Direct answer
An AI answer citing your website is useful evidence. It is not, by itself, proof that a GEO change worked for the business. The page may be cited while the answer describes the offer incorrectly. The answer may be accurate but send no visit. A visit may arrive without becoming a qualified inquiry. Each is a different outcome with a different owner and a different way to check it.
Start with four measurements: whether the answer is accurate, whether the citation actually supports the relevant claim, whether people visit the site, and whether those visits become inquiries that the business considers qualified. Do not merge them into one “AI visibility” number.
This distinction matters because Google includes traffic from its AI features in Search Console’s overall Web reporting, rather than offering a separate AI-features report. A change in Web impressions or clicks can be worth investigating, but it cannot alone show that one GEO edit caused one AI answer or one inquiry. Google: AI features and your website
Use a fixed buyer question, not a brand-mention count
Choose a question that maps to a real decision. For example: “Can this provider process scanned quotations, and which decisions still require human review?” This is a hypothetical example, not a description of a customer project. Record the wording, language, product, search mode, date, answer, and every displayed citation. The question should stay stable through a comparison; otherwise the team is testing a different request.
The first result to assess is answer accuracy. Does the response preserve the actual condition that scanned documents need text recognition? Does it retain the human-review boundary? A brand mention without those qualifications can be commercially worse than no mention because it gives a buyer the wrong expectation.
Keep a claim ledger next to the saved answer. For each material statement, mark it as accurate, inaccurate, incomplete, or not answered, then link the page that establishes the assessment. This gives the content owner a reviewable correction target instead of a vague request to “improve GEO.”
A citation is a link to inspect, not a score to celebrate
The next measurement is citation quality. Open the cited URL and test the precise nearby claim. A company service page may be the right source for current scope and stated conditions. It cannot, on its own, establish independent market recognition or customer satisfaction.
Also record whether the citation is present, whether it supports the claim, and whether it is current. Those are separate fields. The ALCE research evaluates answer correctness separately from citation support and completeness; a linked answer can therefore be well sourced in one place and still wrong or unsupported elsewhere. Enabling Large Language Models to Generate Text with Citations Ten citations to syndicated copies of one press release are not ten independent confirmations. A source can be authoritative for one proposition and unsuitable for another.
OpenAI’s publisher guidance explains that website owners who want their pages considered for ChatGPT Search should allow OAI-SearchBot, while GPTBot is used for a separate training purpose. Allowing access makes a page available for consideration; it does not promise that a target answer will cite it. OpenAI: Publishers and Developers FAQ
Visits show behaviour after the answer
If the platform, analytics setup, and referral information make a visit observable, record it as a visit. Do not quietly convert it into “AI-generated revenue.” A visitor may arrive from a shared link, a browser result, an article, a copied citation, or a direct visit after seeing an answer elsewhere. OpenAI says it appends utm_source=chatgpt.com to ChatGPT Search referral URLs. When that parameter reaches analytics intact, it is useful referral context, not proof that the referral caused a commercial result. OpenAI: Publishers and Developers FAQ
Useful visit measures are modest: landing page, referral information when available, date, engagement with the relevant page, and the next action taken. Compare a defined period before and after a reviewed change, while preserving context such as campaign launches, seasonality, site changes, and product changes. A short-lived movement is an observation, not a causal finding.
Google advises site owners to use its normal Search reporting for AI-feature traffic. That makes Web performance a useful supporting signal, but it does not isolate AI Overview or AI Mode into its own channel. Google: AI features and your website
Bing AI Performance offers a different, platform-specific observation. It reports aggregated citation activity and grounding-query phrases across supported Microsoft AI surfaces. Those phrases are aggregated signals, not the original user prompts; the product does not report referral traffic or organic ranking. A movement in citations can help decide what to investigate, but it does not establish that a page edit caused the movement. Bing: AI Performance Bing Webmaster Blog: AI Performance public preview
Qualified inquiries are a business judgment
An inquiry should be called qualified only under a rule the business already accepts. It might require a named organization, a relevant problem, a reachable contact, and a plausible budget or authority. The exact rule belongs to the commercial team; a web analytics tool cannot infer it reliably.
When a qualified inquiry arrives, record the inquiry date, landing page or stated discovery route, buyer question, and qualification decision. Ask how the buyer found the company, but keep the answer as reported attribution, not proof. A buyer may have used an AI answer in research and later return through another channel. That is valuable context, yet it is not clean causal attribution.
A practical GEO measurement sheet
Use one row for each test question and observation date. Keep four columns or linked records:
| Outcome | Minimum evidence | What it does not prove |
|---|---|---|
| Answer accuracy | Saved answer and source-by-source review | That the answer will repeat |
| Citation support | Cited URL and the passage supporting the claim | That every material claim is supported |
| Site visit | Analytics event and available referral context | Why that person decided to visit |
| Qualified inquiry | CRM or sales record and accepted qualification rule | That one citation caused the inquiry |
This keeps an awkward but important result visible. A page can become easier to cite while the associated answer remains incomplete. Or accurate answers can increase before the business sees a meaningful inquiry pattern. Both findings can guide work, but neither should be overstated.
Scope and boundaries
AI products, web indexes, and displayed citations change over time. A fixed question and saved evidence improve comparability; they do not make a live system deterministic. This method does not reveal a platform’s internal ranking or source weights, and it does not guarantee citation, traffic, or commercial demand.
Use the method for public pages and approved business claims. Keep customer-confidential material, account-only sources, and unverified outcomes out of the evidence set.
Practical steps
- Select a small, stable set of real buyer questions linked to service decisions.
- Save baseline answers and citations under the named product, mode, language, and date.
- Review the official page for the missing capability, condition, or evidence; obtain fact-owner approval before publishing a change.
- Retest the same questions after the change is public, without overwriting the baseline record.
- Review answer accuracy, citation support, visits, and qualified inquiries separately at an agreed interval.
The claims in an answer should also be checked against the current service description, rather than inferred from a brand mention alone.
Verification checklist
- Is the buyer question fixed and recorded with its test conditions?
- Can each material answer claim be checked against an appropriate source?
- Does the displayed citation support the adjacent claim and reflect current facts?
- Are analytics observations separate from attribution claims?
- Does the team have an agreed definition of a qualified inquiry?
- Is the earlier answer preserved for comparison after a change?
Next step
Open GEO Console audits official pages against buyer questions and organizes access, content, answer, and citation gaps into prioritized website work. It does not provide live citation tracking or conversion tracking, and it cannot promise that a platform will cite a page. See the Open GEO project, browse public projects, or contact us with a public site and the buyer questions you want to test.