AI Visibility Audit: What It Should Cover and How to Judge One

The question an audit has to answer

The commercial question is narrow. When a buyer asks an assistant to explain your category, compare providers or shortlist firms for a specific problem, are you named? If you are not, who is, and on what evidence? Everything else in an audit is diagnostics in service of that one answer.

That is why a visibility audit is not an SEO audit with new vocabulary. A page can rank well and still be absent from every generated answer, because ranking and citation are decided on partly different evidence. The audit exists to find which of the two is failing.

The prompt set is the whole test

An audit is only as good as the questions it asks. The prompt set should mirror the buying process rather than the marketing plan: how a buyer defines the category, how they compare approaches, who is best for a specific problem in a specific sector, what the common objections are, what it costs, and what people say about you by name.

Fix that list, then run it repeatedly across the assistants your buyers actually use, and record the raw output with a date each time. Answers vary between runs and between models, so a single screenshot proves nothing. The finding is the pattern across repeats, not any one response.

Check retrieval before you argue about content

An assistant can only cite what it can fetch, parse and quote. Blocked crawlers, content that only exists after JavaScript runs, thinking locked in gated PDFs and webinar recordings, and pages that never state an answer plainly all remove a firm from the answer set before quality is ever considered.

This part is mechanical and worth doing first. OpenAI documents which crawlers fetch pages for its search features and how publishers control them. Google's guidance is that its ordinary Search fundamentals still apply to its AI features. A retrieval fault is usually cheaper to fix than a content problem, and until it is fixed nothing else can be measured.

Audit the evidence about you, not only the evidence from you

Assistants assemble answers from sources they can corroborate. That means the description of your firm held by third parties matters as much as your own copy: directories, review sites, partner pages, conference listings, press mentions and the profiles of the people who do the work.

So the audit should list where the market describes you, and whether those descriptions agree with each other. A firm described as three different things across five sources gives an assistant nothing stable to repeat, and hedged or absent recommendations are the usual result. Inconsistent entity detail is one of the most common findings and one of the least expensive to correct.

What a good deliverable contains

Not a score out of one hundred. A useful audit hands over the prompt set itself, the raw answers with dates, the competitors named in your place, the specific pages and passages that were cited, the retrieval faults found, the inconsistencies in how the market describes you, and a ranked list of changes with an owner against each one.

It should also leave behind a baseline. Without a repeatable prompt set and a recorded starting point, there is no way to tell later whether a change worked, or whether the model simply changed underneath you.

Questions to ask before you buy one

Which assistants do you test, and how many runs per prompt? Do we receive the raw outputs or only your summary? How do you separate a change we made from a change the model made? What happens if the conclusion is that our own content is fine and the problem is that nobody outside our site describes us clearly? And can the recommendations be handed straight to a developer and a writer without further interpretation?

The commercial risk is specific: paying a monthly fee for a monitoring dashboard when the work actually required is publishing, evidence and technical repair. Monitoring is worth having once there is something to monitor. It is not a substitute for the fixes.

When an audit is the wrong purchase

If a firm has no published point of view, no dated case evidence and a site an assistant cannot read, an audit will only tell it that. The money is better spent on the foundations, and the audit is worth running once there is something for an assistant to find.

The same is true immediately after a rebuild or a rebrand. Give the change time to be crawled and reflected in third-party sources, then take the baseline. An audit run in the middle of that transition measures the transition, not the firm.