AI visibility audit

What an AI visibility audit measures

Search used to end in a list. Increasingly it ends in a paragraph, and the paragraph names two or three companies. An AI search visibility audit asks a simple question about that paragraph: are you in it, and if not, who is.

Three things get measured per answer:

That last split has a name and a published number. In a 2026 study of 3,981 domain appearances published in Growth Memo, Kevin Indig and Semrush found 61.7% of citations were ghost citations: the engine used the page as a source and never said the brand. Only 13.2% earned both a link and a name. Measure only whether you are mentioned and you will conclude you are invisible while quietly doing all the work. Measure only whether you are cited and you will conclude you are winning while nobody hears your name.

That's also why the two audits look different. An SEO audit optimises for a position in a ranked list. This optimises for being named inside a synthesised paragraph, where there may be no list, and where the engine can credit your research without crediting you.

How to run an AI visibility audit yourself

This is the method we use, in full. It takes an afternoon for a small prompt set and a week for a real one. Nothing here needs our help and every tool it calls for is free below.

1. 01

   Fix your prompt set before you measure anything

   Write down the questions your buyers type, in their words, and freeze the list. Twenty to forty is enough. The moment you edit the list mid-study you lose the ability to compare this month against last month, which is the only thing that makes the number worth collecting. Include the unflattering ones: "is [brand] a scam", "[brand] vs [competitor]", "cheapest way to do [job]". Those are the answers that lose deals.

   Where it goes wrongDo not use your keyword list. Search keywords are short and headless; questions put to an assistant are long and full of context. They return different answers.

2. 02

   Run every engine separately, five times each

   The six engines do not agree with each other, and none of them returns the same answer twice. Five runs per prompt per engine is the minimum that separates a real pattern from a coin flip. Keep the engines in separate columns forever. A blended "AI visibility score" averages away the only finding that matters, which is where the engines disagree about you.

   Where it goes wrongLogged-in browser checks are not measurement. A personalised answer is not reproducible, so two people at the same company get two different numbers and neither can be compared with itself next month.

3. 03

   Record absence as absence, never as zero

   Sometimes the engine returns nothing usable, or serves no AI Overview at all. That run contributed no sources, so it leaves the denominator. Scoring it zero says "the engine answered and cited nobody", which is a different and false claim. This one rule changes headline numbers by double digits and it is the most common way an AI visibility report ends up wrong.

   Where it goes wrongThe same applies to brands missing from an index. A missing value is null, not zero. Averaging nulls as zeros manufactures a finding out of nothing.

4. 04

   Map the sites used in your place

   When the engine answers your category question without naming you, look at what it did cite. In practice it is a short, stable list: review aggregators, Reddit threads, comparison blogs, a regulator, and occasionally a competitor writing about themselves. That list is your actual target. You get into an AI answer by being described by the sources the model already trusts, not by ranking.

   Where it goes wrongCheck whether the citation is real. Engines cite URLs that 404, redirect elsewhere, or do not contain the claim attributed to them.

5. 05

   Check whether the engines can read you at all

   Before optimising anything, confirm the door is open. Does robots.txt block GPTBot, ClaudeBot, PerplexityBot or Google-Extended? Does the page render its content without JavaScript? Does the mobile version carry the same text as the desktop one? Any of those failing caps everything else, because tidy markup earns nothing on a page that cannot be read.

   Where it goes wrongCloudflare and similar front doors return 403 to a robots.txt fetch, which some parsers read as "block everything". Verify the file rather than trusting the parse.

6. 06

   Turn the gaps into a ranked list, then re-measure

   Sort what you found by how much revenue sits behind the question, not by how easy the fix is. Then set a re-measurement date and use the same frozen prompt set. One measurement is a snapshot and cannot support any claim about whether things improved. Two measurements with the same instrument can.

   Where it goes wrongDo not promise yourself that fixing everything raises citations. Our own data says readiness does not predict citation rate. It removes the reasons you cannot be cited, which is necessary and not sufficient.

Or have us run it

The method above is complete and free. What you cannot practically do by hand is the part that costs money: every engine, five runs per prompt, across a real prompt set, with a person reading every flagged finding before it reaches you.

That review is most of what you would be paying for. The automated pass produces false positives, and in our own testing at least one finding in eleven did not survive a human reading it.

Then we fix what we can fix, and measure it again. Most audits end at the report. Ours runs three more weeks in which we ship the corrections that do not need your developer, your writer or your legal review: your entity records on Wikidata and the knowledge graph, your listings on Crunchbase, G2, Capterra, Product Hunt and your app stores, the databases specific to your category, your licence and registry entries, and correction requests to every third party page we found stating something false about you.

In week six we run the identical question set again. Same questions, same engines, same five runs, same settings. You get the before and the after side by side.

What we will not do is tell you it worked when it did not. Entity resolution either happens or it does not, per engine, and we report it that way. Movement in how often you are named carries its uncertainty printed next to it, because five runs of one question cannot support a precise percentage and anyone quoting you one is quoting noise. If nothing moved, that is the first line of the report.