Before You Claim an AI Assistant Recommends You
Treat 'recommended by ChatGPT', or any assistant, as an advertising claim that needs evidence before it's published, not a screenshot you happened to get. The evidence: buyer questions that don't name you, repeated runs, the rate at which you were named or put first, the dates, and the stored answers. Then write the claim to match what you measured: 'named in 14 of 20 sampled ChatGPT answers to five buyer questions, September 2026' rather than 'ChatGPT recommends us'. Check the assistant owner's brand rules too; OpenAI's published guidelines tell third parties not to imply endorsement. In the US, the FTC expects a reasonable basis for objective advertising claims before they run. This isn't legal advice, and anything customer-facing should go past a lawyer.
A founder asks ChatGPT for the best tool in their category, sees their product named first, and wants it on the homepage by Friday. The screenshot is real. The claim it seems to support, that the assistant recommends them, is a statement about what the assistant generally does. One answer can't carry that.
Why one answer isn't a recommendation
The same prompt can produce different answers on different runs, for different users, in different places and on different model versions. 'Recommends' implies a stable behaviour. The only way to support a statement about general behaviour is to sample it: the same questions, asked repeatedly, with every result kept, including the runs where you didn't appear.
The substantiation worksheet
| Field | What to record | Why a reader would ask |
|---|---|---|
| Exact claim wording | The sentence as it will appear | The wording sets the evidence needed |
| Assistant and surface | Product, app or API, web search on or off | Answers differ by surface |
| Prompts | The buyer questions, none naming you | A prompt naming you tests recall, not recommendation |
| Runs | How many per prompt, over how many days | One run is an anecdote |
| Result | Named in X of Y; named first in Z of Y | The rate is the claim |
| Stored answers | Full text of every run, dated | Lets someone re-check it |
| Expiry | Date the claim is re-tested or removed | Assistants change and the claim ages |
| Approver | Who signed it off, including legal where needed | Someone owns a public claim |
Write the claim to match the rate
'ChatGPT recommends [Brand]' claims a general behaviour. 'In our September 2026 test, ChatGPT named [Brand] in 14 of 20 answers to five buyer questions' claims what you measured. The second is less exciting, and it's the one you can defend when a prospect runs the question themselves and gets a different answer, which some of them will. If the rate is low, there's no claim to make, and that's a finding too.
Brand rules and advertising rules
The assistant's name belongs to the company that makes it. OpenAI's published design guidelines describe the ChatGPT brand as OpenAI's property, and its guidance on using its logos tells third parties not to misrepresent their relationship with OpenAI, imply endorsement or confuse users about sponsorship. Check the equivalent guidance for any assistant you name.
Separately, the FTC's policy statement on advertising substantiation says advertisers need a reasonable basis for objective claims before they are disseminated, and that when an ad states its level of support, the statement's own examples being 'tests prove', 'doctors recommend' and 'studies show', the advertiser should have at least that level. 'ChatGPT recommends' reads a lot like the middle one. That's the US position; other countries have their own rules. We're not lawyers, and nothing here is legal advice.
Set an expiry date
Assistants change underneath published claims. ChatGPT's own release notes show it: the August 6, 2026 entry says GPT-5.6 Luna would become the default model for Free and Go users that week. A claim tested before that date describes what a different default model did for those users. Re-run the worksheet on a schedule, and take the claim down if the rate no longer supports it.
Being straight about our own position
We sell measurement, and a sampled rate is what we produce, so we benefit when people want evidence behind claims like this. The narrower point holds whoever runs the test, including you with a spreadsheet: if you can't state the rate, you can't make the claim.
FAQ
Can I post the screenshot without writing a claim? — A screenshot captioned 'look what ChatGPT said' still implies it's typical. If you share one, date it and say it's one answer.
What rate is high enough to say 'recommends'? — There's no standard, which is the argument for stating the rate instead of the verb. A reader can weigh 14 of 20. Nobody can weigh 'recommends'.
Does this apply to sales decks and cold emails? — Yes. They're marketing to a specific buyer, and the claim gets copied into procurement notes without your caveats attached.
Put this to work on your own website.
Fenn finds what your customers ask, drafts the articles and site fixes, and measures what ChatGPT, Claude, Gemini, Perplexity and Grok say about you — with every change waiting for your approval.