To check what AI says about your brand, do not ask it about your brand. Open ChatGPT in a fresh chat and ask the question your buyer asks — “best payroll software for a 40-person Indian startup”, “which D2C skincare brands actually work” — and see whether your name is in the answer. That one inversion is most of the method. The rest is doing it cleanly enough that the results mean something, and doing it again next month.
This guide is the manual version, end to end: which questions to write, how to ask them so the answers are uncontaminated, what to record, and what to do with each finding. It costs nothing except a few hours a month. If you want the background on why AI answers now decide shortlists, start with the GEO guide and come back.
Why “what do you know about my brand” is the wrong prompt
It is the prompt everyone tries first, and it measures the wrong thing. When you put your brand name in the question, the engine retrieves pages about your brand and summarises them. Your website, your LinkedIn, a press mention or two. Nearly every brand with a functioning site passes this test, which is exactly why it feels reassuring and proves nothing. You did the discovery for the model.
No buyer asks that question. Buyers ask category questions — best X for Y, X versus Z, is X worth it — and the engine has to come up with names on its own. Whether it comes up with yours is the finding that matters. A brand can have a perfect summary when asked by name and be absent from every answer where the buyer has not heard of it yet, which is precisely the moment recommendations get made.
The by-name prompt has one legitimate use: an accuracy check. Ask it once per engine and read the answer for wrong facts — old pricing, a discontinued product still listed, the wrong city. Note anything false. Then move on to the questions that measure visibility.
Step 1: Write the questions your buyers actually ask
Five to ten questions is enough. The goal is not coverage of every phrasing; it is a fixed set you can re-run identically for months. Pull them from places where buyers speak in their own words:
- Sales calls and inbound emails. The sentence after “we were looking for…” is usually a usable prompt verbatim.
- Search Console. Your longest queries — four words and up — read like the questions people now type into ChatGPT instead.
- Comparison intent. “[market leader] alternatives” and “X vs Y, which is better for [situation]” are asked constantly and produce brand lists almost every time.
Keep the constraints in. Real prompts carry a budget, a city, a team size, a use case: “best branding agency in Bangalore for a D2C launch”, not “best branding agency”. Engines answer constrained questions with shorter, more decisive lists, and those are the lists you need to be on. If your buyers are in India, say so in the prompt, because the engines will happily answer with five American brands otherwise.
Step 2: Ask each engine fresh, with no history
This is the part most people get wrong without noticing. ChatGPT’s memory and chat history personalise its answers. If you have spent six months asking it to draft your marketing copy, it knows your brand warmly and will slip it into answers a stranger would never see. You are not your buyer; your buyer’s account has never heard of you.
- Use a temporary chat in ChatGPT, or turn memory off for the session. Logged out is even cleaner.
- One question per chat. Follow-ups inherit the context of everything above them and stop being comparable across engines.
- Ask the same question on ChatGPT, Claude, Perplexity, and Gemini separately. They retrieve from different indexes and disagree often — the disagreement is data, not noise.
- Ask each question two or three times per engine. Models sample their output, and web-searching engines can pull different pages on each run, so a single answer is one draw from a distribution. What you want is the share of draws that name you.
Copy the full answer text out every time, including the sources the engine cites. Your impression of an answer decays in a day; the text does not.
Step 3: Record who gets named, in what order, and what gets cited
A spreadsheet is fine. One row per answer, with columns for: date, engine, question, every brand named in order, your position (or “absent”), the phrase used to describe you, the URLs cited, and whether your own domain is among them. That last column is easy to skip and is the most diagnostic one on the sheet — it tells you whether engines ever read you directly or only hear about you from third parties.
Resist the urge to editorialise while recording. “Mentioned but lukewarm” is a judgement; “listed fifth of six, described as a budget option, no citation” is a record. You can form judgements at review time, against the whole sheet.
Step 4: Re-run the same set on a schedule
One check is a photograph of weather. Models get updated, retrieval indexes refresh, a new roundup gets published and enters the sources, and the answer changes with no warning. What you want is climate: the same questions, the same engines, re-run monthly at minimum, weekly if you are actively doing visibility work and want to see whether it lands.
Do not improve the questions between runs. A prompt set that changes each month measures nothing; the value of the boring fixed set is that month four is comparable with month one.
What to look for in the answers
Five things, in rough order of importance:
- Presence. In what share of answers are you named at all? Early on, this is the only number that matters.
- Position. First or seventh? Answer lists are read the way search results always were — the top carries the intent, the tail is a courtesy.
- Description. “The strongest choice for small teams” and “also available” are both mentions. They are not the same mention.
- Citations. Which pages did the engine actually read? If the same three roundups appear behind every answer in your category, those pages are the electorate. Whether you are on them decides most of the rest.
- Wrong facts. Pricing, location, features, founding year, products you discontinued. Note each error together with the citation that carried it, because the citation is where you will fix it.
What to do with each finding
| Finding | What it usually means | What to do |
|---|---|---|
| Absent everywhere | The corpus about you is thin; engines have nothing to retrieve | Get listed in the roundups and directories that already rank for your buyer questions — the playbook in how to rank in ChatGPT |
| Named, but last | You appear in few sources, or with vague descriptions | More independent mentions; sharper, more concrete positioning in the sources that already list you |
| Described wrongly | A stale or incorrect source is being retrieved | Fix the cited source first, then publish the correct fact plainly on your own site with structured data |
| Named, own site never cited | Third parties carry your story; your pages are not extractable | Answer the buyer question verbatim on your own pages — see the AEO guide |
| Strong on one engine, absent on another | Different retrieval indexes, or a crawler you block | Check robots.txt for that engine’s crawler; broaden beyond the sources that only one engine reads |
Notice that almost none of the fixes happen inside a chat window. You cannot argue an engine into recommending you; you change what it reads. That is the core difference between this discipline and classical search work — unpacked properly in GEO vs SEO and AEO vs GEO.
The mistakes that make the check worthless
- Asking about your brand by name. Covered above, and worth repeating because it is the default instinct. It tests summarisation, not recommendation.
- Checking once. A single run tells you what one sample of one week looked like. Every conclusion drawn from it is fragile.
- Judging from one engine. ChatGPT, Claude, Perplexity and Gemini disagree with each other routinely. A brand that checks only ChatGPT can be invisible on the engine its actual buyers use.
- Checking from your own logged-in account. Memory and history tilt the answers toward you. The flattering result is an artefact.
- Treating one answer as the answer. Sampling variance is real. Named in one run out of three is a very different position from three out of three, and you only see that by asking more than once.
- Rewriting the questions every month. The set is only useful because it is fixed.
The automated version
Everything above works. It also decays: the spreadsheet that gets filled diligently in month one is abandoned by month three, because twelve questions times four engines times three runs is a tedious morning, every month, forever. That is the honest case for tooling — not that the manual check is wrong, but that the schedule is the whole value and schedules are what humans drop.
BrandAuditor’s AI visibility checker runs this exact process: it writes the buyer questions for your category, asks ChatGPT, Claude, Perplexity and Gemini fresh, and returns the leaderboard of brands each engine names, your rank and share of voice, whether your own site is cited, and a citation-gap comparison of the pages the engines trusted against your own page — with the raw answers attached. The lite check is free: one buyer question on ChatGPT, no card, and you will know in a minute whether you are in the answer or not.
Frequently asked questions
How do I see what ChatGPT says about my brand?
Do not ask ChatGPT about your brand by name. Open a fresh chat with no history, ask the question your buyer would ask (for example 'best CRM for a small agency in India' or 'which running shoe brands are worth it'), and see whether you are named in the answer, in what position, and with what description. Asking 'tell me about [brand]' only tests whether the model can summarise pages about you, which almost every brand passes. The commercial question is whether it names you unprompted.
Why does ChatGPT give a different answer each time I ask?
Two sources of variance: language models sample their output rather than computing one fixed answer, and engines that search the web can retrieve a different set of pages on each run. So the same prompt can produce a different brand list, or the same list in a different order. The fix is to ask each question more than once and track the share of answers that name you, instead of treating any single answer as the verdict.
Does asking ChatGPT about my own brand change its answers?
It can change the answers you see, not the answers strangers see. ChatGPT's memory and chat history personalise responses within your account, so an account that has discussed your brand for months gets warmer answers than a fresh one. Use a temporary chat or turn memory off when checking, and remember that your buyers are asking from accounts that have never heard of you.
What should I do if AI says something wrong about my brand?
Trace it before you try to fix it. If the answer cites sources, the wrong fact usually came from one of them: an outdated pricing page, a stale directory listing, an old news article. Correct the source, then make sure your own site states the true fact plainly and in structured data, so the next retrieval has something accurate to read. If there is no citation, the error is in the model's training knowledge, which you counter the same way: consistent, correct facts published where the engines read.
How often should I check my brand's AI visibility?
Monthly at minimum, with the same fixed set of questions each time. AI answers drift as models update and retrieval indexes refresh, so a one-off check is a snapshot that expires. If you are actively working on AI visibility (pitching roundups, fixing pages), check weekly so you can see whether the work is landing.
Is there a free way to check my brand's AI visibility?
The manual check in this guide is free: write your buyer questions, ask each engine in a fresh chat, and record the results in a spreadsheet. BrandAuditor also runs a free lite check (one buyer question on ChatGPT, no card required) which shows you the brand list and where you sit in it.