Audit how assistants describe your brand, in a form you can repeat
A prompt set for auditing what ChatGPT, Claude and Perplexity say about your brand, recorded so the run you do next month is comparable to this one.
- Works in
- ChatGPT, Claude
- You need
- Your brand and domain · Three or four competitors · The questions a buyer asks before choosing
- Written for
- ai visibility audit
Scored by our own engine
This page, run through the audit we sell. Measured 6 August 2026.
An AI visibility audit is a measurement, and most of them are not, because somebody asked ChatGPT a few questions, felt bad about the answers, and wrote a summary. The prompt below makes each assistant answer normally and then record what it did in a fixed set of fields, so the run you do next month sits in the same sheet as this one and the difference means something.
The record is the point
The answers themselves are interesting once. What you need is rows: which question, which assistant, which run, named or not, in what position, on what basis. Fixed fields are what make that possible without a person rereading paragraphs and forming a slightly different judgement each time.
That is also why the prompt asks the model to fill the record from the answer it just gave. Asked fresh, a model will produce a tidier and different account of the same market. Asked to transcribe, it stays close to what a real user would have seen.
Run everything twice
These systems are not deterministic. The same question, twenty minutes apart, in the same assistant, can name a different set of companies. One run is a sample of one, and a finding built on it is noise reported as a result.
Two runs is not a confidence interval either. It is enough to notice when an answer is unstable, which is itself a finding worth recording: a market where the named set changes between runs is a market where nobody is established, and that is the easiest kind to enter.
Basis is the column that tells you what to do
Position tells you where you stand. Basis tells you what is holding you there. Being named because a third party roundup listed you is a different situation from being named because an assistant read your own pricing page, and it needs a different response. The first is fragile and cheap to lose. The second is durable and slower to build.
If the Sources column never contains your own domain, stop and check whether the assistant crawlers can reach the site at all before commissioning anything. A blocked crawler is a robots file fix, and it is the one finding here that no amount of writing can work around.
To turn a run like this into a number you can track from month to month, our AI visibility score publishes the weights we use and, more usefully, the limits of what any single number about AI visibility can honestly claim.
I am auditing how AI assistants describe a market. Answer as you normally
would for a user asking this, using web search if you have it. Do not try
to be diplomatic or complete: answer the question the way you would answer
it for somebody who asked it in earnest.
After your normal answer, add a section headed AUDIT RECORD containing
exactly these fields, and fill them from the answer you just gave rather
than from a fresh consideration:
- Named: every company, product or website you named in the answer above,
in the order you named them.
- Position of [BRAND]: the place [BRAND] appeared in that list, or "not
named" if it did not appear.
- Described as: the words you used about [BRAND], quoted from your answer.
Write "not described" if you did not mention it.
- Sources: the URLs you actually used, if you searched. Write "no search"
if you answered from training data. Do not list URLs you did not open.
- Basis: in one sentence, what your description of [BRAND] rests on. A
specific page, general impression, a third party list, or nothing.
- Confidence in the [BRAND] description: high, medium or low, and one
clause on why.
Ask me nothing. If you do not know something, record that rather than
filling it in.
The question: [ONE BUYER QUESTION, ASKED THE WAY A BUYER WOULD ASK IT]
The brand being audited: [BRAND] at [DOMAIN]What to change
Everything in square brackets is yours to replace. Nothing else needs editing.
[ONE BUYER QUESTION, ASKED THE WAY A BUYER WOULD ASK IT]- One question per run, phrased as a person would type it into an assistant. "Best project management tool for a small construction firm", not "project management software". Run the prompt again for each question rather than pasting a list, because a list makes the model summarise across all of them and the answer stops resembling anything a real user would see.
[BRAND]- The brand exactly as you want it recognised. If you trade under two names, run both separately, because the answer will often differ and knowing which name the assistants hold is itself the finding.
[DOMAIN]- Your domain, so the model can distinguish you from a similarly named company. Include it even when the name seems unmistakable. Collisions with an unrelated business are common and they are the single most misread result in this whole exercise.
How to run it
- 01Write eight to twelve buyer questions
Cover the range: broad discovery questions with no brand in them, comparisons against a named competitor, and the problem shaped questions somebody asks before they know what category they need. Discovery questions are where you are usually absent and comparisons are where you are usually described wrongly.
- 02Run each question separately in each assistant
One question, one fresh conversation, one assistant. Use ChatGPT, Claude and Perplexity at minimum. Never continue in the same thread, because the earlier answer becomes context and the second result is contaminated by the first.
- 03Run each question twice
These systems are not deterministic and the same question can name different companies twenty minutes apart. Two runs will not give you a confidence interval, but they will stop you reporting a single unlucky answer as a finding, which is the mistake almost every published AI visibility audit makes.
- 04Paste the AUDIT RECORD blocks into one sheet
One row per question, per assistant, per run, with the date. The record fields are fixed so the rows line up. This sheet, not the individual answers, is the audit, and the point of the fixed shape is that next month drops straight underneath it.
- 05Read the Basis column before the Position column
Position tells you where you stand. Basis tells you why, and it is the column that decides what to do. Being described from a third party listicle rather than from your own site is a different problem from not being named at all, and it is a more fixable one.
- 06Check the assistants can reach you at all
If the Sources column never contains your domain, run the crawler check before writing any content. A site that blocks the assistant crawlers cannot be cited by them regardless of how good the pages are, and that is a robots file fix rather than a content project.
- 07Repeat on a fixed date
Same questions, same assistants, same sheet, monthly. A single audit is a snapshot of a system that changes weekly, and the value of this exercise is entirely in the second run and everything after it.
Questions people ask
What is an AI visibility audit?
A structured check of what AI assistants say when somebody asks a question your business should be an answer to. It records whether you are named, in what position, in what terms, and on what basis, across several assistants and several runs. It is a measurement exercise rather than an optimisation one, and it is only useful if it is repeatable.
How do I audit my brand visibility in ChatGPT?
Ask the buyer questions your customers ask, in fresh conversations, and record the answer in a fixed format rather than reading it and forming an impression. Run each question at least twice, because the answers vary between runs. Then repeat the same set on a schedule, since the number that matters is the change rather than any single result.
Why does the prompt ask the model to record its own answer?
Because reading answers by hand does not scale and does not stay consistent. Asking for a fixed record after the answer gives you rows that line up across assistants and across months. The record is drawn from the answer already given rather than from a fresh consideration, which keeps it a transcription rather than a second opinion.
Can I trust what an assistant says about why it named a company?
Partly. The Sources field is reliable when the assistant actually searched, because those URLs are in the transcript. The Basis field is a plausible account rather than an inspection of its own reasoning, and it should be read as a hint about where to look, not as evidence. Treat the URLs as data and the explanation as a lead.
How often should an AI visibility audit be repeated?
Monthly is enough for most businesses and weekly is worth it while you are actively changing pages. What matters more than the interval is that it is fixed and that the questions do not drift, since changing the questions between runs makes the comparison meaningless and it is a very easy thing to do accidentally.
What do I do if I am not named anywhere?
Check reachability first. If the assistants cannot crawl your site, nothing else you do will register, and that is a robots file rather than a content problem. If they can reach you and still do not name you, the gap is usually that no page answers the buyer question directly and no third party page names you in that context.
Unsubscribe in one click. We never pass your address on.