AI visibility audit - how to check whether ChatGPT knows your company

The question “does AI see me” has three different answers and each one is checked differently - which is why this is an audit, not one question to ChatGPT. You can be known to a model and never be cited. You can be cited with an out-of-date price in the answer. You can not exist at all, while your site has been in Google for a decade.

This piece is a measurement procedure: what to ask, where, what to record and how to turn it into a number you can compare a quarter from now. You will not find an explanation of what AI Overviews are, or how to write content so AI cites it - that is a separate topic, set out in the piece on SEO and AI. Here I assume you know the stakes and want to know where you stand.

Three questions, not one

Before you start, separate what you are auditing. Three different states, three different fixes:

1. Does the model know you exist? Test: ask about the company name directly. “I have no information about that company” is state zero.
2. Does it recommend you when somebody asks about your service? Test: a buying question with no company name in it. That is the real stake - that is what a genuine customer’s question looks like.
3. Is what it says true? Test: fact-check the answer. An out-of-date price, an old address or a service you do not provide does more damage than silence.

Most people only check the first one and draw a conclusion from a single query. That is like judging your SEO by typing your own company name into Google.

Step 1: build a fixed set of questions

The key to a meaningful measurement: the questions have to be the same every time. Otherwise you are comparing things that cannot be compared. A dozen or so is enough, in five categories:

Brand (2-3 questions) - do you exist for the model at all:

• “What does [company name] do?”
• “What do you know about [name] from [city]?”

Category (3-4) - are you in the consideration set:

• “Who does [your service] in [country]?”
• “Name providers of [service] for small businesses.”

Buying (3-4) - the most important ones, because that is how a customer with money asks:

• “I am looking for someone who can [specific customer problem]. Who do you recommend?”
• “What does [your service] cost and who does it properly?”

Local (2-3) - if you work a specific area:

• “[Service] in [city] - who should I pick?”

Comparative (2) - these show who the model puts you next to:

• “[Your company] versus [competitor] - what is the difference?”
• “What are the alternatives to [competitor]?”

Write them down once, in a sheet, and do not change them. Rewording a question changes the answer more than a quarter of work on content does.

Step 2: ask them in five places

Every engine answers from a different source, and each one is worth checking for something different:

You do not have to do all five the first time. A sensible minimum is three: ChatGPT, Perplexity and AI Overviews. Perplexity is the most instructive of them, because it shows sources outright - you see not only that you are absent, but who is there instead of you.

How to ask so the result is comparable

Five rules, without which the measurement is worthless:

• Fresh session, logged out or in a private window. A model that remembers you and your company answers differently than it would a new user. That is the same trap as checking your own Google position in your own browser.
• Turn off personalisation and memory where the tool allows it. In ChatGPT that is the memory setting; in the browser, private mode.
• One question per session. A conversation builds context, and context changes the answer. A new chat for every question.
• All on the same day. Engine indexes shift week to week; a measurement spread over a fortnight mixes two different realities.
• A screenshot of every answer. You will not reconstruct this from memory in three months, and models are not always repeatable - the same question can produce two different answers.

Step 3: exactly what to record

Five columns in the sheet, one row per question and engine:

1. Mention: yes/no. Whether the company name appears at all.
2. Position in the answer. First on the list, fourth, or just a mention at the end. A model usually names three to five providers and position matters.
3. Whether your domain is in the sources. Separate from a mention - you can be named without a link and linked without being named.
4. Factual accuracy. Write out every error explicitly: wrong price, a service that does not exist, an old address, somebody else’s work credited to you.
5. Who is there instead of you. The list of providers the model named. This is the most valuable column in the whole sheet - it shows who the engines consider the answer to your buying question, and where they get their knowledge about them.

At the end, one number to compare next quarter: on how many of the dozen-plus questions the company name appeared. Simple, honest, repeatable.

Step 4: check whether the bots are allowed in at all

The most banal and most frequently missed cause of absence: the site blocks AI bots in robots.txt. It usually happens after installing a “protect my content from AI” plugin, or after copying somebody else’s file off the internet.

Open yourdomain.com/robots.txt and look for agent names. It pays to know which is which, because one name is not the same as another:

• GPTBot - OpenAI, data collection. Blocking it does not remove you from live answers, but it cuts you out of the model’s knowledge.
• OAI-SearchBot - OpenAI, the index behind search in ChatGPT. This is the one that decides citation.
• ChatGPT-User - an on-demand page fetch, when a user supplies a link or the model decides to check a source.
• PerplexityBot - Perplexity’s index.
• ClaudeBot - Anthropic.
• Googlebot - the ordinary Google bot. This is what feeds AI Overviews.
• Google-Extended - a separate switch covering the use of your content by Gemini. Blocking it does not remove you from AI Overviews and does not affect your search rankings.

That distinction is the source of the most common misunderstanding in this area. People block Google-Extended believing they are opting out of AI in search results, then wonder why nothing changes. It works the other way too: blocking Googlebot cuts you out of search and out of AI Overviews at the same time.

While you are there, check two things beyond robots.txt: whether your hosting firewall or CDN is cutting those agents off server-side (a block invisible in the file), and whether key pages carry a noindex tag.

Step 5: proof from the server logs

robots.txt says what you forbid. Server logs say who actually turns up - and that is the only hard evidence in the whole audit.

Ask your host for access logs (access.log) or download them from the panel, and search for the agent names listed above. Three things matter: whether a given bot came at all, how often, and with what response code. A 200 is success; a run of 403s or 429s means something is rejecting or throttling it, even though robots.txt permits it.

A caution on interpretation: a bot in the logs does not mean citation, and its absence does not always mean a block - some engines query only their own index, built at some earlier point, and do not come back for fresh content on every question.

Step 6: what Search Console will and will not tell you

Search Console is a free source of truth about search, but on the AI side it has limits you need to know before drawing a conclusion:

• Clicks and impressions from AI answers in Google search are not broken out into a separate report in a way that would let you subtract them from the rest - they land in the search statistics. Google keeps adding breakdowns here, so check what your panel shows today rather than relying on an article from a year ago.
• Traffic from ChatGPT, Perplexity and other engines does not exist in Search Console at all - it is not Google search. You see it in your site analytics as visits from those domains, in the traffic source report.
• An indirect signal worth watching: rising impressions with flat or falling clicks on question-shaped phrases. That usually means the answer is being given above the results and the user has no need to click through.

For an AI visibility audit, then, Search Console is a supplement, not a foundation. The foundation is the hand-asked questions from step one. How to read the rest of the Search Console data, and how it differs from what you see in your own browser, I set out in the SEO audit checklist.

How to read the result

Three scenarios and three different decisions:

“The model does not know me.” The most common result at small companies and there is nothing dramatic in it. The cause is usually prosaic: not much content that answers questions, and not many places on the web where the company is described in words. The fix is years of work, not a week, and it overlaps for the most part with decent SEO.

“The model knows me but gets facts wrong.” The most urgent to fix, because it does active damage. Find out where the model gets the wrong information - usually an out-of-date page of yours, an old directory entry or a profile on a service you forgot about. Fix the source, not the answer; a model cannot be corrected directly.

“The model knows my industry and cites my competitors.” The most instructive result. Open the pages being cited and see how they differ from yours: do they carry concrete answers to questions, prices, data, a named author, an update date. That is a ready-made list of gaps - and the best content brief you will ever get for free.

Rhythm: once a quarter, the same sheet

AI visibility moves slowly and unevenly, so measuring weekly is a waste of time. A quarter is a good window: long enough to see the effect of work on content, short enough to react to a factual error.

One sheet, the same questions, the same engines, a column for the date. After a year you have four measurements and you can see the direction - which is the only thing that matters here. A single measurement tells you where you are; only the second one tells you whether the work is working.

FAQ

Can you “optimise” a company for ChatGPT the way you do for Google? Not in the sense of flipping a switch. AI engines draw on the same material as search engines: content on the web and what other people write about you. There is no tag or file that makes a model start recommending you. What actually influences it - and what no plugin will handle - I described in the piece on SEO and AI.

I asked ChatGPT about my company and it answered correctly. Does that mean I am visible? Not necessarily. Asking about the company name is the easiest test - the model can often simply visit your site and summarise it. The stake sits in the question with no name in it: “who does [service] in [city]”. If you are not there, you are not visible, however well the brand test goes.

Does blocking AI bots make sense? It depends what you sell. A content publisher has a reason not to hand material over for free. A service company that wants to be recommended shoots itself in the foot by blocking bots - models cannot recommend something they have no way of reading. If you block, block deliberately and by name, not wholesale.

How long does an audit like this take? The first round with a dozen or so questions across three engines is two to three hours including recording the results and checking robots.txt. Later quarters take an hour, because the sheet and the questions already exist. Server logs add half an hour, if you have access to them.

Does an SEO audit cover AI visibility? It should, though not every one does yet. The technical layer is shared - indexability, bot access, structured data, speed - and what differs is the measurement layer and the test questions. What a decent audit has to contain, and how to spot one that only looks like an audit, I set out in the piece on SEO audit cost and scope.

Want your company to show up in Google and in ChatGPT answers? I work on technical SEO and AI visibility from the code side, not the reporting side. Tell me which questions matter to you and I will quote it.