Most audits fail before the first question is asked, because they measure the wrong thing and record too little to prove anything afterwards. This is a timeboxed version that produces a number you can defend in a client meeting.
An AI visibility audit answers one question: when a customer asks an assistant for what you sell, does it name you? Everything else — share of voice, sentiment, trends — is a derivative of that. If the audit does not produce a defensible yes-or-no per question, it is decoration.
Which is why the deliverable is not a dashboard. It is a table with one row per question and one column per assistant, and four recorded facts per answer. Thirty minutes gets you that table. Anything longer and you will not repeat it next month, and an audit you run once is a screenshot, not a measurement.
Remove your brand from the question.
Typing "tell me about Acme" into an assistant produces a paragraph about Acme. Every time. That is not visibility, it is autocomplete — the prompt did the work, not your reputation. A buyer who has never heard of you types a category, a need and often a place: "web design agency for a small manufacturer in Nuremberg". Your brand cannot appear in that string, so the model either volunteers you or it does not.
If you take one thing from this page: the only question worth measuring is the one your brand is not in.
Five is the right number. Three cannot separate a pattern from a fluke; ten makes the audit long enough to skip. Write them the way a buyer types, not the way a marketer thinks.
A useful set covers five different shapes of intent, because assistants behave differently across them:
Write them down once and freeze them. Changing the questions between audits invalidates the comparison, and it is the single most common self-inflicted wound in this discipline.
Run each question on at least two engines, and make sure one of them searches the web while the other answers from memory. They fail in different ways, and knowing which failure you have determines what you fix.
If you only check one, you will draw the wrong conclusion about which lever to pull. An answer with no citations is not evidence that nothing can be done — it is evidence that this engine did not look anything up.
Copy the answer, but record only these four things. This is the part almost everyone skips, and it is the part that makes the audit worth anything a month later.
| Field | Why it matters |
|---|---|
| Engine + date | Different assistants disagree. An unlabelled number cannot be reproduced or defended. |
| Searched the web? (yes/no) | Separates a retrieval problem (fixable in days) from a training-data problem (months). |
| Who it named instead | The competitive set you are actually up against, which is often not the set you assumed. |
| Pages cited | The specific URLs to influence. This is your to-do list, not a diagnostic footnote. |
A fifth field is optional and useful: a verbatim sentence from the answer. Not for sentiment analysis — for proving to yourself later how the phrasing changed. "They are a small agency in Bavaria" becoming "a well-regarded specialist for manufacturers" is a real movement, and you only see it if you kept the sentence.
A spreadsheet with one row per question-engine pair. Two engines and five questions is ten rows. If your audit produces more rows than that in the first pass, you are measuring more than you can act on.
Two numbers, nothing else:
Then choose exactly three actions. More than three and none happen:
Discussed above, and still the most common. It manufactures mentions.
ChatGPT, Gemini and Perplexity read different sources and disagree routinely. Choosing the one that flatters you is not an audit; it is a press release.
New month, new questions, no comparison. Freeze the set for at least a quarter.
Sentiment without a mention is worthless, and sentiment is the easiest thing to hallucinate. Count mentions; quote sentences when they change.
Retrieval fluctuates. A single run that shows three mentions where last month showed one is a hint, not a result. Two consecutive runs is a pattern; four is a trend.
Three outcomes are possible, and they call for different work.
Named on the searching engine, absent on the memory one. Your retrieval footprint is fine and your long-run familiarity is not. That is a corroboration problem: more independent mentions, more consistent naming, more quotable pages. Expect months, not weeks.
Absent on both, with citations pointing at directories. You are not in the lists that define your category. Fix the listings before writing anything new — no article outranks a directory that every answer already cites.
Named but never cited. The assistant knows your name and prefers someone else's page. That is an asset problem: you need one page that is worth quoting for that question, not more pages generally.
Here is what the audit actually looks like when it is done, because the format matters more than the theory. A small web design studio in southern Germany ran it in a single sitting.
The five frozen questions: which web design companies would you recommend to a small business; I need to hire a web designer, who should I contact; what are the best web design agencies in Nuremberg; how do I choose a web designer without getting burned; which web designer is good for a manufacturer with a small budget.
The result after fifteen minutes: zero mentions on the memory-based engine across all five questions, one mention on the searching engine, and that mention came from a directory profile the studio had not looked at in two years. The cited domains were a review platform, two comparison listicles and one forum thread.
What that told them: not a messaging problem, and not a content problem. They had no presence in the four sources that answer their category. The single mention they did get came from a stale listing, which is the cheapest possible signal that the listing route works.
The three actions they chose: correct and complete the profile on the most-cited review platform; write one page answering the fourth question properly, because every cited page answered it badly; and fix the two spellings of their brand name that appeared across their own site.
What they deliberately did not do: publish ten articles, buy links, or re-run the audit the next day. None of those would have changed the answer, and all three would have consumed the time that the listing work needed.
That is the shape of a useful audit: a number, the reason for it, and three things to do this week. If your first audit produces a twenty-page report and no actions, the report is the problem.
Cut in this order, never by shortening the questions:
What you must not cut is the freezing of the questions. An audit whose questions drift is not a shorter audit, it is a different audit every month with the same name.
If you serve several cities, run the five questions per city but keep the category constant. Answers are strongly location-sensitive: an assistant asked about Nuremberg and one asked about Hamburg will cite different local directories, and a single national question hides both.
Run the set twice, once per service, and do not average the two scores. A firm that is visible for tax advice and invisible for audit work has two different problems, and an averaged number conceals exactly the one worth fixing.
The value of this audit is not the first run; it is the fifth. Three habits keep the series comparable:
Do those three things and a folder of ten audits becomes a genuine asset - the only record of how an assistant changed its mind about your brand, and why.
Skip the manual version. Our free check runs the buying questions on a real assistant and shows you every answer with the pages it cited — no signup for the first one.
Run the free AI visibility check →Five is enough to see a pattern and few enough to finish. Fewer than three cannot separate a trend from a fluke; more than ten makes the audit long enough that you will not repeat it — and repetition is what makes the data useful.
No, not for the visibility score. A prompt containing your brand forces the assistant to discuss you, so every answer looks like a mention. Keep brand probes in a separate section and never add them to the percentage.
At minimum one that searches the web and one that answers from memory, because they fail differently. ChatGPT, Gemini and Perplexity read different sources and routinely disagree, so a single-engine audit is a coin flip rather than a measurement.
Monthly, with the same five questions, the same engines and the same recorded fields. Weekly checks add noise; quarterly is too slow to tell whether last month's work did anything.
There is no universal benchmark — it depends on how competitive your category is. What matters is direction and gap: whether the number moves after you publish or correct something, and which competitors occupy the answers you are missing from.
Written by the OptiLink team. We publish our own numbers, including the unflattering ones. Related: How to get ChatGPT to recommend your brand · How the check works