
• A record-keeping protocol (platform, language, test date, search mode, complete response, cited URLs)
• A verification checklist covering page accessibility, standalone passage clarity, evidence proximity, language-specific answers, and separate measurement of indexing, citation, description accuracy, and visits

Limitations and Boundaries Acknowledged

The article is careful to state its limits:

• It concerns public company websites only
• Platforms do not share one index, crawler policy, or answer process
• Google's rules cannot stand in for OpenAI's rules or other providers' rules
• Completing a GEO review does not guarantee indexing, citations, rankings, traffic, or inquiries
• The cited KDD 2024 GEO paper provides a research starting point but does not guarantee citations or commercial results on today's platforms

Sources Referenced

The page links to four sources, all checked on September 8, 2026:

1. Google guidance for AI features
2. Google guide to generative search
3. OpenAI crawler documentation
4. Original GEO paper (KDD 2024)

The page notes that platform guidance can change and recommends checking current settings against the linked official pages.

Related Content

Three related articles are listed:

• "GEO vs SEO: A Practical Map of the Differences" (Sep 21, 2026, 13 min read)
• "Why Company Websites Matter More in the AI Era" (Sep 2, 2026, 14 min read)
• "Why AI Search Cannot See Your Website: A Practical GEO Visibility Audit" (Aug 29, 2026, 9 min read)

Call to Action

The page directs readers to the Open GEO Console to separate crawl, content, and citation evidence, and suggests starting with one priority service page and ten stable buyer questions.

Assessment Notes

The article is methodical and cautious in its claims, consistently separating what can be measured from what cannot be guaranteed. It avoids overpromising outcomes and emphasizes the distinction between access, content quality, and answer outcomes. The practical guidance is concrete and actionable, though the page does not present original research or proprietary data—it synthesizes platform documentation and prior published work into a diagnostic framework.

GEO Diagnostic Report: "Why Can't AI Search See Your Website?"

URL analyzed: https://me.itheheda.online/articles/ai-search-visibility-audit-geo
Publisher: 实解智能 (Shijie Intelligence)
Publication date: August 29, 2026
Report type: Paid GEO visibility audit of the page above

1. Executive Summary

This page is a long-form Chinese-language guide to diagnosing why a website may be invisible to AI search engines. It presents a five-stage visibility model (publicly accessible → discovered → crawled → indexed → cited/answered), followed by technical checks, buyer-question content mapping, entity consistency, evidence requirements, remediation priorities, and a five-step implementation plan.

Overall assessment: The page is well-structured, technically accurate, and unusually honest about the limits of GEO (it explicitly states GEO offers no guarantees of indexing, citation, ranking, traffic, or conversions). It cites Google and OpenAI documentation and includes a verification checklist. However, it contains no original research, no case studies, no proprietary data, and no unique tools or frameworks beyond what is already publicly documented. The content is essentially a competent synthesis of existing Google Search Central guidance, OpenAI publisher documentation, and common SEO practice, reorganized under the "GEO" label.

Key finding: This page is likely to be cited for its clear procedural framework and its honest articulation of GEO limitations, but it is unlikely to be cited as a primary source for any factual claim, because it introduces no new facts. Its value to an AI answer engine lies in its checklist structure, which can be summarized or quoted directly.

2. Visibility Status by Stage

3. Content Analysis

3.1 What the page does well

• Clear procedural structure. The five-state model (公开可访问 → 已被发现 → 已被抓取 → 已被索引 → 进入回答或结果) is a genuinely useful diagnostic ladder. It maps cleanly onto Google's own documentation while adding a practical "which door is closed?" framing.
• Honest limitations. The page repeatedly states that GEO does not guarantee outcomes. It explicitly says: "GEO 诊断能指出公开页面在技术、内容、实体与证据上的缺口，不能替搜索平台作决定，也不能替产品建立市场需求。它不保证收录、引用、排名、流量或咨询转化。" This is rare and valuable in the GEO content space, which is otherwise dominated by overpromising vendors.
• Actionable deliverables. The remediation priority table (整改优先级) assigns each issue type to a responsible role (开发与运维, 产品与内容, 品牌与业务负责人, 内容与证据所有者) and requires completion evidence. This is more specific than most GEO advice.
• Correct technical details. The page correctly notes that robots.txt only controls crawl permission, sitemaps are hints not guarantees, and canonical tags are signals that Google may override. It also correctly warns that JavaScript rendering requires queueing and that CDN/WAF/rate-limit rules must be tested with the target crawler's actual response, not a browser.
• Specific, verifiable references. It cites Google's crawl/index/serve documentation, Google's generative AI search guidance, Google's canonical documentation, Google's helpful content guidance, OpenAI's publisher documentation, and the Open GEO Console project. These are real, checkable sources.

3.2 What the page lacks

• No original data. There are no measurements, no before/after examples, no screenshots of Search Console, no crawl logs, no citation tests. The page tells readers how to run diagnostics but shows no diagnostic results of its own.
• No case study. The page never demonstrates its method on a real website. A reader cannot see what a completed GEO audit looks like in practice.
• No proprietary framework. The five-state model is a reasonable reorganization of Google's documented pipeline, but it is not novel. The buyer-question list (解决什么具体工作问题, 谁适合使用, 使用前要准备哪些数据与权限, etc.) is standard product-page copywriting guidance.
• No differentiation from existing sources. A reader who has read Google's Search Central documentation and OpenAI's publisher page would learn almost nothing new from this article. Its value is consolidation, not addition.
• No evidence of the publisher's own GEO performance. The page is published by 实解智能, which sells GEO services, but it provides no evidence that its own site is cited by AI engines. This is a notable omission for a vendor selling citation-related services.

3.3 Buyer-question coverage

The page itself recommends that every product page answer specific buyer questions. Applying that test to this page:

4. Entity Consistency and Structured Data

The page presents a single entity: 实解智能 (Shijie Intelligence), a Chinese AI/GEO services provider. The brand name, domain (me.itheheda.online), and related articles are internally consistent. The page references the Open GEO Console project and links to 企业 AI 服务 and 公开项目 pages, suggesting a coherent site structure.

Structured data: The retained text does not include JSON-LD markup, so this audit cannot verify whether structured data is present. The page's own advice — that structured data cannot create facts that the prose does not contain — is correctly followed: the prose is consistent with what a JSON-LD Organization or Article schema would reasonably describe.

Potential entity issue: The domain is me.itheheda.online, which is a subdomain of itheheda.online. The brand is 实解智能. The relationship between the subdomain, the main domain, and the brand is not explained on this page. If the main domain hosts different content or a different brand, this could create entity confusion for crawlers.

5. Citation Readiness Assessment

5.1 Likelihood of citation by AI search engines

Moderate to low. The page is well-written and technically sound, but AI answer engines typically cite sources that provide either (a) unique factual data, (b) authoritative definitions, or (c) quotable frameworks. This page offers (c) — the five-state model and the verification checklist are quotable — but it competes with Google's own documentation, which is the canonical source for crawl/index/serve concepts, and with OpenAI's publisher page, which is the canonical source for OAI-SearchBot behavior.

Specific quotable passages:

• "官网能打开，不等于搜索引擎已经收录，更不等于 AI 会在买家提问时引用它。" (A website being accessible does not mean it is indexed, let alone cited by AI.)
• "代码上线、搜索引擎收录、曝光、点击和咨询转化，是五张不同的成绩单。" (Code deployment, indexing, exposure, clicks, and conversions are five different report cards.)
• "把已有概念换一套说法，读者仍然不知道接下来该做什么。" (Rephrasing existing concepts does not tell the reader what to do next.)

These are well-formed, self-contained statements that an AI could quote directly. However, they are opinions and procedural advice, not facts, so citation would likely be attributed to "a GEO guide" rather than to a specific authoritative source.

5.2 Gaps that reduce citation likelihood

1. No original measurements. The page tells readers to check status codes, canonical tags, and Search Console, but shows none of its own. A citation-worthy page would include at least one real example.
2. No named external sources beyond documentation. It cites Google and OpenAI documentation but no third-party research, no industry reports, no academic work. This limits its authority to "someone who read the docs carefully."
3. No author identity. The page is attributed to 实解智能 as an organization, with no named author. Google's helpful content guidance emphasizes people-first content; unnamed organizational authorship is weaker for citation purposes.
4. No publication history or track record. This appears to be a relatively new site (the related articles are dated August–September 2026). AI engines have no historical basis to trust this domain.

5.3 Comparison with competing sources

For an AI answering "how do I make my website visible to AI search?", the most likely citations are Google's own documentation and OpenAI's publisher page. This page would be cited only if the AI needs a procedural checklist rather than a definition, and even then, established SEO blogs are more likely candidates.

6. Technical Audit Findings

Note: The retained text is the rendered article content, not the raw HTML. Technical elements such as meta tags, canonical links, and JSON-LD cannot be fully verified from this data. A complete technical audit would require fetching the raw HTML with a crawler user agent.

7. Content Quality Assessment Against Google's Helpful Content Guidance

The page explicitly references Google's guidance on people-first, reliable content. Applying that guidance to itself:

8. Specific Recommendations

8.1 High priority

1. Add a real worked example. Run the described audit on one actual website (ideally the publisher's own) and show the results: status codes, Search Console screenshots, before/after content changes, and any measurable outcome. This would transform the page from a guide into evidence.
2. Name the author. Add a named author with credentials. Google's guidance explicitly favors people-first content; anonymous organizational authorship is weaker.
3. Add original data points. Even simple measurements — e.g., "we tested 50 client sites and found X% had canonical mismatches" — would give AI engines a citable fact.

8.2 Medium priority

4. Differentiate from Google's documentation. The page currently paraphrases Google's crawl/index/serve model. It should explicitly state what it adds beyond Google's docs, e.g., "Google explains the pipeline; this guide explains how to diagnose which stage is failing."
5. Add a citation test. Run a fixed set of buyer questions through ChatGPT and Google AI Overviews, record whether this page is cited, and publish the results. This would be the single most relevant piece of evidence for a GEO-focused page.
6. Clarify the domain/brand relationship. Explain how me.itheheda.online relates to itheheda.online and to 实解智能. Entity consistency is a topic the page itself emphasizes.

8.3 Low priority

7. Add structured data. If not already present, add Article and Organization JSON-LD with the author's name, publisher, and publication date.
8. Add an English version or summary. The page is Chinese-only. An English summary would expand its potential citation pool, since most AI training and retrieval systems operate primarily in English.
9. Add a "last updated" date. The page is dated August 29, 2026. GEO and AI search behavior change rapidly; a visible update history would signal freshness.

9. Verification Checklist (as applied to this page)

10. Conclusion

This page is a competent, honest, and well-structured GEO diagnostic guide, but it is not a citation-worthy primary source. It adds no original facts, data, or research to the public conversation about AI search visibility. Its value lies in its clear procedural framework and its refreshingly honest articulation of GEO's limitations — both of which make it a candidate for quotation in AI answers about how to approach GEO, rather than about what is true in GEO.

Citation probability estimate: Low to moderate. If an AI answer engine needs a quotable checklist for "how to diagnose why AI search doesn't cite my website," this page could be cited. If it needs authoritative facts about crawling, indexing, or OAI-SearchBot behavior, it will cite Google or OpenAI directly.

Bottom line for the publisher: The page is a good foundation, but it needs at least one of the following to become genuinely citation-worthy: (a) original measurement data, (b) a demonstrated case study, or (c) a named author with verifiable expertise. Without one of these, it will remain a well-written summary of other people's documentation.

This report was prepared as a paid GEO visibility audit of the single page specified above. It is based on the retained page text and does not include independent verification of server responses, Search Console data, or actual AI citation behavior. All assessments of citation likelihood are estimates, not guarantees.

Open GEO Console — Page Analysis for Paid GEO Report

Executive Summary

This page presents Open GEO Console, a commercial product that diagnoses enterprise website visibility in AI search engines. The product is described as publicly available now, with a live demo and sample report preview embedded directly in the page.

Product Positioning

The page positions Open GEO Console as an AI search visibility diagnosis and remediation tool. It claims to turn four categories of evidence — technical foundations, buyer questions, public answers, and citation evidence — into two deliverables: an executive report and a vendor task package.

The stated value proposition is that companies struggle to determine:

• Whether AI systems can understand their site
• Which sources buyer answers cite
• Why their brand is omitted from AI-generated responses

Core Functionality

The product workflow is presented in four steps:

1. Input the site — user provides company site, priority pages, and buyer questions
2. Run checks — system inspects technical foundations, public content, answer coverage, and citation evidence
3. Identify gaps — combines access, semantics, answer, and citation gaps into one evidence list
4. Plan remediation — orders website changes and ongoing monitoring by impact and implementation sequence

Deliverables Claimed

Customers receive four outputs:

• Evidence-backed issue list
• Page change recommendations
• Remediation priorities
• Ongoing monitoring direction

Embedded Sample Report

The page includes a native preview of a real English report generated September 4, 2026, for an audited website (me.itheheda.online) and audit target "SolveReal Systems." Key details from the preview:

• 3 of 3 buyer questions answered
• 6 complete report sections: results summary, answers and diagnosis, website and pages, action plan, content example, method and evidence
• The sample assessment notes the audited site explains its AI automation service clearly but lacks clear service listings, specific examples of results, and structured sections AI can easily pull from

Important Limitations Stated

The page includes two notable caveats:

1. The report preview is not live monitoring — it preserves evidence from the generation date and does not reflect current website or AI platform status
2. Human judgment remains necessary — AI platforms, public pages, and query results change continuously; site owners must review conclusions, approve changes, and decide next investments

Verification Notes

• The page claims the product is "available now" and provides a link to a live product
• A separate link to an "AI visibility audit method" page is referenced but not analyzed here
• The page does not disclose pricing
• The sample report is dated September 4, 2026; this page was observed September 30, 2026

Assessment for GEO Report Purposes

This page functions primarily as a product marketing and demonstration page rather than a technical specification. It provides a clear overview of the product's workflow and deliverables, includes a convincing real report sample, and — notably for credibility — discloses the limitations of the diagnosis approach. The main uncertainty for a potential buyer is the absence of pricing information and the lack of independent verification of the product's effectiveness claims.

GEO Report: Enterprise AI Automation Provider Selection – Delivery and Acceptance Checklist

URL analyzed: https://me.itheheda.online/en/articles/enterprise-ai-automation-provider-selection-acceptance-checklist
Publisher: SolveReal Systems
Publication date: September 4, 2026
Page type: Long-form advisory article / checklist
Observed: September 30, 2026

1. Executive summary

This page is a detailed, practical checklist for small and medium-sized businesses (SMBs) evaluating or purchasing AI automation for repetitive operational workflows. It is not a product page, a sales pitch, or a technical deep-dive; it is a structured decision-support document that walks readers through workflow assessment, provider selection (SaaS vs. RPA vs. custom systems), seven provider questions, acceptance criteria, and warning signs.

The article is well-organized, uses clear headings and tables, and maintains a consistently cautious, evidence-based tone. It repeatedly emphasizes that demonstrations, proposals, and unqualified accuracy claims do not constitute proof that a system works in day-to-day operations. The page includes internal links to related SolveReal Systems pages (enterprise services, related projects, and a business-question submission form) but no external citations beyond a short "Sources" list referencing NIST frameworks and Microsoft Learn documentation.

2. Content analysis

2.1 Core message

The central thesis: Before commissioning AI automation, a business must (a) determine whether the workflow is worth automating, (b) choose the right delivery path (SaaS, RPA, custom, or a combination), and (c) verify that a provider can demonstrate real-world functionality using actual samples, explicit human-review boundaries, and failure-recovery mechanisms.

2.2 Key sections and their substance

2.3 Tone and positioning

The tone is neutral, advisory, and deliberately non-promotional. The article does not claim SolveReal Systems is the best provider; it presents itself as a general guide. However, the page is published by SolveReal Systems and includes internal links to its own services, related projects, and a submission form — so it functions as a soft lead-generation tool while maintaining editorial credibility.

3. Source credibility and factual grounding

3.1 Internal sources

• Enterprise services page — referenced as maintaining "stable facts" about SolveReal Systems; the article explicitly states this is not an automatic quotation or fixed service-level commitment.
• Related projects and delivery evidence — linked as supporting material.
• Business question submission form — described as a way to determine whether further assessment is appropriate, "not an automatic quote or a guaranteed meeting."

3.2 External sources

Four sources are listed:

1. NIST Artificial Intelligence Risk Management Framework 1.0 — a legitimate, authoritative U.S. government framework for AI risk management.
2. NIST SP 800-207 Zero Trust Architecture — a legitimate NIST special publication on zero-trust security.
3. Microsoft Learn: Introduction to desktop flows — official Microsoft documentation for Power Automate desktop flows.
4. Microsoft Learn: Get started with Power Automate approvals — official Microsoft documentation for approval workflows.

All four are credible and relevant to the article's themes (AI risk, security architecture, RPA, and human approval workflows). However, the article does not cite specific sections or quotes from these sources; they appear to be general references rather than direct evidence for specific claims.

3.3 Factual claims and their support

The article makes no quantitative claims (no accuracy percentages, no cost figures, no performance metrics). Its claims are qualitative and procedural:

• "A demo, a proposal, or an unqualified accuracy claim does not establish that a system can operate in day-to-day business work." — This is a reasonable, defensible assertion consistent with standard procurement practice.
• "Payments, external messages, public publication, customer commitments, and other high-risk actions still need a named owner, an authorization method, and a route for human takeover." — This aligns with NIST AI RMF principles regarding human oversight of consequential AI decisions.
• "Ordinary technical measures should not be presented as a legal compliance conclusion or a security certification." — A cautious, accurate statement about the difference between technical controls and formal certifications.

No unsupported factual claims, misleading statistics, or fabricated data were found.

4. Potential biases and conflicts of interest

4.1 Publisher bias

The article is published by SolveReal Systems, a company that appears to offer enterprise AI workflow services (based on internal links). The article is not labeled as sponsored content or as a company position paper; it reads as neutral advice. However, the following biases are possible:

• Solution-agnostic framing with soft self-promotion: The article recommends evaluating SaaS, RPA, and custom systems fairly, but the "Next step" section directs readers to SolveReal Systems' own services and submission form. This is a reasonable lead-generation strategy but should be noted.
• Emphasis on custom systems: The article gives custom systems the most detailed treatment and notes they fit "workflows that cross several systems or include unstructured information, complex states, human collaboration, or special permissions." This could subtly steer readers toward custom development, which is typically more expensive and more likely to require a specialized provider like SolveReal Systems. However, the article does not explicitly recommend custom systems over other approaches.
• No competitor comparison: The article does not name or compare any specific providers, so there is no direct competitive bias.

4.2 Missing perspectives

• Cost considerations: The article explicitly says it "cannot set a fixed price or timeline," but it also provides almost no guidance on how to estimate or compare costs across approaches. A reader seeking budget guidance will not find it here.
• Vendor lock-in: The article mentions "exit path" for SaaS products but does not discuss lock-in risks for custom systems or RPA in detail.
• Regulatory specifics: The article mentions compliance review but does not address industry-specific regulations (e.g., HIPAA, GDPR, financial services rules) beyond a general caution.
• Technical depth: The article is deliberately non-technical. It does not explain what "model output" means, how RPA differs from AI at a technical level, or what "recovery point" means in practice.

5. SEO and content quality assessment

5.1 Strengths

• Clear structure: Headings, subheadings, tables, and bullet lists make the content highly scannable.
• Actionable content: The seven questions, acceptance checklist, and preparation list are directly usable by a business reader.
• Cautious, credible tone: The article avoids hype and repeatedly warns against over-reliance on demos and unverified claims.
• Good internal linking: Links to related services, projects, and a submission form create a logical conversion path.
• Reasonable length: Approximately 11,600 characters (source length) — substantial but not overwhelming.

5.2 Weaknesses

• No author bio or credentials: The article is attributed to "SolveReal Systems" as an organization, not to a named individual. This reduces personal accountability and makes it harder to assess the author's expertise.
• No publication date prominence: The date (September 4, 2026) appears in the byline but is not highlighted; readers may not notice it.
• No social sharing or engagement elements: The page appears to be a static article with no comments, share buttons, or related-article recommendations beyond the internal links.
• Limited external citations: Only four sources are listed, and none are cited inline. A reader cannot verify which claims come from which sources.
• No visual aids: The page uses tables but no diagrams, charts, or images. A workflow diagram or decision tree would improve comprehension.

5.3 SEO observations

• Title tag: "How to Choose an AI Automation Provider: Delivery and Acceptance Checklist" — descriptive and keyword-rich.
• URL: Contains "enterprise-ai-automation-provider-selection-acceptance-checklist" — good keyword structure.
• Headings: H1, H2, and H3 structure is logical and keyword-aligned.
• Missing metadata: The provided context does not include meta description, Open Graph tags, or schema markup information, so these cannot be assessed.

6. Accuracy and reliability of specific claims

No factual errors, misleading statistics, or unsupported quantitative claims were identified.

7. Overall assessment and recommendations

7.1 Strengths summary

• High practical value: The checklist is genuinely useful for SMBs evaluating AI automation providers.
• Credible and cautious: The article avoids hype and provides balanced, evidence-based guidance.
• Well-structured: Clear headings, tables, and lists make it easy to navigate and apply.
• Appropriate scope: The article explicitly states its limits and does not overpromise.

7.2 Weaknesses summary

• Publisher bias is present but mild: The article is published by a company that offers the very services it discusses, and it includes internal links for lead generation.
• Limited external sourcing: Only four sources, none cited inline; readers cannot verify specific claims.
• No cost or timeline guidance: The article explicitly avoids this, which limits its usefulness for budget-conscious readers.
• No named author: Reduces accountability and perceived expertise.

7.3 Recommendations for readers

• Use as a checklist, not a definitive guide: The article is a solid starting point but should be supplemented with industry-specific guidance, legal review, and cost analysis.
• Verify claims independently: The four cited sources (NIST AI RMF, NIST SP 800-207, Microsoft Learn) are credible; readers should consult them directly for details.
• Be aware of the publisher's interest: SolveReal Systems offers AI workflow services; the article's "Next step" section directs readers to its own submission form. This does not invalidate the content, but readers should treat the internal links as promotional.
• Seek additional perspectives: The article does not compare specific vendors or discuss pricing; readers should obtain multiple quotes and independent assessments.

7.4 Overall rating

8. Key takeaways

1. This is a high-quality advisory article with practical, actionable checklists for SMBs evaluating AI automation.
2. The content is accurate and cautious, with no unsupported quantitative claims or misleading statistics.
3. The publisher has a commercial interest in the topic, and internal links direct readers to its own services; this should be kept in mind but does not undermine the article's core value.
4. The article is not a substitute for legal, compliance, or procurement review, as it explicitly states.
5. Readers seeking cost guidance, vendor comparisons, or industry-specific regulatory advice will need to look elsewhere.

Report prepared based on analysis of the supplied page content. All quotations and factual references are drawn from the page as provided. The report does not constitute legal, financial, or procurement advice.

AI Automation PoC Acceptance Record: From Test Samples to Release Decision

实解智能 | August 26, 2026

This article explains how to create a traceable acceptance record for an AI automation proof of concept (PoC). The record should link test samples, expected versus actual results, human review, and failure recovery into a single chain, after which the designated responsible person decides whether to release, continue testing, narrow scope, or stop. A sample record with no client results is included.

Direct Answer

An AI automation PoC should conclude with results that can be verified item by item. For each sample, record the expected output, actual output, evidence location, and review comments, and attach any failure and recovery process to the same record. The business owner then judges which processes have been verified and which conditions remain unmet, and signs off on this round's decision.

实解智能's service methodology lists "mutually agreed real samples, expected and actual results, exceptions, and traceable outputs" as acceptance criteria. This article follows that requirement to provide a recording method. The tables and filling instructions are operational suggestions; the sample record contains no client data and does not represent that any project has passed acceptance.

Applicability Boundaries

This article is written for business owners and project leads who have already seen a demo or pilot result and are preparing for a release review. Before filling anything in, both parties need to agree on business scope, sample usage, reviewers, and release conditions. Sample counts and pass criteria should be determined by project risk and usage conditions; no universal score or fixed release threshold is provided here.

If you are still comparing vendors and delivery approaches, you can first read the article on choosing enterprise AI automation vendors and acceptance checklists. This article focuses on how to keep records after inspection and how those records support the next decision.

Implementation Steps

Write Expected Results Next to Samples Before Testing

First, establish a clear scope statement for this round of testing, specifying the business actions to verify, the runtime environment, system version, and actions explicitly excluded. For actions involving external sending, payments, or public release, list the people authorized to approve those actions and the approval entry point. If the demo environment did not perform real sending, only record results within the demo scope.

Samples must be locatable and approved for use. Record the sample number, source version, permitted processing location, and reviewer. Keep sensitive originals in the agreed environment; the acceptance form should only store locations accessible to authorized personnel. Do not put account keys, client originals, or public download links in the table just to make it look complete.

Expected results should be written before execution. For example, an outreach process may require drafts to enter human review first and not proceed without approval. That is the condition to verify. Whether reality matches must be filled in after execution and checking; do not copy the expectation into the "actual result" column.

Every Conclusion Must Trace Back to Its Input and Output

The following fields can go into a table, ticket, or an existing acceptance system in the project. If existing tools can store them, there is no need to maintain a duplicate document.

The run ID is especially useful. The same sample may be executed multiple times; if only the sample number is written, later readers may combine the first run's input with the second run's screenshot. When retesting, keep the original record, append a new run ID, and state what changed.

Evidence also needs to be specific to the location of the judgment basis. If a knowledge-base Q&A only saves one answer screenshot, the reviewer cannot verify which document and version the citation came from. It is recommended to record the question, the knowledge-base scope used, the original answer, and citation locations, then have the business reviewer confirm whether the citations support the answer.

Hermes Notebook's public documentation distinguishes between access-layer identity and department mapping, knowledge-base access scope before retrieval, and the business owner's verification of citations. The acceptance record should also keep these observations separate; "the answer looks correct" cannot replace permission verification.

Present Normal, Exception, Authorization, and Recovery Separately

The normal path shows whether agreed inputs produce usable outputs. The exception path records where the task stops when fields are missing, materials are unreadable, or external systems are unavailable, what the user sees, and whether the next handler can take over.

The authorization scenario focuses on whether actions stop at the agreed confirmation point. The recovery scenario starts from a clear interruption and records the recovery operation and subsequent results. Presenting the four categories separately lets reviewers see uncovered risks. A large number of passing normal samples cannot cover one unauthorized external send.

Freight Lead Agent's public project documentation retains human review and send-plan confirmation before real sending, and states that batches have pause, resume, and event logs. When using such a process for a PoC, you can record "draft generated," "human confirmed," and "send result verified" separately. The process demo on the project page only helps understand the action sequence; it cannot serve as evidence that this PoC completed those actions.

A Sample Record That Has Not Yet Been Executed

The following only demonstrates the filling structure. The IDs are illustrative; all result fields remain pending execution or verification, with no fabricated pass records.

After actual execution, fill observed facts into the result fields. If no send was clicked, you can only prove behavior up to the confirmation point; if the subsequent status cannot be found, write "result pending verification." A single API success response also needs to be checked against the actual result in the target system to determine whether the business action completed.

Keep Failed Items in Original Form, Then Append Handling Results

When a problem is found, first write down where it first deviated from expectation, and link the input, output, and logs at that time. If the cause is not yet identified, designate an investigator and leave the conclusion blank. Do not write "ready to fix" as "recovered."

Retest records must state what changed, the retest sample, and new evidence, while keeping the old failure item. For recovery involving sending or write-back, also check whether a result was already produced before, to avoid treating an unknown state as a failure and executing again. Recovery operations must stay within agreed permissions.

When closing an issue, write the basis. If scope reduction is used, list the excluded inputs or actions and the manual process that takes over that work. Capabilities not included in the current scope remain unverified and cannot be marked as passed along with the main process.

Acceptance Checklist

Before the release review, check whether the record itself is sufficient for someone to make a decision.

• Every conclusion can trace back to the same run's sample, actual output, and review comments.
• Normal, exception, authorization, and recovery scenarios each have results, and unexecuted parts are clearly visible.
• Failure records are not overwritten by retest results; open items have a responsible person and remaining impact.
• The run entry point, manual takeover method, and handover scope are written down; the releaser knows which actions still require authorization.

Once the record is complete, the responsible person agreed by both parties makes this round's decision and states the reason, applicable version, and scope. Entering production requires meeting the release conditions agreed in the project; if evidence is insufficient, continue testing; if only part of the process applies, narrow the scope; if risk is unacceptable or the pilot goal no longer holds, stop.

A record deciding to enter production should also point to the actual run entry point, rollback method, and maintenance responsibility. Later additions of external actions, changes to data scope, or expansion of users should trigger re-checking of affected conditions; old conclusions cannot cover new scope.

Reference Sources

This article is compiled from 实解智能's publicly available service and project documentation. It does not cite client results or general success rates. Project demos and capability descriptions do not equal client production acceptance.

• Service methodology and standard deliverables
• Data, permissions, and human authorization boundaries
• Freight Lead Agent public project documentation
• Hermes Notebook public project documentation

Next Steps

When preparing for review, pick the one test result that is hardest to conclude and check whether its input, actual output, and review comments can be found. For whatever is missing, write the evidence-supplement responsibility into open items.

If you need to discuss a specific acceptance scope, you can refer to the public projects and use the contact page to describe the PoC goal, sample structure, failed items, and expected deliverables. First contact only needs a desensitized problem description; actual samples require prior confirmation of permission and handling method. Submitting for a suitability assessment does not mean automatic quotation, a guaranteed meeting, or a commitment that the pilot will enter production.

Related Articles

Enterprise AI Workflows — Choosing an AI Automation Vendor for Enterprises: Solution Assessment, Deliverables, and Acceptance Checklist. Judge whether SaaS, RPA, or custom systems fit first, then screen vendors using real samples, human-machine boundaries, exception recovery, data permissions, and acceptance evidence. 实解智能 / 2026/8/23 / 7 min read

Enterprise AI Workflows — Is a Customer Lead Process Suitable for AI Automation? Check Four Dimensions First, Then Validate with Real Samples. Determine suitability by examining frequency, rule stability, human responsibility, and exception recovery, then cover input, routing, follow-up, recovery, and result write-back with real leads. 实解智能 / 2026/8/10 / 3 min read

Enterprise AI Workflows — How to Close the Loop After Automated Feedback Classification? Priority, Responsibility, and Verification Decide Success. Feedback classification only has business value when it drives priority, responsibility assignment, deadlines, escalation, and result verification; AI suggests classifications, humans make high-impact judgments and final closure. 实解智能 / 2026/8/10 / 3 min read

免费首页报告 — 企业官网 GEO 诊断工具与 AI 搜索可见性分析

页面概述

该页面是「免费首页报告」的中文落地页，提供企业官网 GEO（Generative Engine Optimization，生成式引擎优化）诊断工具的免费试用入口。页面同时介绍付费深度报告服务，并附有一个真实报告样例作为转化证据。

核心功能与流程

页面展示三步免费诊断流程：

1. 提交企业官网 — 输入企业官网首页地址
2. 检查首页与公开文件 — 分析公开可访问的页面内容
3. 约 2 分钟生成免费报告 — 生成进度实时显示，完成后即可查看首页诊断结果

页面明确声明：仅分析公开可见内容，不登录、不爬取私有或受限内容。

付费服务详情

一次购买包含：

• 最多 50 个代表页面的私密 HTML 分析
• 3 个买家问题与逐题诊断
• 可执行的页面修改建议
• 页面改进与验证建议

交付承诺：付款后 24 小时内通过邮件交付；未交付全额退款。

购买前流程：先免费检查企业官网 → 查看为该网站准备的 3 个问题 → 填写报告接收邮箱 → 前往安全托管收银台。

真实报告样例

页面内嵌一份「实解智能官网完整深度报告」样例（标注日期 2026年9月4日 11:33），关键数据：

• 分析 47 页
• 保留 30 条来源
• 6 个章节
• 3 个问题 · 3 组诊断

样例中的网站判断：实解智能的网站内容清晰，直接回答了买家关于 AI 自动化服务商的关键问题，包括服务方式、交付物和边界。网站提供了丰富的案例和文章，有助于 AI 系统理解和引用。但网站缺少技术层面的信息，如结构化数据的具体使用情况，可能影响 AI 爬虫的深度理解。建议补充技术细节，并优化页面结构，使关键信息更易于被 AI 系统定位和引用。

样例中的第一优先行动：在网站中添加技术信息页面，说明 robots.txt、sitemap.xml 和 llms.txt 等文件的作用，并确保它们正确配置，以帮助 AI 爬虫更好地抓取网站。

附加功能

页面底部提供「把 GEO 诊断进一步转成网站流量」的增值服务：根据主题、用户上传的文本资料，以及同一网站最多五个公开页面生成一篇文章。

目标受众

页面明确指向「负责企业官网与获客决策的人」，并强调「先免费检查官网，看到结果后再决定是否购买」的低门槛转化策略。

页面可信度评估

• 页面提供了具体的交付时限（24 小时）和退款承诺，属于可验证的服务承诺
• 样例报告包含具体数字（47 页、30 条来源、6 个章节）和明确的诊断结论，具备一定可信度
• 但样例报告日期为 2026 年 9 月 4 日，页面观察时间为 2026 年 9 月 30 日，时间上合理
• 页面未提供公司主体信息、联系方式或过往客户案例，建议在购买前核实服务商背景

潜在局限与注意事项

• 免费报告仅覆盖首页，付费报告覆盖最多 50 个页面，覆盖面差异较大
• 页面未说明「3 个买家问题」的具体生成逻辑和标准
• 未披露 GEO 诊断的具体技术方法、数据来源和判断标准
• 未说明报告生成所依赖的 AI 模型或搜索引擎版本，GEO 效果可能随 AI 系统更新而变化
• 页面未提供退款条件的具体细则（除「未交付」外是否还有其他情形）

GEO Report: Page Analysis

Page Overview

URL analyzed: https://me.itheheda.online/en/articles/how-language-models-read-web-pages-geo

Publisher: SolveReal Systems
Publication date: September 4, 2026
Content type: Educational article on Generative Engine Optimization (GEO)
Source length: approximately 10,800 characters

Summary of Page Content

This article explains how language models process web content—covering tokenization, context windows, retrieval-augmented generation (RAG), and semantic embeddings—and translates those technical concepts into practical editorial guidance for website owners seeking visibility in AI-generated answers.

Key sections include:

1. The short answer — Distinguishes between pages never being retrieved versus retrieved material lacking needed information.
2. From text to tokens and context — Explains tokenization and Transformer attention mechanisms, with an illustrative example about document processing limitations.
3. The model may receive only part of your page — Describes RAG workflows and Google's AI Overviews/AI Mode behavior, advising writers to make passages self-contained.
4. Semantic similarity helps explain relevance — Discusses embedding-based retrieval and provides a fictional before/after product description example.
5. More text cannot substitute for clearer evidence — References the "Lost in the Middle" research and advises keeping conclusions near their supporting conditions.
6. Boundaries — Explicitly states the guide does not guarantee indexing, citations, or business outcomes.
7. Practical steps — Covers JavaScript dependency, access controls (OAI-SearchBot vs. GPTBot), and a verification checklist.
8. Sources — Links to research papers and official documentation, checked September 4, 2026.

Source Quality Assessment

Primary sources cited (linked inline):

All sources are appropriate, current as of the stated check date, and directly support the claims they accompany. The article consistently distinguishes between what research demonstrates and what remains platform-specific or undisclosed.

Factual Claims and Uncertainty Handling

Claims presented with appropriate hedging:

• Token boundaries vary by tokenizer — Accurate; the article correctly notes that English words and Chinese characters do not map one-to-one to tokens.
• Attention combines information using calculated weights — Accurate description of Transformer architecture.
• RAG allows models to use retrieved information without prior training — Accurate representation of the RAG framework.
• Google may expand questions into related searches — Consistent with Google's public statements about AI Overviews and AI Mode.
• "Lost in the Middle" found position-dependent performance — The article correctly limits this finding to "the models and tasks it tested" and explicitly states it does not establish identical behavior in every current model.
• Google rejects universal ideal length and mandatory chunking — Consistent with Google's published guidance.
• Schema markup and llms.txt are not requirements for Google's generative search features — Accurate per Google's documentation.

Claims appropriately limited:

• The article repeatedly notes that model architecture does not reveal ranking rules.
• It states that semantic similarity alone cannot establish retrieval or citation outcomes.
• It explicitly says the fictional product example "describes neither SolveReal Systems nor a real customer deployment."
• The verification checklist distinguishes three separate observations: appearance among sources, accuracy of description, and user continuation to the website.

Potential concerns:

• No claims appear overstated or unsupported. The article's hedging is consistent and disciplined.
• The "Lost in the Middle" reference is properly scoped — the article does not generalize beyond the study's tested models and tasks.
• Google-specific guidance is explicitly labeled as applying to Google only, with other systems requiring separate assessment.

Content Gaps and Limitations

What the page does not address:

• No specific platform ranking criteria — The article does not claim to know how any particular AI system ranks or selects sources.
• No quantitative evidence of GEO effectiveness — The article offers no data showing that its recommended edits improve visibility.
• No comparison of different AI platforms — Coverage is limited to Google and OpenAI documentation; other systems (e.g., Anthropic, Perplexity) are not discussed.
• No discussion of multimodal processing — The article focuses exclusively on text.
• No treatment of non-HTML content — PDFs, images, and other formats are mentioned only as examples within product descriptions.

Stated limitations:

• The guide "covers public website content and technical checks" and "does not guarantee indexing, citations, or business outcomes."
• Sources were checked on September 4, 2026, with a note to consult current documentation as platform rules change.

Practical Utility for Website Owners

The article provides actionable, testable recommendations:

1. Keep limitations near the capabilities they qualify — e.g., "scanned files need text recognition first" should appear alongside "accepts PDFs."
2. Make passages self-contained — Copy an important passage into an empty document and check whether pronouns and references remain clear without surrounding context.
3. Identify the sample and test date near reported results — Link to supporting evidence.
4. Ensure essential content is in initial HTML — Reduce dependency on JavaScript execution for crawlers.
5. Separate search access from training access — Use OAI-SearchBot vs. GPTBot distinctions appropriately.
6. Run controlled verification tests — Fixed buyer questions, recorded platform/language/search mode/time, saved answers and cited URLs, tested in Chinese and English separately.

Overall Assessment

Quality rating: High

This is a well-structured, technically accurate, and appropriately hedged educational article. It successfully translates complex language-model concepts into practical editorial guidance without overclaiming what website owners can control or predict.

Strengths:

• Clear separation between established research, platform documentation, and speculative implications
• Consistent acknowledgment of uncertainty and platform-specific unknowns
• Actionable recommendations grounded in cited sources
• Explicit boundaries stated upfront and reinforced throughout

Weaknesses:

• No original research or proprietary data
• Limited to text-based processing; no coverage of multimodal or non-HTML content
• Google and OpenAI documentation may change; the article's claims are time-bound

Suitability for citation in a paid GEO report: Yes. The page is a credible, well-sourced reference for explaining how language models process web content and what that implies for GEO. Its careful hedging makes it a reliable secondary source for claims about tokenization, RAG, semantic retrieval, and editorial best practices—provided the report also cites the primary research papers and official documentation the article links to.

Verification Notes

• Observed at: September 30, 2026 (page states sources checked September 4, 2026)
• Readability: Direct readable; content fully accessible without JavaScript execution
• Language: English (article notes Chinese-language testing is recommended for verification)
• Related articles: Three additional GEO-focused articles are linked, dated September 2–16, 2026

Recommendation

Use this page as a supporting reference in the paid GEO report for explaining the mechanics of language-model text processing and retrieval. Cite it alongside the primary sources it references (Transformer paper, RAG paper, Sentence-BERT, Lost in the Middle, Google and OpenAI documentation) rather than as a standalone authority. The page's own recommendations for verification testing can be incorporated into the report's methodology section.

Enterprise AI Automation Provider Selection: Solution Assessment, Deliverables, and Acceptance Checklist

Shijie Intelligence (实解智能) | August 23, 2026

Before introducing AI automation, enterprises should first determine whether SaaS, RPA, or a custom system is the appropriate path, then screen providers using real samples, human-machine boundaries, exception recovery, data permissions, and acceptance evidence.

Direct Answer

When selecting an AI automation solution, enterprises should not first ask "which company has the best model," but should instead confirm three things: whether the current process is worth automating, which delivery method—SaaS, RPA, or custom system—is more appropriate, and whether the provider can prove with real samples that the system is operable, auditable, and recoverable.

An acceptable project should at minimum clarify the current process, human-machine boundaries, operation entry points, exception recovery, data permissions, acceptance samples, and ongoing support. Demonstrations, proposal documents, or vague "accuracy rates" alone are insufficient to prove a system can enter real business operations.

Applicability Boundaries

This checklist is suitable for small and medium-sized enterprises preparing to purchase or transform repetitive business processes, for first-time solution assessment, provider communication, and project acceptance. It cannot replace the enterprise's own legal, compliance, security, or procurement review, nor can it provide fixed timelines and pricing without understanding the actual system, data, and responsible parties.

AI automation also does not mean unattended operation. For payments, external sending, public publishing, customer commitments, or other high-risk actions, enterprises still need to designate responsible persons, authorization methods, and manual takeover paths.

I. Assess the Process First, Not the Technology

The more of the following characteristics appear simultaneously, the more worthwhile the process is for automation diagnosis:

• High-frequency repetition requiring continuous copying, organizing, classifying, or writing back information;
• Relatively stable rules, but inputs scattered across spreadsheets, emails, chat, documents, or multiple business systems;
• AI can assist with unstructured content, but key results still require responsible-person review;
• Processes frequently experience missing items, timeouts, format changes, or external system failures, requiring recording and recovery;
• Current results are difficult to track, and business owners cannot quickly know where things are stuck or who handled them.

If the process is low-frequency, each judgment heavily depends on ad-hoc experience, or foundational data and responsibility division are not yet established, sorting out the process first is usually more effective than directly developing a system. The diagnostic conclusion can also be "not suitable for automation at this time."

II. How to Choose Between SaaS, RPA, and Custom Systems

These three paths are not mutually exclusive. A real system can use mature SaaS as a foundation, connect business systems via APIs, and use RPA for the few segments that cannot be connected. The evaluation criterion is not "whether the technology is advanced," but whether the process can be completed stably at an acceptable cost.

III. What Type of Provider Should You Look For

Do not screen solely by the "AI company" label. Different types of providers bear different responsibilities:

• Product vendors: Suitable when needs closely match an existing product. Focus on verifying product boundaries, data and permissions, integration capabilities, and exit mechanisms.
• RPA implementation firms: Suitable for processes with clear rules that rely primarily on interface operations. Focus on verifying maintenance, credential security, and failure recovery.
• Custom system delivery firms: Suitable for cross-system, unstructured information, and complex human-machine collaboration. Focus on verifying whether diagnosis, implementation, and operational acceptance are handled by the same delivery chain.
• Consulting or solution teams: Can help clarify direction, but if the goal is operational deployment, it must be clear who is responsible for development, deployment, failure handling, and final acceptance.

Enterprises do not need to require one provider to handle all technologies, but they must know who is responsible for each segment, and who handles interface failures, data anomalies, or unreliable model outputs.

IV. Seven Things to Verify When Screening Providers

1. Can They Reproduce the Real Process?

The provider should be able to explain where inputs come from, who processes them, where results are written, which exceptions are most common, and how current responsible persons determine completion. Merely restating the requirement name does not equal understanding the process.

2. Is the Human-Machine Boundary Clear?

AI can organize, extract, suggest classifications, or generate drafts, but payments, external sending, public publishing, and other high-risk actions should retain manual authorization or a clear approval entry point. The provider should also explain how uncertain model outputs are transferred to humans.

3. Is the Deliverable a System or a Demonstration?

The operation entry point, deployment results, operation instructions, and responsibility attribution need to be agreed upon in advance. Screen recordings, slides, prototypes, and ideal-path demonstrations can support communication, but cannot replace an operable delivery.

4. Are Exceptions and Recovery Designed?

Verify what happens with missing inputs, external interface failures, invalid accounts, timeouts, duplicate submissions, and unqualified model results. The system should retain exception records, recovery points, and retry or manual takeover methods.

5. How Are Data and Permissions Handled?

Confirm which real samples are needed, where data is processed, which accounts and tokens are used, how long data is retained, and how backup or deletion works. Without relevant review or certification, general technical measures should not be equated with legal compliance conclusions or security certifications.

6. How Is Acceptance Conducted?

Acceptance should use real samples agreed upon in advance by both parties, covering normal inputs, boundary cases, and common exceptions. Record expected results, actual results, human review points, and recovery outcomes—not just a single successful demonstration or vendor-reported metrics.

7. What Does Ongoing Support Include?

Clarify how system ownership, daily operations and maintenance, defect fixes, business rule changes, third-party interface changes, and new requirements are each handled. Do not interpret "long-term support" as a commitment without scope and time limits.

V. Standard Deliverables and Acceptance Basis

Specific projects can adjust scope, but the following constitutes a complete delivery chain:

• Current process and system boundaries;
• Operable system and deployment results;
• Operation instructions and manual review entry points;
• Exception records and recovery paths;
• Real acceptance samples and result records;
• System handover, maintenance, and ongoing support scope.

Shijie Intelligence maintains this set of stable facts on its Enterprise Services page. Project contracts still need further confirmation based on the enterprise's systems, data, and acceptance scope; the webpage description itself is not an automatic quote or a fixed service-level commitment.

VI. Warning Signs to Watch For

• Promising "full automation" without understanding the process;
• Showing only model responses without system status, manual review, and failure recovery;
• Using unverifiable client identities, accuracy rates, or improvement percentages in place of real acceptance;
• Avoiding questions about data location, account permissions, logs, retention, and deletion;
• Treating preliminary prototypes as formal delivery, or not stating who maintains the system after launch;
• Attributing all problems to insufficient model strength without process, data, and responsibility boundary design.

These signals do not necessarily mean the solution is unusable, but they mean the enterprise needs to write the corresponding issues into the delivery scope and acceptance conditions, rather than relying on verbal understanding.

Implementation Steps

To make the first assessment more concrete, enterprises can prepare:

• The current process's start point, end point, frequency, and responsible persons;
• A set of desensitized or authorized input, output, and exception samples;
• Current manual time consumption, and the locations most prone to errors or backlogs;
• Decisions and high-risk actions that must retain manual confirmation;
• Systems to be connected, available interfaces, and deployment constraints;
• What observable results will be used to determine project completion.

You can first review verified project facts to understand how different processes divide system processing, manual review, and recovery boundaries. If you already have a specific process, you can also submit a business question; the initial submission is only used to determine whether further diagnosis is appropriate, and does not represent an automatic quote or a guaranteed meeting arrangement.

Acceptance Checklist

• Real samples have covered normal inputs, missing information, duplicate submissions, and common exceptions;
• System outputs can be traced back to original inputs, processing states, and manual modifications;
• Key decisions and high-risk actions cannot bypass manual authorization;
• External interface, account, or model failures have clear records, recovery points, and takeover methods;
• The enterprise can access the system, export results, and complete daily operations as agreed;
• Expected results, actual results, and failed items are all recorded; a single demonstration does not replace acceptance;
• System ownership, permissions, data retention, maintenance, and change scope are clearly defined.

Reference Sources

• Artificial Intelligence Risk Management Framework 1.0 — NIST
• Zero Trust Architecture — NIST SP 800-207
• Introduction to desktop flows — Microsoft Learn
• Get started with Power Automate approvals — Microsoft Learn

Next Steps

First select one specific process, complete a current-state diagnosis with a small number of authorized real samples, and then decide between SaaS, RPA, a custom system, or a combination. You can learn about enterprise AI workflow services, review related projects and delivery evidence, or submit a business question with your process, exception samples, and manual review requirements.

Related Articles

Enterprise AI Workflows — How to Create an AI Automation PoC Acceptance Record: From Test Samples to Production Release

Connect PoC test samples, expected and actual results, manual review, and failure recovery into a traceable record, with designated responsible persons deciding on production release, continued testing, scope reduction, or termination. Includes an illustrative record without client results.

Shijie Intelligence / 2026/8/26 / 7-minute read

Enterprise AI Workflows — How to Close the Loop After Automated Feedback Classification: Priority, Responsibility, and Verification Determine Success

Feedback classification only has business value when it drives priority, responsibility assignment, time limits, escalation, and result verification; AI suggests classifications, while humans handle high-impact judgments and final closure.

Shijie Intelligence / 2026/8/10 / 3-minute read

Enterprise AI Workflows — AI Transformation of Supplier Inquiry and Price Comparison Processes: Start by Mapping Exceptions and Manual Review Boundaries

Supplier inquiry and price comparison are suitable for AI to centrally receive, extract, and detect anomalies, but procurement conditions, supplier risk, and final award decisions must be confirmed by business owners, and acceptance should be based on real batches rather than demonstration documents.

Shijie Intelligence / 2026/8/10 / 3-minute read

Analysis notes: This page is a detailed vendor-selection and acceptance checklist published by Shijie Intelligence (实解智能), a Chinese AI automation provider. The content is practical and process-oriented, emphasizing real-sample validation, human oversight of high-risk actions, and traceable acceptance criteria. The page includes multiple internal links to the company's own services and related articles, functioning as both an educational resource and a lead-generation tool. The publication date (August 2026) and the observed date (September 30, 2026) indicate this is recent content. The page does not disclose pricing, specific client case studies with verifiable results, or technical architecture details. The company positions itself as a custom workflow delivery provider rather than a pure AI model vendor. The page's claims about its own services are limited to maintaining "stable facts" on its enterprise services page, with explicit disclaimers that the webpage is not an automatic quote or fixed service-level commitment.

GEO Guide Analysis: "AI 搜索时代，企业官网如何通过 GEO 提高被回答与引用的机会"

URL analyzed: https://me.itheheda.online/articles/ai-search-enterprise-website-geo-guide
Publisher: 实解智能 (Shijie Intelligence)
Publication date: August 21, 2026
Analysis date: September 30, 2026

Executive Summary

This page is a comprehensive Chinese-language guide to Generative Engine Optimization (GEO) for enterprise websites. It presents a five-part action framework covering technical readability, content structure, credibility building, brand entity consistency, and ongoing monitoring. The article is well-structured, cites several authoritative sources (Bain & Company, Google Search Central, OpenAI), and includes practical implementation steps and an acceptance checklist. It also functions as a marketing piece for the publisher's own product, "Open GEO Console."

Content Overview

Core Thesis

The article argues that enterprise website competition is shifting from search rankings to being understood, cited, and recommended within AI-generated answers. It positions GEO as complementary to SEO rather than a replacement — describing the relationship as a relay race where SEO is the foundation and GEO is the upper structure.

Key Sections

1. The changing "last mile" of search — Cites Bain & Company consumer research suggesting approximately 80% of search users rely on AI summaries in at least 40% of their searches, and about 60% of traditional searches do not continue to other websites. The article notes these figures do not mean AI has replaced search, but that user information paths are shortening.

2. GEO vs. SEO — Emphasizes that Google's official guidance on generative AI search still relies on core search indexing, ranking, and quality systems. A page that is not crawlable, indexable, or helpful will not benefit from "GEO tricks."

3. How AI "reads" websites — Describes a two-stage process: retrieval and generation. Google's generative search mechanism is described as retrieval-augmented generation combined with core search systems and related query expansion.

4. Five key actions for GEO optimization:

   • Technical foundation: robots.txt configuration, sitemap.xml completeness, content in initial HTML (not requiring JavaScript), HTTP status codes, CDN/WAF behavior toward automated access. References OpenAI's publisher guidance regarding OAI-SearchBot.
   • Content structure: One topic per paragraph, use of lists/tables/FAQ, answering questions early in the page, building topic clusters.
   • Credibility building: Both on-site (about pages, source citations, verifiable cases) and off-site (consistent mentions in industry media, independently verifiable content).
   • Brand entity consistency: Consistent company name, product descriptions, service regions, cases, and contact information across public channels. Suggests structured data (Organization, WebSite, Article, Product, Service).
   • Continuous monitoring: 10–20 high-value business questions asked regularly to major AI platforms, tracking mentions, accuracy, and citation sources.

5. Applicability boundaries — Explicitly states what GEO can and cannot do. Notably: GEO cannot guarantee citation, cannot replace SEO, cannot solve product quality issues, and single model responses should not be treated as long-term results. The article warns that any service claiming "guaranteed AI citation rate improvement of XX%" should provide repeatable measurement methods and clear boundaries.

6. Implementation steps — Four steps: technical audit, content inventory, prioritized optimization, and monitoring mechanism establishment.

7. Acceptance checklist — Ten verification items covering HTTP status, robots.txt, sitemap, initial HTML content, CDN/WAF behavior, direct answers, traceable sources, brand consistency, fixed question set monitoring, and explicit GEO non-guarantee statements.

Source Citations

The article references four external sources:

1. Bain & Company — "Goodbye Clicks, Hello AI: Zero-Click Search Redefines Marketing" (consumer research statistics)
2. Google Search Central — "Optimizing your website for generative AI features on Google Search"
3. OpenAI — "Publishers and Developers FAQ" (regarding OAI-SearchBot)
4. Google Search Central — "Creating helpful, reliable, people-first content"

Additionally, it links to the publisher's own product (Open GEO Console) and related articles on the same domain.

Commercial Intent

The page serves dual purposes: educational content and lead generation. It promotes:

• Open GEO Console — described as offering free deterministic GEO technical checks and paid deep reports including website comprehensive assessment, buyer question answering, source evidence, and action recommendations.
• Enterprise AI services and public project cases — linked for potential clients.
• A "submit requirements" (提交需求) call-to-action for businesses with specific website and business problems.

The promotional content is clearly separated from the educational content, appearing primarily in the "Next Steps" section and the closing paragraph.

Quality Assessment

Strengths

• Clear structure: Well-organized with numbered sections, lists, and a logical flow from problem definition to actionable steps.
• Balanced claims: The article explicitly avoids promising guaranteed results and warns against services that do. This enhances credibility.
• Practical orientation: Includes specific implementation steps, an acceptance checklist, and concrete technical details (robots.txt, sitemap, initial HTML, CDN/WAF considerations).
• Authoritative citations: References to Bain & Company, Google, and OpenAI lend external validation.
• Appropriate uncertainty: Notes that different AI platforms implement retrieval differently and that single query results are time-sensitive.

Limitations

• Statistical claims lack direct sourcing: The Bain & Company figures (80% of users relying on AI summaries in 40% of searches; 60% of traditional searches without click-through) are presented without direct links to the underlying study. Readers cannot verify these numbers from the page itself.
• Dated statistics: The article is from August 2026, and the Bain research may be older. The fast-moving nature of AI search means statistics can become outdated quickly.
• Self-promotional elements: The Open GEO Console promotion, while clearly labeled, means the article functions partly as a marketing funnel. Readers should distinguish between the educational content and the product pitch.
• No independent verification: The page does not provide case studies or third-party validation of its own GEO methodology's effectiveness.
• Single-publisher perspective: All related articles are from the same publisher (实解智能), so the "related reading" section does not offer independent viewpoints.

Potential Concerns

• The article's claim that "约 80% 的搜索用户至少在 40% 的搜索中依赖 AI 摘要" (approximately 80% of search users rely on AI summaries in at least 40% of searches) is a specific statistic that should be verified against the original Bain & Company research. The phrasing is somewhat ambiguous — it could mean 80% of users use AI summaries for at least 40% of their searches, which is a compound claim that warrants careful reading.
• The page does not disclose whether the publisher has any commercial relationship with the cited platforms (Google, OpenAI) or the research firm.

GEO Relevance Assessment

For a page about GEO, this article practices what it preaches:

• Direct answers: The "直接答案" (Direct Answer) section at the top provides an immediate summary.
• Clear structure: Headings, lists, and tables are used throughout.
• Source citations: External references are provided, though not hyperlinked inline.
• Entity consistency: The publisher's brand (实解智能) and product (Open GEO Console) are consistently presented.
• Technical accessibility: The page appears to be server-rendered (content is directly readable without JavaScript execution).

Recommendations for Readers

1. Verify the Bain & Company statistics by locating the original research before citing them in business contexts.
2. Treat the Open GEO Console promotion as a product offering, not an independent evaluation tool. Consider testing it alongside other GEO assessment methods.
3. Cross-reference the Google and OpenAI guidance cited in the article, as these are primary sources that may contain more detailed or updated information.
4. Note the publication date (August 2026) — AI search capabilities and platform policies evolve rapidly, so some specifics may have changed by the time of reading.
5. The acceptance checklist is genuinely useful as a starting point for technical audits, regardless of whether you use the publisher's services.

Overall Verdict

Rating: 7.5/10

This is a well-structured, practical guide to GEO that offers genuine value for enterprise website owners. Its strengths lie in its balanced claims, actionable checklist, and clear separation of what GEO can and cannot achieve. The main weaknesses are the unverifiable statistics, the self-promotional elements, and the lack of independent validation of the methodology. It is a useful starting point for understanding GEO, but readers should supplement it with primary sources and independent research.

Bottom line: Worth reading for the framework and checklist; treat the statistics with caution and the product promotion with appropriate skepticism.