Answer
What is in an AI visibility audit?
An AI visibility audit measures whether assistants name your brand, whether they cite your site as a source, and how much of each answer competitors occupy. A credible one runs a fixed question set across every major assistant, stores each raw answer as evidence, and ends with the specific pages you need to create.
The term is used loosely. Some audits are a genuine measurement programme; others are a handful of screenshots from a chat window. The difference is whether the result is repeatable.
What gets measured
Three numbers carry the weight. Mention rate: how often your brand is named at all. Citation rate: how often your site is used as a source. Share of answer: of every brand named, how much of the answer was you. A competitor above you on that third number is being recommended where you are not.
What makes it repeatable
Four things. The question set is fixed before the first run, so today's audit still means the same thing next quarter. Every question is repeated, because a single response is noise. The models are frozen per run, so a change in score is not quietly a change in model. And every raw answer is stored, so any finding can be traced back to the exact text it came from.
Without those, you have a snapshot you cannot compare against anything — which means you can never prove the work you do next made a difference.
What you should be handed at the end
A diagnosis is not a deliverable. The audit should end with the specific buyer questions you lose and have no page capable of winning, turned into a prioritised list of pages to create, rewrite or strengthen — plus the third-party sources being cited instead of you, which is your earned-media target list.
You should also get a technical read on whether assistants can reach and parse your site at all. If they cannot, nothing else in the audit can change an answer, which is why it is checked before anything is acted on.
Related questions
How often should it be re-run?
Quarterly suits most brands. It is long enough for published work to be crawled and start influencing answers, and short enough to catch a competitor moving on you. Re-running against the same locked question set is what makes the movement readable.
Can I do this myself?
Partly. Asking assistants buyer-style questions and recording the answers costs nothing but time, and is well worth doing. It becomes impractical when you need many questions across several assistants with repeats, stored evidence, and results comparable quarter to quarter.
Do first-party tools cover this?
Only partly. Bing Webmaster Tools reports citations from Microsoft Copilot for domains you own, and it is genuinely useful. It cannot tell you which questions you were absent from, who was named instead, or anything about ChatGPT, Claude, Gemini or Perplexity.