The GEO audit: what a real one contains, and how to tell it from a site scan

What a GEO audit measures, and what it does not

It measures answers. A fixed set of your buyers' questions goes to an assistant, every answer is kept word for word, and the reading is dated. That half produces the facts no site tool can: whether you are named, whether you are named first, who is named instead, and whether what is said about you is true.

It explains the answers from the causes a business controls: what its pages state, what a crawler receives from them, how the business is described elsewhere, and what the named competitors' pages have that its own do not. What it cannot do is look inside an assistant. No audit sees how an answer was put together, and any audit that claims to, or promises a place in the answers afterward, is claiming something nobody outside the companies that build the assistants can know.

The parts of a real GEO audit

The panel follows the method on our methodology page: twelve questions, each put once to one assistant, Claude, with live web search on. The record behind it is a narrower form of the one described in setting up prompt monitoring: whether you were named, whether you were named first, and who else was. The competitor part is the method in competitor analysis in AI answers, done in full.

How to tell an audit from a scan

Ask any provider, us included, these questions before you pay. A real audit answers every one of them with a yes, except the last.

1. Does it ask the assistants anything, or only read my site?
2. Will I see the question panel, and is it fixed so a later reading compares?
3. Does it name the assistant it asked and date the reading, so you know what the answers stand for?
4. Is every answer kept word for word, so any verdict can be checked against its text?
5. Does it name the pages the assistants drew on for each question?
6. Does the fix list say exactly what to change, in order, in a form my web person can use?
7. Will the same questions be asked again once the fixes are live?
8. Does it promise a place in the answers? If so, walk away.

A site scan answers only the sixth question, and often not that one. It can still be useful as a first pass, and some of the free tools on this site, such as the crawler access check and the checklist, are scans of that kind. But a scan is not an audit, and it cannot tell you whether any of what it finds is why the assistants leave you out.

What the crawl half checks

The crawl looks for the same failures our audit of 9,483 small business websites counted across 40 trades in September 2026, on every page rather than three. Access first: 7.2% of those sites served our crawler a challenge page instead of the site, and 3.2% disallowed AI crawlers on purpose. Then form: 3.0% rendered their content only in JavaScript, and 2.1% kept the real content in a PDF.

Then identity and description. 43.1% were inconsistent about who and where they are across their own three main pages, 17.8% had no schema.org markup at all, and a further 13.7% had business markup missing the fields that make it usable. Last, the content an answer can use: 61.5% had no question-and-answer content, and 64.7% showed a crawler no review signal. The GEO checklist and the consistency audit cover the same ground as things you can check yourself.

Reading the result

The visibility score is the least useful number in the report on the day it arrives, because it is a baseline with nothing yet to compare against. Its value comes a month later, on the same questions. The parts to act on first are the source map and the fix list, read together: a fix that closes a gap on the pages the assistants actually read, for questions you are losing, goes before one that tidies a page nobody fetches.

Order the fix list by reach. Access and identity touch every question, so they come first; an answer block for one service touches one. Change what the list says, then wait for the re-run rather than asking the assistants every day, because single answers vary and a day's worth of them proves nothing either way.

Doing a first pass yourself

• Ask the assistants: six of your buyers' questions, each put to ChatGPT, Claude, Gemini and Perplexity three times, with every answer pasted into a sheet.
• Check access with the crawler access check, and view your main pages with scripts switched off.
• Compare your name, address and phone across your own pages and two listings.
• Look at the source of your homepage for business markup, and read your service pages for questions answered in their first sentence.
• Open the sites of the businesses named instead of you, and write down what they state that you do not.

That is most of a first audit, done by hand in a few mornings. What it lacks is scale, the source map, and a second pair of eyes that has no reason to see your name where it is not.

What ours costs, and what you keep

The AI Visibility Audit is $490, one-time, delivered as a report in about a week. If you go on to have us do the optimization work, the whole $490 is credited against it. It contains every part in the table above, and the fix list is written for whoever maintains your site, so it is yours to act on with or without us. It promises no place in any answer, because no one can honestly promise that.