The GEO Audit: What It Checks, and How to Run One Yourself

Nobody can promise you a placement in ChatGPT or Google AI Overviews, and anyone offering that is selling something other than what they claim. What you can reliably find out: whether, how and in what words an AI system mentions you today, and whether your website supplies the facts such a mention would have to be built from. That is what a GEO audit checks.

What a GEO audit checks, and what it doesn't replace

A GEO audit has two halves that are often mixed up. The first is the query side: what do ChatGPT, Perplexity, Gemini and Google AI Overviews answer today when a prospective customer asks a typical question from your industry. The second is the site side: does your website actually supply the raw material an AI would need to build a correct, citable answer, or is it missing exactly the information that would be asked for.

An audit is therefore not a bet on an outcome but a stock-take at a point in time. It doesn't replace the ongoing monitoring that follows afterwards, which we cover in Measuring AI Visibility. The difference: monitoring repeats a lean set of questions monthly to catch change over time. An audit goes deeper, once or quarterly, additionally checks the website itself, and gives the monitoring something meaningful to compare against in the first place.

The test questions: what to ask ChatGPT, Perplexity, Gemini and Google AI Overviews

Four question types together give you a usable picture, because they simulate different situations:

• Recommendation question without a brand name: "Which [profession] in [town/region] would you recommend for [problem]?"
• Category question: "What does [your service] typically cost in [region]?", to see who gets named as a reference source
• Own-name question: "What do you know about [company name]?", to surface errors and outdated information regardless of popularity
• Comparison question: "[Your company] or [a competitor], which fits better for [situation]?"

Ask the same four questions in all four systems, because they behave differently by design. Perplexity and Google AI Overviews show a source list by default. Gemini can answer differently depending on whether it's invoked as an AI Overview extension inside Google Search or as a standalone chat. ChatGPT without browsing enabled usually answers from trained data with no traceable source, while with browsing enabled it more often includes references. Test only one system and you see only a slice of how your industry is actually treated in AI answers.

Citation or mention: the distinction that decides what the nod is worth

Not every mention is worth the same, and this distinction is missing from most self-run checks. A mention is your name inside the body of an answer, with no traceable origin. A citation is a mention with a recognisable, usually clickable source attribution, the way Perplexity and Google AI Overviews display by default.

The distinction is more than semantics. A citation can generate a click you'll recognise as a referral in your analytics. A plain mention generates no measurable traffic, however favourable it reads, and it's barely verifiable outside the chat window itself. So for every mention in your audit, record whether it was a name-only mention with no source, or a citation with attribution. Collapse both into a single "mentioned: yes/no" column and you systematically overstate where you stand.

How to log the result so it stays comparable

A test run with no log is unrecoverable four weeks later. For each query, record at least: date, system and model version where shown, whether you tested signed in or anonymously, the exact wording of the question, mention or citation, position within the answer, and which competitors were named alongside you.

What matters for comparability over time: the question has to be asked in identical wording on the next round, not just the same idea. Even a different phrasing can produce a different answer, and then you can't tell whether your visibility changed or just the question did. Also log the account state, because a signed-in account with history can return different answers than an anonymous session, and both states are realistic for different groups of your actual users.

What actually makes a site quotable

An AI system can only reproduce what your site states unambiguously. Vague copy like "your partner for everything around..." offers nothing usable; a sentence like "Plumbing and heating firm in Borken, since 2015, specialising in bathroom renovation and heating replacement" does. Four traits matter most:

• Clear structure: standalone FAQ sections and headings that directly answer a question
• Unambiguous, dated facts instead of promotional phrasing
• Entity clarity: company name, location and services match across every page and in your schema.org markup
• Open technical paths: an llms.txt file with the core facts, and a feed that surfaces new content quickly

How this technical foundation gets built in practice, from crawler access to structured markup, is covered on our GEO optimisation service page; which crawler rules in robots.txt actually make sense for this is covered in robots.txt 2026.

Why part of any measurement stays unreliable

Two sources of error can't be engineered away, only accounted for. First, the session: the same question can be answered differently on two consecutive queries, due to randomness built into the answer model itself. Second, region and language: a query from Germany can be answered differently than the identical query routed through a location abroad, and a system can update its training basis without announcement, so a baseline from three months ago may already sit on a different model version.

Treat a single test result as one data point among many possible ones, not as proof. The same caution applies to figures circulating online about the share of German searches that involve an AI answer: most name neither a method nor a sample size. The one with a stated method is the SISTRIX analysis of roughly 100 million German keywords (as of February 2026), which found that an AI Overview appears for around 20% of the keywords examined. That is a statement about how often AI Overviews appear at all, not about how often any one business is named inside them, but it's the only one of the circulating figures this research could verify against a stated method.

What to do in the next 90 days

• Days 1-30: fix four to six test questions, ask them in ChatGPT, Perplexity, Gemini and Google AI Overviews, and log a baseline with date, mention or citation, and exact wording
• Days 31-60: check your website against the four quotability criteria, close the biggest gap (usually missing FAQ structure or unclear entity details), publish an llms.txt file
• Days 61-90: ask the identical questions again, compare against the baseline, then decide on monthly monitoring or outside support

Conclusion

A GEO audit isn't a product that promises a placement, but an honest stock-take: how AI systems talk about you today, and whether your website gives them enough to work with in the first place. Go through both halves properly once, and you'll know more than any single metric a tool dashboard could show you. What a guided build-out costs afterwards is covered in AI Visibility; what fundamentally sets GEO apart from classic SEO is covered in GEO Instead of SEO.