What a Useful GEO Audit Should Deliver Beyond a Visibility Score

Website owners use generative engine optimization (GEO) to help AI-assisted search products find and describe their sites when answering questions. A GEO audit checks this work. The visibility score helps only if the audit shows what it measured, what evidence supports it, and what the owner should do next.

AI answers can mention a business, link to one of its pages, recommend it, or draw on information from another site. These outcomes mean different things. A useful audit records each one separately, checks whether cited pages support the answers, and reports technical access and business results separately.

What the score leaves out

The score gives one number for several distinct checks: Did the brand appear? Was its own page cited? Did the answer recommend the business, or merely mention it? Was the information accurate? Did the same result appear again when the prompt was tested another time?

An audit should preserve enough detail to answer those questions. For each test, it should identify the product and feature tested, the exact prompt, the date and relevant location or language, and the answer with its cited links. It should also say how many times the prompt was run and define what counts as a mention, recommendation, or supporting citation.

That record matters because answer behavior varies across products and test runs. A study of generative search systems found differences in retrieval and answer synthesis, including variation over time and between runs. One response is evidence of what happened in that test, not proof of a stable result.

The prompt set also needs a clear purpose. Questions about learning a topic, comparing options, and choosing a provider test different needs. An audit that uses only prompts naming the brand cannot show whether the brand appears when a person asks a broader category question. Keep the prompt wording and groupings fixed when comparing results over time, or report changes to the test set alongside the results.

Evidence that makes the findings usable

A useful GEO audit includes a test record that others can repeat, a record of answers and cited links, separate technical checks, and an action plan in priority order. A score can summarize these findings, but it should not replace them.

For each cited URL, the auditor should check whether the page supports the claim next to the link. A link to a company page does not automatically mean that page supports the answer’s recommendation or key facts. The audit record should also show whether the answer describes the business correctly and which competitor or third-party sources appeared, so the site owner can distinguish a factual error from a visibility gap.

Technical checks need their own evidence. An auditor can check whether important pages load successfully, whether access rules block automated programs that fetch pages, whether internal links lead to those pages, and whether search tools report that the pages are in their index, the list of pages they can show in search results. Each check answers a different question. A page loading in a browser does not show that an automated program fetched it, and access for that program does not prove that a search tool indexed or cited the page.

For example, OpenAI’s crawler documentation distinguishes OAI-SearchBot, an automated program that supports ChatGPT search, from GPTBot, which OpenAI uses for possible model training. The settings are separate. This documentation describes OpenAI’s automated programs, not every AI search product. Allowing one to fetch a page does not guarantee that the page will appear in an answer.

Likewise, Google says its AI search features use existing Search foundations and do not require special technical changes or structured data just for AI. An audit should not call a particular type of structured data or a file such as llms.txt a required GEO fix without evidence for that product. Google’s technical requirements also state that meeting those requirements does not guarantee that Google will fetch or index a page.

Turn each finding into a specific action

A useful recommendation points back to evidence. If a report says an important page is missing from a search engine’s index, first check whether the page is eligible for indexing in Google Search Console or Bing Webmaster Tools. Do this before rewriting the page’s content. If an answer repeatedly gets a product detail wrong, compare the cited pages and the site’s own wording, then correct the conflicting or outdated information. If relevant prompts surface competitors and third-party comparison pages, inspect those sources and decide whether the site has a useful, substantiated page that addresses the same buyer question.

The audit should name the affected URL, the prompt or technical check that revealed the issue, and what to do next. “Add more FAQs” is not a diagnosis. “This pricing page does not state which plan includes the feature named in the answer” identifies a concrete gap a reader can verify. Recommendations about clearer structure or better-supported claims are reasonable when they address such a gap, but they do not guarantee more citations.

Reports from the search providers add useful evidence, but each covers only that provider’s products. Google announced generative AI performance reports in Search Console, with worldwide rollout completed by August 31, 2026. The reports show visibility in Google’s generative search features. Impressions alone do not explain why a page appeared or whether anyone who saw it made an enquiry or purchase. Bing Webmaster Tools’ AI Performance report reports citations, the pages cited, and the queries used to ground answers. Microsoft also says citation counts do not show a page’s position or role in an answer. Its later AI visibility insights were announced as preview features.

When your analytics support it, compare those reports with measures closer to business activity, such as relevant visits from search, enquiries, sign-ups, or sales. Keep the measures separate. An answer impression differs from a citation, and a site visit and a conversion are two other outcomes.

Check an audit before acting on it

Read the report and ask whether you could repeat its test. Can you see the exact prompts, AI search features tested, dates, number of test runs, and scoring definitions? Can you open the cited pages and check whether they support the claims? Does each recommendation refer to an observed answer, access issue, indexing report, or business measure? If the report gives a score without those details, treat it as a vendor-specific indicator rather than a complete diagnosis.

You can do an initial check yourself by choosing a small set of real customer questions, including comparison or selection questions, and running the same wording in the specific AI search experiences you care about. Save the answers and links, then repeat the tests later under comparable conditions. Separately review your search performance and indexing reports. Google’s AI search guidance and Bing’s reporting describe different product scopes, so do not combine their figures into one cross-platform score without explaining the method.

A small sample cannot represent every query, location, user, or answer mode. A request from an automated program or a successful HTTP response does not prove that a provider added the page to its index, retrieved it, cited it, or relied on it. And a visibility change after a site edit does not, by itself, prove the edit caused the change. A good audit states those limits and dates product-specific findings so readers know what was tested.

Key Takeaways

• Demand the evidence behind the score: Keep the prompts, product surfaces, dates, answers, citations, and scoring rules so the result can be checked and repeated.
• Separate different outcomes: Track mentions, recommendations, supporting citations, technical access, traffic, and conversions as distinct measures.
• Tie recommendations to findings: Name the affected page and the observed gap before proposing a content or technical change.
• Treat access as one step: A page that loads or permits a crawler is not necessarily indexed, cited, or used in an answer.

What I'm building

Delegate tasks. Get software.

Give Vroni a GitHub issue, bug report, spec, or rough idea. It reads the repo, plans the change, writes code, runs checks, and works toward a review-ready pull request.

Take a look at vroni.com