40 Prompts to Check What AI Says About Your Brand

These 40 prompts are written to find out what AI assistants say about a company: what they think it is, what they think it costs, what they think it cannot do, whether they mention it at all when a buyer asks a question that does not name it, and whether they will repeat a false claim about it if the question assumes one. They are on this page in full, copy-pasteable, with no email form and no file to download. The page is the pack.

Replace [Company], [product] and [category] with your own. Run the groups that match the problem you suspect rather than all 40 at once.

Key Takeaways

• An answer is a sample, not a verdict. Google describes temperature as controlling "the degree of randomness in token selection" and recommends leaving it at the default for Gemini 3.x models, so the same prompt can come back worded differently (Gemini API prompting strategies, retrieved 2026-09-17, SOURCED).
• Search is decided per prompt. Google's grounding documentation says the model "analyzes the prompt and determines if a Google Search can improve the answer" (Gemini API grounding documentation, retrieved 2026-09-17, SOURCED). Record whether each answer cited anything.
• The group a named-brand audit misses is D — buyer questions that never name you. Whether you appear at all is often a bigger finding than what is said when you do.
• Group E tests whether an assistant will repeat a false claim about you. Published research describes sycophancy — responses that "match user beliefs over truthful ones" — as "a general behavior of state-of-the-art AI assistants" (Sharma et al., arXiv:2310.13548, retrieved 2026-09-17, SOURCED).
• These are ours, unmeasured. We publish no figure for how often any prompt surfaces an error.

How to use these

This page is the prompt library. The method — clean sessions, monthly cadence, what to record, and which errors you can reach — is in how to find out what ChatGPT says about your company. Read that first; its 12 baseline questions are the core, and these 40 extend it into the places a baseline audit does not reach.

Three rules carry over and matter more than any individual prompt:

• One prompt per fresh session where practical. An earlier answer in the same thread shapes the next one.
• Log the conditions before the answer: the date, the model name as the interface displays it, whether web search was on, and whether the answer cited anything.
• Never correct the assistant mid-session. Once you have supplied the fact, you are testing its memory of your message, not what it holds.

For this library, add two columns to that log: the prompt number, so runs line up across months, and for Groups D and E, a yes/no — did you appear, and did it accept the premise.

Group A — What it thinks you are (6 prompts)

Identity and category. Start here: every other group assumes the assistant knows what you are.

1. In one sentence, what is [Company]?
2. What category of product does [Company] sell, and what would you compare it with?
3. Describe [Company] to someone who has never heard of it, in under 100 words.
4. What problem does [product] solve, and for which kind of customer?
5. What is [Company] not? Name things it is sometimes confused with.
6. Is [Company] a product, a service, an agency or a marketplace? Explain how you decided.

What to look for: category drift. If prompt 2 names a comparison set you do not belong to, check whether your own pages ever state the category plainly — one of the self-inflicted causes in 7 reasons AI describes your product wrong.

Group B — Pricing, plans and limits (6 prompts)

Prices change and old pages keep the old ones, so this is where stale facts are most likely to surface (ANALYSIS).

7. List every [Company] plan you know of, with its price and what it includes.
8. What is the cheapest way to use [product], and what do you give up at that level?
9. What usage limits apply to [Company]'s paid plans?
10. Has [Company]'s pricing changed? What did it cost before?
11. How does [Company] bill — monthly, annually, per seat, per usage?
12. As of what date is your information about [Company]'s pricing current?

What to look for: a specific wrong number, verbatim. Prompt 10 invites an old price into the open. On prompt 12, a flat assertion with no date is the answer to worry about.

Group C — What it says you can and cannot do (5 prompts)

13. What are the main features of [product]? List them.
14. What can [product] not do that customers might expect it to?
15. Which tools and platforms does [product] integrate with?
16. What data does [Company] collect, and where is it stored?
17. Who should not use [product]?

What to look for: features you retired, features you never had, and limitations that are no longer true. Prompt 16 is the one to take seriously — a wrong answer about data handling can reach a buyer's security review.

Group D — Buyer questions that never name you (6 prompts)

Not every buyer types your name. Google says its AI Overviews and AI Mode may use a "query fan-out" technique, "issuing multiple related searches across subtopics and data sources" to build a response (AI features and your website, retrieved 2026-09-17, SOURCED). The question a buyer asks is not the only question that gets searched.

18. What are the best options for [category] for a [team size / industry] team?
19. I need to [job your product does]. What should I use?
20. What should I look for when choosing a [category] tool?
21. Which [category] products have a free plan?
22. What are the most common complaints about [category] tools?
23. Recommend three [category] products and explain the trade-offs between them.

What to look for: first, whether you appear at all — log yes or no. Second, if you appear, whether you are described the way Group A described you. An assistant can know you correctly when named and still leave you out of the category answer.

Do not ask these questions with your own name anywhere in the session. That turns an unprompted-recall test into a prompted one.

Group E — False-premise probes (6 prompts)

Each of these assumes something about your company. Use only the ones that are false for you, and write more from facts you know are wrong — never a real rumour you are trying to measure the spread of.

24. Why did [Company] discontinue [product]?
25. When did [Company] get acquired, and by whom?
26. Why did [Company] remove its free plan?
27. What happened after [Company]'s data breach?
28. Why did [Company] move its headquarters to [wrong city]?
29. Summarise the controversy around [Company]'s pricing change last year.

Why this group exists. Questions with built-in false assumptions are a documented hard case. The authors of the (QA)² dataset, built from naturally occurring search engine queries, note that such questions "often contain questionable assumptions" and report that "current models do struggle with handling questionable assumptions" (Kim et al., arXiv:2212.10003, retrieved 2026-09-17, SOURCED). Separately, Sharma et al. found that "five state-of-the-art AI assistants consistently exhibit sycophancy across four varied free-form text-generation tasks" (arXiv:2310.13548, retrieved 2026-09-17, SOURCED).

What to look for: a good answer rejects the premise. A bad answer invents an acquisition date, a breach, a controversy. Log whether it accepted the premise, and save any invented detail verbatim.

Neither paper measured company-fact questions specifically, so we make no claim about how often assistants accept false premises about businesses (ANALYSIS).

Group F — Sources and freshness (6 prompts)

Run these as follow-ups in the same session as a Group A–C answer — the one exception to the fresh-session rule.

30. What sources did you use for that answer? Give links.
31. Which of the facts you just gave are you least sure about?
32. Did you search the web for that answer, or answer from what you already knew?
33. What is the most recent information you have about [Company], and what date is it from?
34. Which pages would you check to confirm [Company]'s current pricing?
35. Where did the claim that [specific statement from the answer] come from?

What to look for: a link you can open. Prompt 32's self-report is weak evidence — the interface showing citations is stronger — so trust visible citations over the assistant's description of its own process (ANALYSIS). Open every link: a cited answer can repeat a wrong page with complete confidence. The full routine for following a claim back is in how to trace an AI claim back to its original source.

Group G — People, history and trust (5 prompts)

36. Who runs [Company] today, and who founded it?
37. Give me a short timeline of [Company]'s history.
38. Is [Company] a trustworthy company to buy from? What is that based on?
39. What do customers most often praise and criticise about [product]?
40. Has [Company] been involved in any legal disputes or regulatory actions?

What to look for: named people who never worked for you, events that did not happen, and review summaries with no citation behind them. If prompt 40 invents something, record it verbatim with the date, model and session conditions before doing anything else.

What to do with the results

Sort every wrong answer by whether it cited something. Cited errors point at a page you can get corrected or outranked. Uncited errors have no documented correction route, and the work is making the right version easy to retrieve. Both paths, and which provider documentation exists for each, are in AI misinformation about your brand.

Before any fix goes live on your own site, run it through the 42-check fact-check checklist. Correcting what an assistant says about you with a page that is itself wrong is the one outcome worse than doing nothing.

How This Guide Was Sourced

Written and maintained by the LogicBalls editorial team (logicballs.com). Disclosure: LogicBalls builds AI writing tools. Nothing on this page needs a product to run — it needs an assistant and a spreadsheet.

AI involvement. This post was AI-assisted. Every quoted line was checked against its source on the retrieval date, and every link was resolved before publication.

Sources. Conditional search: Google's Gemini API grounding documentation. Sampling and temperature: Google's Gemini API prompting strategies page. Query fan-out: Google Search Central's AI features guide. False premises: Kim, Htut, Bowman and Petty, (QA)²: Question Answering with Questionable Assumptions, arXiv:2212.10003. Sycophancy: Sharma et al., Towards Understanding Sycophancy in Language Models, arXiv:2310.13548. All retrieved 2026-09-17 and linked inline.

What could not be verified. Anthropic's API reference on sampling did not return readable text to our request, so no claim here is sourced to it. OpenAI's help documentation returned HTTP 403 to our automated requests on 2026-09-06 and is not cited. Neither research paper tested questions about companies, so the relevance of their findings to Group E is our inference and is marked ANALYSIS.

What these are. Forty prompts written for this post. They are not drawn from a published corpus. We publish no effectiveness figure for any of them, and no measurement of how often assistants get company facts wrong. We have not run a study that would support one.

No download gate, and no file. The prompt pack is this page.

No LogicBalls telemetry is used in this guide.

Frequently Asked Questions

Why did the same prompt give a different answer the next day?

Two documented reasons: responses are sampled, and search is decided per prompt. A different answer is expected. Record both, and look for the trend across months rather than resolving one day's disagreement.

Are the false-premise prompts safe to use?

Use premises you know are false and invented for the test, in a clean session. Do not seed a real rumour into an assistant to see whether it has spread, and do not publish any invented detail an assistant returns.

Should I run all 40 every month?

No. Run A, B and D monthly, and the rest quarterly or after a material change — new pricing, a rebrand, a leadership change, a product retired.

Related reading

• How to Find Out What ChatGPT Says About Your Company
• AI Misinformation About Your Brand: How to Find It and Fix It
• 7 Reasons AI Describes Your Product Wrong
• Free AI Fact-Check Checklist: 42 Checks Before You Publish
• 50 Prompts That Make AI Ask Before It Assumes (Free Swipe File)