How to check if ChatGPT recommends your business

Open a fresh chat, ask the question your customer would ask, and see whether your name comes back. That is the method. The complication is that the answer moves — between people, between phrasings, between one Tuesday and the next — so a single run tells you very little. Here is how to run the check so the result means something, and how to read it without flattering yourself.

The short version

1. Open a new chat. Sign out, or use a temporary chat, so your own history is not in the room.
2. Ask the question a customer would ask, in their words. Do not mention your business.
3. Tell it to search the web, so you are testing what it can find rather than what it happens to remember.
4. Write down three things: whether you were named, how you were described, and which sources it cited.
5. Repeat — different phrasings, different days — and date every run.

Steps 3 and 5 are the two people skip, and they are the two that turn an anecdote into a reading.

Which prompts to run

Ask what your customers actually type, not what your marketing wishes they typed. Five shapes cover most of it.

• Category and city. 'Best [category] in [city]'. The blunt one, and the one a competitor is most likely to already own.
• Service-specific. 'Best [category] in [city] for [specific service]'. Narrower questions are easier to win and usually sit closer to a booking.
• Situation-first. 'I [problem] — who should I see in [city]?' People describe the problem far more often than they name the category.
• Comparison. '[Your business] or [competitor] for [job]?' This tells you what difference the engine thinks there is between you.
• Brand check. 'Is [your business] any good?' and 'Does [your business] do [service]?' Verification is a different job from discovery, and it runs on your reviews and listed facts rather than on your prose.

Those are close to the shapes our own batch work used. The queries are recorded per batch: 'best [practice] lawyer in Boulder' run across six practice areas, 'best craft beer in Boulder', 'best med spa Boulder', 'best CrossFit Longmont'. Pick four to six and keep them fixed. A small set you re-run beats a wide set you run once, because the whole value is in the comparison over time.

Vary the wording, the location and the session

Three variables move the answer, and you want all three deliberately under control.

Wording. 'Best', 'top-rated', 'who should I hire', 'recommend me a' — these are not one query. Run at least two phrasings of the same intent before concluding anything.

Location phrasing. 'In Boulder', 'near me', 'near [landmark]', 'in [neighborhood]', '[city] area'. 'Near me' behaves differently from a named city, because the engine has to infer a location — and where it thinks you are is not where your customer is.

Session. A fresh chat, signed out or temporary, is the closest you get to a stranger's view. A thread in which you have already mentioned your business is worthless as a test: you have handed it the answer. If the assistant carries memory between chats, that memory is in the result too.

Non-determinism, and what a small sample proves

The same prompt does not reliably produce the same answer. It varies by person, by session and by day, and it moves again whenever the underlying model or index changes. That is not a flaw in your test. It is a property of the thing you are testing, and the fix is to size your conclusion to your sample.

• One run of one prompt. Proves only that this output was possible once. Do not act on it, and do not send it to your team as news.
• Three runs of the same prompt, in fresh sessions. Now you can say something weak but real: named every time, never named, or inconsistent. All three are useful, and the third is the most common.
• Four to six prompts, three runs each, logged with dates, repeated monthly. A trend you can act on. It is still a small sample, and you should say so out loud whenever you report it.

We hold our own work to the same limit. The Front Range AI Visibility Index graded 70 businesses from answer-engine output captured over three working days in June 2026, and the page says plainly that a re-run would not reproduce every row.

What to write down

One row per run: date, engine, whether the web was searched, the exact prompt, named yes or no, roughly where in the answer, the words used to describe you, which competitors were named, and which sources were cited.

The last two columns are the ones that turn a score into a to-do list. The competitors tell you who the engine already trusts in your category. The cited sources tell you which pages and profiles it is actually reading — which is usually a shorter and stranger list than owners expect.

Screenshot anything surprising. Answers move, and 'I'm sure it said something different last month' is not evidence.

ChatGPT is not one check

It matters whether an answer came from the model or from the live web, because the two fail for different reasons and you fix them differently.

• Answered from the model. No sources, no links. You are seeing what the model absorbed in training, which is old by definition and which nothing you publish this week will change. A business that is small, new, or simply not written about much will not be there.
• Answered from the web. The answer carries linked sources. This is the one that reflects your site and your listings as they stand now, and the one worth re-running after you fix something.

Tell them apart by looking for citations, and get the second kind by asking it to search the web.

The agent that matters is the one behind that second mode: OAI-SearchBot, OpenAI's crawler for ChatGPT search. GPTBot is the training crawler — blocking it opts you out of model training, not out of ChatGPT search. ChatGPT-User fetches a page because a person asked for it in the moment, and OpenAI notes robots.txt rules may not apply to it. If you check one line in your robots.txt, check that OAI-SearchBot is not disallowed. The crawler guide covers the rest.

The same check, on the other engines

• Perplexity. The easiest to read, because it shows its sources alongside the answer. Run the identical prompts and note which of your pages or profiles it pulls. PerplexityBot is its index crawler.
• Google AI Overviews and AI Mode. Run the query in Google itself. AI Overviews is the generated summary above the results; AI Mode is the conversational surface. Both are generated from Google's ordinary Search index, so Googlebot is the crawler that matters — and your Google Business Profile carries weight here that it does not carry everywhere.
• Claude. Can search the web and cite what it used. Claude-SearchBot is the search agent; ClaudeBot is the training crawler.

Do not over-read the differences between engines. They draw on different indexes and they will disagree with each other. But a business that is clearly described and consistently listed tends to do better in all of them, and we have not measured the size of any per-engine gap, so we will not sell you one.

Reading the result

Four outcomes, four different jobs.

• Named and described accurately. You are winning. Protect it — keep the listings current, keep reviews coming, and re-run the log monthly.
• Named, but the details are stale or wrong. A data problem, and the most fixable of the four. Chase the wrong fact back to its source: your site, your profile, or a directory nobody has updated since 2019.
• Not named, while competitors are. The gap that costs work. Look at which competitors come back and which sources get cited — together they describe what the engine trusts and what you have not given it.
• Nobody named, or a generic non-answer. An open lane. The engine has nothing specific enough to recommend in your category, and the first business that becomes legible takes it.

What a Front Range result usually looks like

Of the 66 owner-led Front Range businesses we audited and could sell to, exactly half graded Invisible, Partial or Compromised — 70 businesses in total, across seven batches covering breweries, restaurants, fitness studios, salons and spas, and law offices. So a weak result is ordinary rather than shameful. Those grades read the signals on the site and the profile, not the quality of the work, which is why a bad one is a to-do list and not a verdict.

What this check cannot tell you

• How often the question is asked. Being named in a query nobody types is not visibility.
• What other people see. Your result is one draw from a distribution shaped by location, account and history — yours, not your customer's.
• Why you were or were not named. The check records the outcome, not the cause. Our own scans recorded which businesses the engines named, not which URLs they cited; that is the more valuable question and we have not answered it yet.
• Whether a fix has landed. Indexes update on their own schedule. Re-run monthly, not daily, and do not read one bad Tuesday as a regression.
• Anything about the mechanism. No engine publishes its weighting. Anyone selling you a ranked list of ChatGPT ranking factors is guessing, and so would we be.

What the free scan does instead

The manual check tells you what an engine said. It does not tell you what to change. The free scan is the other half of the picture: it reads the page you give it plus robots.txt, sitemap.xml and llms.txt and grades the site-side signals that decide whether an engine could name you — 26 scored checks across eight dimensions, in seconds. It does not query ChatGPT, Perplexity or Google, and it cannot see your Google Business Profile. Those are separate jobs, and we would rather say so than let a score imply a check it never ran.

Run both. The log tells you whether you have a problem. The scan tells you where it probably lives — and if the answer is 'the engine cannot read who you are', start with structured data and what ChatGPT pulls from.

Where we apply this

The trades and cities our research covers, and the service behind it.

More in GEO

Find out what an answer engine can actually read about you.

The free scan reads your site the way an answer engine's crawler does and shows you the gaps to close first. It does not query ChatGPT, Perplexity or Google — that check is the human deep-dive.

Run the free scan How we do GEO