How to check whether your brand appears in ChatGPT and AI answers

You do not need a tool or a budget to find out whether AI assistants mention your brand. You need a fixed set of questions, a clean testing setup, and about an afternoon. Doing it by hand first is worth more than buying a tool, because it teaches you what the results actually mean before you automate them.

To check whether your brand appears in AI answers, run a fixed set of the questions your buyers ask through each assistant they use, in a clean session with personalisation turned off, and record for each one whether you are cited, mentioned, or absent, and how you are described. The important discipline is to check each surface on its own. ChatGPT, Gemini, Perplexity, and Google AI Overviews retrieve and compose answers differently, so a good result on one tells you almost nothing about the others. What follows is the method, surface by surface, and how to read what you find.

Build the question set first

Do this before you open a single assistant, because the questions decide whether the exercise is useful or just interesting.

You are trying to see yourself the way your buyers do, so write the questions the way they would ask them, not the way you would describe your own company. Three kinds of question matter. Category questions, where the buyer is looking for a type of provider or a way to solve a problem and does not know your name yet, such as "how do I stop losing direct bookings to the big travel platforms" or "who helps mid-sized hotels with AI search visibility". Brand questions, where they already know your name and want to know more, such as "what does [your company] do" and "is [your company] any good". And problem questions, the specific pains you solve, phrased without reference to any provider.

Write between fifteen and thirty of these and then freeze the list. The wording has to stay identical every time you run it, because the only way to know whether anything changed later is to compare like with like. If you reword the questions on the next run, you have thrown away the comparison. Keep the list in a simple sheet with a row per question, because that is where the results will go.

Set up a clean test

The single most common mistake is testing in your own logged-in account, where the assistant already knows you, has your history, and may have been primed by earlier questions. That tells you what the assistant says to you. You want to know what it says to a stranger who is your buyer.

So test clean. Start a fresh session with no prior conversation, because earlier messages influence later answers. Turn off memory and personalisation in the assistant's settings, or use a logged-out or private window where the surface allows it. Do not lead the model: ask the buyer's question as written, and do not follow up with "what about [your company]" until you have recorded the unled answer first, because the moment you name yourself you have changed the result.

Then run each question several times, not once. These systems are non-deterministic, which means the same question can produce different answers on different runs for reasons that have nothing to do with your site. One answer is an anecdote. What you are recording is how often you appear across several runs, which is a signal you can actually trust. Three to five runs per question per surface is enough to see a pattern.

Checking in ChatGPT

ChatGPT is usually the first surface people check, and it behaves in two quite different modes depending on whether it browses the web for the answer.

When it browses, it runs searches, retrieves live pages, and typically shows citations you can click, so you can see exactly which URLs it drew on. This is the mode that reflects your current site, and it is the one where recent changes can show up within weeks. When it answers from training alone, without browsing, it is drawing on what it learned when it was trained, which means it may describe you as you were a year ago, or not at all, and no change you make to your site this month will affect that answer until a future training run.

So when you test in ChatGPT, note which mode produced the answer. If citations appear, record the URLs, because that tells you which of your pages is doing the work, or which competitor's page is doing it instead. Ask your category questions first and record whether you appear unled. Then, separately, ask a direct brand question, because an assistant that cannot place you in the category but can describe you when named is telling you something specific: you are understood but not retrieved for the questions that matter.

Checking in Google AI Overviews

Google's AI Overviews are a different surface with a different catch: you cannot fully isolate them.

An AI Overview appears above the normal results for many queries, composed from pages Google has indexed. To check it, run your category and problem questions as ordinary Google searches and see whether an Overview appears, whether it names you, and which sources it links. The links matter, because they show whose content Google chose to build the answer from.

The catch worth knowing is that Search Console does not break out AI Overview impressions or clicks as a separate line. You cannot open a report that says "you appeared in this many Overviews". You can only observe the Overviews directly by running the searches, and infer the traffic effect indirectly from the gap between impressions and clicks at a stable ranking position, which I have covered in why your organic traffic is falling while your rankings hold. For this check, treat the Overview as something you observe by hand, query by query, and record the same way as the others.

Checking in Perplexity

Perplexity is the most transparent surface for this exercise, because it is citation-first by design: it shows its sources prominently for almost every answer.

That makes it the easiest place to see which pages, yours or a competitor's, are being used as sources for your category. Run your questions and read the citation list as carefully as the answer text. If a competitor is cited and you are not, open their cited page and look at what it does that yours does not: whether it states the answer plainly near the top, whether it is structured for extraction, whether it is a page that directly addresses the question. Perplexity will often show you the specific competitor page that is beating you, which is a gift, because it turns a vague problem into a concrete comparison.

Checking in the other assistants your buyers use

Do not stop at the ones you personally use. Check Gemini, and check Microsoft Copilot if your buyers are in a Microsoft-heavy environment, and check any assistant that is common in your specific market. Each has its own retrieval behaviour and its own defaults, and your visibility can differ sharply between them.

The principle is the same everywhere: clean session, unled question, several runs, record whether you are cited, mentioned, or absent, and capture how you are described and who is named instead of you. The point of covering multiple surfaces is that buyers are spread across them, and being strong on one while invisible on another is a real and common pattern. You want the full picture, not the flattering slice of it.

Scoring what you find

Now turn the answers into something you can act on. For every question, on every surface, record four things in your sheet.

First, the outcome, in three levels: cited, where you are named and linked as a source; mentioned, where you are named in the text but not linked; and absent, where you do not appear. The gap between cited and mentioned matters, because a mention without a link builds awareness but does not send the buyer to you.

Second, the description. When you are named, what does the assistant say you do, and is it accurate. A wrong or outdated description is its own problem, separate from being absent, and it points at an entity-understanding gap rather than a retrieval one.

Third, the competitors. Which companies are named instead of you, and how consistently. A competitor that appears across most of your category questions on most surfaces is not a coincidence, and their cited pages are a map of what good extraction looks like in your category.

Fourth, the source, where the surface shows it. Which URL was used, yours or theirs. This is the difference between knowing you have a problem and knowing where it lives.

What the results tell you to do next

The pattern across your sheet points at the fix. If you are absent across most surfaces even on questions that are practically about you, the problem is upstream, at retrieval or entity understanding, and no amount of content will help until that is resolved. If you are mentioned but described wrongly, you have an entity-consistency problem to fix across your site and profiles. If a competitor is consistently cited from a specific page for questions you also answer, but only in prose, you have an extraction problem on your own pages. And if you are simply outgunned on general authority questions, you are looking at corroboration, the slowest gap to close.

This hand-run check is genuinely useful, and I would rather a team did it than nothing. Its limits are worth naming, though. It covers the prompts you thought of, it is run by hand so it is easy to do inconsistently, and it captures a moment rather than a trend. When you want it done with fixed controls, enough runs to be statistically meaningful, and a proper baseline you can measure against later, the rigorous version of the method is set out in how to measure whether your brand appears in AI answers.

The mistakes that make a check misleading

Most checks that produce a wrong conclusion fail in one of a few predictable ways, and all of them are avoidable once you know to watch for them.

Testing while logged in is the most common. Your own account knows you, and the assistant may lean toward you for reasons that have nothing to do with what a stranger sees. Always test in a clean or logged-out state. Leading the model is the next: the moment you ask "what about my company", you have told the assistant the answer you want, and it will often oblige. Record the unled answer first, every time. Running each question only once is the third, and given non-determinism it turns noise into a false signal, so run several times and record the frequency rather than the last thing you saw. And rewording the questions between runs quietly destroys the comparison, because you are no longer measuring the same thing. Freeze the wording and change nothing but the date.

There is a subtler error too: treating a single surface as the whole picture. Your buyers are spread across assistants, and being strong on the one you happen to use while invisible on the one they use is a real and common pattern. The check is only honest if it covers the surfaces your buyers actually use, not the ones convenient for you.

The takeaway

Checking whether AI assistants mention your brand is a manual exercise anyone can run: fixed questions, clean sessions, several runs, one surface at a time, scored into a simple sheet. Do it yourself first, because it teaches you what the results mean and where your gap actually is, before you spend on tools or fixes.

When the check tells you there is a problem and you want it diagnosed properly, with controlled conditions across every surface your buyers use and a clear read on which fix will move the outcome, that is what an AI search visibility audit does at scale.

This article is part of the SEO in the AI Era: The Complete Guide guide.