What an AI Visibility Audit Should Contain
Seven steps to run yourself or demand from a supplier, from the language of each market to the raw answers behind every number.
You are probably suspecting that an AI visibility audit is someone typing your company name into ChatGPT, taking a few screenshots and invoicing for a PDF. Some audits are exactly that. The suspicion is fair, and the answer to it is not a better-looking PDF. It is a protocol: a fixed list of questions, languages, assistants and repetitions that anyone can rerun and get a comparable result.
What follows is that protocol in seven steps. You can run it yourself with a spreadsheet and an afternoon per market, or hand it to a supplier and ask which steps their audit covers. A step they skip is not a detail. It is a number you cannot check.
1. Write the questions down before anyone runs them
An audit starts with a list of buying questions, written before the first answer is read. These are the questions a buyer asks when they do not yet know you: which suppliers make a given component, who certifies to a given standard, which firms serve a given region. Your brand name does not appear in them.
Questions that contain your name belong in a separate group. They show what an assistant says about you once someone already knows you exist. That is useful, but it says nothing about whether you get recommended. An audit that mixes the two groups reports a visibility the buyer never sees.
The list is then frozen. If the questions change between runs, two audits cannot be compared. How to build the set for each market is covered in building a prompt set for every export market.
2. One language per market, asked in that language
A buyer in Lyon asks in French. A buyer in Stuttgart asks in German. An audit that runs every question in English and translates the conclusions has measured the English-speaking market and nothing else.
Each market in scope gets its questions written natively, not translated word for word, because the category name a German engineer types is often not the literal translation of the English one. And language is only half the setting. Where the question appears to come from is the other half, which is the next step.
3. Name the assistant, the model and the route
"We checked ChatGPT" is not a specification. For every answer, an audit records:
- The assistant: ChatGPT, Claude, Gemini, Perplexity, or whichever others are in scope.
- The model and the date. The default model inside an app changes without notice, so a result without a date cannot be placed.
- The route: the consumer app, logged in or out, or the API.
- Web search on or off. In the OpenAI and Anthropic APIs, web search is a tool that has to be added to the request. Without it, the model answers from what it learned in training.
- The location. Both APIs accept an approximate user location with a two-letter country code. In the app, location comes from the account and the connection.
None of these routes is the correct one. Each measures a slightly different thing. What makes an audit useless is not choosing the API or the app. It is not saying which.
4. Repeat every question, and report counts
The same question, asked twice, can return a different list of companies. That is how these systems work, as explained in why one check proves nothing. A single run per question is not a finding. It is an anecdote.
The result for a question should read as a count: named in 6 of 10 runs, on Claude, in German. Counts carry their own warning, too. With ten runs, a result of five out of ten is compatible with a true rate anywhere between roughly two and eight in ten, as any binomial confidence calculator will show. That is why a sound audit spreads its runs across many questions, languages and assistants rather than piling them onto one, and why a gap of one or two mentions between two audits is not a trend.
5. Record the sources, not only the names
A mention is your company name in the text of an answer. A source is a page the assistant consulted or linked. They are different measurements and an audit needs both. The names tell you whether you are recommended. The sources tell you which pages the recommendation was built from, which is where any further work would have to happen.
Expect sources to be uneven from one assistant to the next. In the 17,719 answers PSentry collected through the APIs between December 2025 and October 2026, Claude reported at least one source in 97 answers out of 100, ChatGPT in 44 out of 100. The sample has limits: the questions are those of PSentry users, mostly in B2B services and industry, and it does not represent all searches. It is still enough to say that an audit reading sources from ChatGPT alone will see far fewer of them than one that reads all four assistants.
6. List the competitors named instead of you
When an answer does not name you, it names someone. Those names, counted market by market with the same method as your own, are often the most useful page of the whole audit. They are frequently not the competitors your sales team would list: a local distributor, a firm from a neighbouring sector, a directory that has become the answer itself.
The rule is symmetry. If you count your own mentions over ten runs, count theirs over the same ten. A competitor list built from one memorable answer is a story, not a measure.
7. Hand over the raw answers
The deliverable of an AI visibility audit is not the score. It is the full text of every answer, with date, assistant, model, language, question and the sources it listed, in a file that opens in a spreadsheet. A summary and a score can sit on top. Without the raw file nobody can check the summary, and next year's audit cannot be compared with this one.
What is not an AI visibility audit
- A folder of screenshots. They show one run, in one account, on one day.
- A "position" or "rank" in ChatGPT. Answers do not have a fixed order to rank in. A count over repeated runs is the closest honest equivalent.
- A single score with no answers behind it.
- A report written in English for a company that sells in four languages.
- Anything that arrives with a promised improvement. An audit measures. Whoever promises what the next measurement will show is selling something else.
Once, or every month
The protocol above produces a snapshot. Rerun with the same questions, it becomes a series, and only a series separates a real change from the noise described in step 4. PSentry runs fixed questions across ChatGPT, Claude, Gemini and Perplexity, in each language and market you monitor, twice a month, and keeps the competitors named and the sources cited, with a CSV export of the results. It does not replace step 1: the questions are generated for each site, and someone who knows the buyers still has to read them and decide whether they are the right ones. And whatever tool you use, ours included, ask how it handles step 3. The answer belongs in writing.