What is an AI visibility audit?

What a complete audit contains

The questions, chosen properly. Not brand searches — those flatter everyone. The questions that generate shortlists: category questions, problem questions, comparison questions, and the "who should we talk to about X" phrasing buyers actually use. Getting this set wrong invalidates everything downstream.

Verbatim responses across systems. What ChatGPT, Claude, Gemini, Perplexity and Google’s AI Overviews return, recorded as written, with the date and the exact prompt. Paraphrase is not evidence.

The competitive set the model believes in. Who is named alongside you. This is often the most surprising output, because the model’s comparison set frequently differs from the one in your board deck.

Cited sources, kept separate. The sources an assistant showed, recorded apart from sources you infer may have influenced the answer, with an honest note of what cannot be traced. That turns a finding into a plan without pretending to a precision nobody outside the providers has.

A legibility check. Whether your content exists in the served HTML without JavaScript, whether the entity graph resolves, whether crawler policy permits what you intend. This is where problems fully within your control turn up.

Variance disclosure. These systems differ between sessions. An audit that reports one run per question, with no account of variance, is reporting noise as signal.

What an audit cannot tell you

It cannot tell you why a model said what it said. The systems are not interpretable from the outside, and any vendor presenting a causal explanation is presenting a hypothesis. Cited sources show what an assistant displayed; anything beyond that is inference — useful, and not the same thing.

It cannot tell you your revenue impact. There is no impression data, no click path, and no way to know which buyers consulted an assistant and quietly moved on. Anyone converting an AI visibility score into a pipeline figure is performing arithmetic on an assumption.

It cannot promise that a change will move the answer. It can identify the changes most likely to, and then the answer has to be re-sampled to find out. Treat the whole exercise as an experiment with a measurement plan, because that is what it is.

How to judge a vendor selling one

Five questions separate a serious audit from a dashboard with a proposal attached.

Will you show me the verbatim responses, or only a score? Will you show me which sources the assistant cited, separately from your own inferences? How many times did you run each question, and what was the variance? What will you recommend if the finding is that our positioning is the problem rather than our markup? And: what would make you tell us not to buy the follow-on engagement?

The last question is the most revealing. A firm whose audit always concludes that a long programme is required has an audit that is a sales instrument. That is not automatically disqualifying — but you should price the finding accordingly.