AI visibility audit

What an AI visibility audit includes

An audit implies a fixed scope: somebody looks at a defined list and reports what they found. Here the list is not fixed. What you get ranges from a report on where you appear to a diagnosis of why you do not.

Despina Gavoyannis at Ahrefs sets out eight steps and not one of them is diagnostic. Another audit is mostly diagnosis.

An ecommerce platform splits its audit four ways: query mapping, manual prompting, tracing each cited source back to where it came from, then an automated technical scan. That third stage is what tells you why you are missing, not just that you are.

The six audits published openly, read 4 September 2026Which of the two deliverables you are being quoted for

What each audit covers.

One provider is explicit about which half it sells. Its page opens by calling the audit a diagnostic. It lists structural and content gaps preventing citations alongside the presence check, and that is the sentence Google’s own answer quotes when it defines the term. Not every audit sold under that name contains a diagnostic step at all, so the definition Google is repeating fits some of them and misdescribes the rest. The gap is not a quibble about labels, because the two deliverables answer different questions and cost different amounts of work. The measuring half is the same work as how to measure AI visibility, run once instead of continuously. The diagnosing half asks why you are absent. That is a question about your pages, not about the engines. That does not make its findings permanent: what counts as well-structured or as fresh is a moving standard, and one of the published checks asks exactly that.

What one reading can carry

The diagnosing half looks at your own pages, so re-running it does not depend on what an engine answered today. The measuring half is a single reading of something that moves, and that is where the missing sample size in all six starts to matter.

Three separate measurements of how far one reading can move.

There is a condition under which one reading is enough, and it is the strongest defence of the format. Profound ran 753 prompts once a day against ten times a day for two weeks. The two readings landed within about two percentage points, because a set that size averages the movement out. So a one-off audit over a large enough prompt set is a reasonable instrument. Whether any particular audit meets that condition cannot be checked, because none of the six says how many prompts it used.

How long an audit takes: thirty days or five

The clearest way to see the tension in the format is to put two sentences from this category side by side. Both are written by companies selling into it, and they are about the same measurement.

One is an observation window. The other is a delivery promise. They are not the same kind of number, which is exactly why they are worth putting together.

SE Ranking (23 April 2026) and Ariad Partners, read 4 September 2026How long the number under your audit was observed

Two vendor statements about the same kind of measurement.

This does not make the five-day audit dishonest. A diagnosis of your pages asks whether your product information is structured, whether your expertise claims are verifiable, and whether the content matches how people ask. None of that needs thirty days of observation. It is the measuring half that inherits the thirty-day problem, and the audits that are entirely measurement inherit all of it.

Running that diagnosis on your pages, and then doing the work it names, is our AI visibility service.

Four questions to ask before buying one

The useful version of this purchase is specific about which half you are buying and about the sample behind the half that has one. Four questions cover it.

Buy the diagnosis for the fixes, and the measurement for the baseline.

The third question is the one nobody asks, and it comes from a measurement. If the prompt set is built entirely from opening questions, the audit has measured how engines answer people who have not asked anything yet. One vendor’s data puts a size on the difference, with a single follow-up question changing 62 percent of the brand recommendations. Whether your own buyers ask a second question is a fact about them, and this measurement does not establish it. What it does establish is that a set of opening questions and a set including follow-ups return substantially different brand lists. On the diagnosing half, the most concrete list published anywhere is Search Engine Journal’s. Whether expertise claims are verifiable. Whether content matches how people query engines. Whether product information carries machine-readable markup, and whether content that looks fresh to you is fresh to an engine. Those are checks on your own pages, not on an engine’s output, so re-running them tomorrow does not depend on what an engine happened to answer. Whether the answer stays useful is a different question: what counts as fresh, or as well-structured, is a judgment that can change.

Search Engine Journal, read 4 September 2026Which of these your own pages would fail today

The questions this audit asks. The last one is measurement, and it is here because the same audit lists it beside the others.

That list is one audit’s wording, not a standard. It is still the most concrete published version of the diagnosing half.

What you can check yourself afterwards

Whatever the report says, the next question is where an improvement would show up, and the answer is narrower than the report implies. One half of an audit produces something you can verify on your own pages; the other produces a number that your own Google reporting will not isolate.

Google documents that its own share of this is not broken out.

Google counts appearances in its AI features inside overall search traffic, under the Web search type. So on that surface there is no line to watch after an audit. What the other engines report to publishers is published separately by each of them and is not covered here. That leaves three checks you can run yourself. The first is repeating the audit’s own prompts yourself a few weeks later and seeing whether the answers moved, which is the procedure in how to measure AI visibility. The second is the diagnostic half: the markup either exists now or it does not, and that is verifiable on your own pages.

There is a third check worth naming because it is free and nobody sells it: re-read the report’s own claims about your pages. If it says your product markup is missing, that is verifiable in a browser in a few minutes, and if it says your content does not match how people ask, the prompts it used are either in the report or they are not. An audit whose diagnostic claims cannot be traced to something on your own site has given you an opinion in the shape of a finding. If the report gave you a percentage, the thing to record beside it is the sample it came from. What that percentage is a share of, and why the denominator decides whether it can be compared to anyone else’s, is AI share of voice.

Four ways an audit misleads

The first two come from the same root, which is a one-off reading of a moving measurement presented as a state of affairs. The third comes from the word not naming its contents, and the fourth from expecting a platform to report something it documents that it does not.

A report with no sample size cannot be re-read later.

The only published sample size, and what it adds up to

1. 1The three lines below the total come to between 35 and 60 prompts. The total above them says 20 to 40, so the headline figure and the split it introduces cannot both hold.
2. 2Take the number you can act on and keep the arithmetic in view. Whichever of the two you adopt, write it down beside your report, because a figure with no sample cannot be compared to itself next month.

The free ones show you what that looks like. A method described in one sentence is the common shape, and the sentence typically settles none of the three: how many prompts, which engines, how many repeats.

A fifth problem is not yours to solve. Almost every audit here is published by a company selling into this category, and the field measurements come from vendors too. The one thing you can check without trusting anyone is what each audit says it covers, which is what the figure at the top compares.

When an audit is the right purchase

Given that everything above is vendor-published, the safe reading is the narrow one. Nothing here shows that commissioning either half improves anything; what it shows is which half answers which question, so ask for the one whose question you have.

One half asks about your site. The other asks about an engine’s output.

The asymmetry is the whole decision. Whether your product pages carry machine-readable markup is a fact about your site, checkable by you at any time and not dependent on what four engines said last Tuesday. What good markup looks like will keep changing; whether you have any is not a matter of opinion. Whether four engines named you last Tuesday is a draw from a distribution, and a report that does not say how many prompts produced it cannot be checked by you, by them, or by the version of you reading it next quarter.

An audit ends where the work starts. Knowing which half of the question you have does not change either answer.

The markup still has to be written. The pages still have to say something an engine will repeat, and that second job is our answer engine optimization service.