Does ChatGPT recommend your business? You cannot answer that by typing one question and reading one reply. The answers change from run to run. The honest check is a sample: the questions your buyers ask, asked several times in ChatGPT, Perplexity and Gemini, and reported as a share with a range.
Why one answer proves nothing
A founder asks ChatGPT for “the best fractional CMO in the UK”, sees a competitor named, and concludes the business is invisible. Or sees their own name and relaxes. Both readings are wrong for the same reason. One answer is one draw from a distribution.
Language models sample the next word rather than look it up. OpenAI says plainly that its models are non-deterministic by default, and that even its reproducibility setting is best effort. Three more things move the answer:
- Search. When the engine searches the web, the pages it retrieves vary. Google documents how Gemini grounds answers in Google Search and returns the queries and sources it used. Different queries, different sources, different shortlist.
- Context. Memory, custom instructions, location and the earlier turns of a chat all shape the reply. Your own account knows you run the business.
- Wording. “Best fractional CMO” and “who can fix our HubSpot reporting” are different questions with different answers.
So the useful question is not “does it recommend me?” It is “in what share of the answers my buyers would see am I named, and is that share moving?”
Write the questions buyers ask
Start from the buyer, not the brand. Write ten to fifteen questions someone would ask before they know your name. Three kinds cover most of it:
- Category. “Who are the best fractional CMOs for B2B SaaS in the UK?”
- Problem. “Our pipeline dropped after we cut paid search. Who can diagnose it?”
- Comparison. “Fractional CMO or agency for a £3m B2B company?”
Leave out questions that contain your name. Being named when someone asks about you proves only that the engine can read. Keep the list fixed once you start: changing the questions each month makes the trend meaningless.
Run the check
The method is dull on purpose. That is what makes it repeatable.
- Use a clean session each time: signed out, or a temporary chat with memory off. Note the country you run it from.
- Ask every question three times in each engine: ChatGPT, Perplexity and Gemini. Fifteen questions, three engines and three runs gives 135 answers.
- For each answer, record four things: were you named, where in the list, which competitors were named, and which sources were cited.
- Store the date and the model or mode shown, because engines change underneath you.
The cited sources are the most useful column. Perplexity and ChatGPT search show their links, and Gemini’s grounding returns them. If the same directory, review site or trade article keeps appearing, that is where the engines learn who is in your category.
Read the result honestly
Count the answers that name you and divide by the answers collected. Then put a range on it, because a small sample is noisy. The standard tool is a Wilson score interval, which behaves sensibly when the counts are small. The arithmetic is worth seeing once:
- Named in 3 of 20 answers: 15%, but the 95% range runs from about 5% to 36%.
- Named in 7 of 45 answers: 16%, with a range of about 8% to 29%.
- Named in 1 of 1 answer: 100%, with a range of about 21% to 100%.
The last line is the founder who checked once. It tells you almost nothing. The rule I hold to: report “named in 7 of 45 answers, likely 8% to 29%”, and only claim a change when this month’s range no longer overlaps last month’s. Anything less is reading tea leaves. The full rules are in how AI visibility is measured.
Read the engines separately. They search different indexes and favour different sources, so a blended number hides the engine where you are absent. And never write “ChatGPT does not recommend us” on the strength of a sample. Write what you measured.
What moves the number
Two separate things sit underneath the share, and they need different work.
Can the engines read you? Your pages need to be crawlable, indexed and clear about who you serve. That is technical readiness, and it is covered in llms.txt, AI crawlers and structured data. It is necessary and rarely enough.
Do credible sources describe you? The cited-sources column tells you which pages shape the shortlist. Being named there, accurately, is slower work: useful writing on your own site, a clear category description, and real mentions in the publications your buyers read. It is the same work that makes B2B SEO create pipeline, measured in a new place.
The honest limitation: this check measures what the engines say, not what buyers do. Track AI referral visits and enquiries alongside it. A rising share with no enquiries is a vanity number. If the measurement and the sequencing need an owner, that is the kind of system I install as a fractional CMO.
FAQ
How do I check if ChatGPT recommends my business?
Write ten to fifteen questions a buyer would ask before they know your name, ask each one several times in ChatGPT, Perplexity and Gemini with a clean session, and record whether you are named and which sources are cited. Report the share of answers that name you, with a range, not a yes or no.
Why does ChatGPT give a different answer each time?
Language models sample their words, so the same prompt can produce different text. OpenAI's own documentation says its models are non-deterministic by default. When the engine searches the web, the pages it finds can also change between runs, and memory, location and settings change the context.
How many answers do I need before the result means anything?
More than most people collect. Three mentions in twenty answers is consistent with a true rate anywhere from about 5% to 36%. Collect at least forty to sixty answers per engine before you compare months, and only claim a change when the ranges stop overlapping.
Should I check signed in or signed out?
Signed out, or in a temporary chat with memory off. Your own account knows who you are and what you sell, so it flatters the result. Your buyers' accounts do not.
Key takeaways
- One answer is one draw. It proves nothing either way.
- Fix the buyer questions, then sample: several runs per engine, clean sessions.
- Report a share with a range, and claim a change only when the ranges separate.
- The cited sources show where to work; enquiries show whether it mattered.
- Next step: read how AI visibility is measured, the full method behind these numbers.
