What does ChatGPT know about your company?

In short: Test what ChatGPT knows about your company with a dated canonical fact sheet and 20 fixed prompts covering identity, category, audience, offer, proof, and alternatives. Score correct facts, omissions, ambiguity, stale claims, and unsupported details separately. Recognition is the starting point. Accurate reconstruction requires more evidence.

This is a public-information reconstruction test. Use it to answer the business question “What does ChatGPT know?” with observable evidence. Keep confidential company data out of the prompt panel.

Create the canonical company fact sheet

Before prompting, document the facts against which answers will be reviewed. Include:

• legal name, trading name, former names, and parent relationship
• official domain, locations, and service regions
• current products and discontinued products
• category and the problem each offer solves
• intended customers and material exclusions
• current capabilities, integrations, pricing scope, and limits
• dated primary evidence, certifications, or public proof
• direct alternatives and complementary categories

Every fact needs an authoritative URL, effective date, owner, and scope. If your own pages conflict, mark the fact unresolved rather than choosing the most favorable version. Leaf’s content audit for AI search can help reconcile duplicate claims.

Run this 20-prompt entity test

Replace bracketed fields without adding confidential information. Freeze exact wording for later retests.

Identity and relationships

1. What is [Company], using [domain] to disambiguate it?
2. What products or services does [Company] currently offer?
3. Is [Product] a company, product, or service, and who owns it?
4. Is [Company] connected to any similarly named organizations?

Category and problem

5. Which category best describes [Company], and why?
6. What customer problem does [Company] claim to solve?
7. Which adjacent categories should not be confused with [Company]?
8. What terminology does [Company] use differently from competitors?

Audience and fit

9. Who is [Company] designed for?
10. Which organization sizes, regions, or sectors does [Company] serve?
11. Who may be a poor fit for [Company]?
12. What buyer requirements would determine whether [Company] is suitable?

Offer and limitations

13. What are [Company]’s current main capabilities?
14. What important limitations or eligibility conditions apply?
15. What is publicly known about [Company]’s pricing model?
16. Which integrations, delivery methods, or service boundaries are documented?

Proof and alternatives

17. What primary evidence supports [Company]’s material claims?
18. Which public sources describe [Company], and how current are they?
19. What are credible alternatives to [Company] for [specific buyer need]?
20. Under what conditions would [Company] be recommended or excluded?

Run prompts in a fresh, recorded context. Record product, visible mode, account state, date, market, and run number. Repeat high-risk questions and preserve all responses and citations.

The website understanding test traces access and source evidence, while the broader AI brand perception audit adds framing, sentiment, recommendations, and business value.

Distinguish recognition from understanding

A direct prompt supplies the brand name, so a plausible answer may show only recognition under a cue. Accurate entity understanding requires correct relationships, category, audience, offers, limits, and alternatives.

Score each material claim against the dated fact sheet:

• Correct: supported by a current authoritative source within the same scope.
• Incorrect: contradicted by current authoritative evidence.
• Stale: formerly correct but no longer current.
• Ambiguous: missing scope or mixing entities in a misleading way.
• Omitted: asked for and materially absent.
• Invented or unsupported: specific claim with no adequate supporting evidence found.
• Unverifiable: neither confirmed nor contradicted with available evidence.

An uncited statement may still have supporting evidence elsewhere. Search reasonable primary sources before assigning “invented or unsupported,” and retain the “unverifiable” label when the evidence is inconclusive. For cited claims, check whether the linked page supports the specific statement.

Score by entity dimension

Create a matrix with rows for the 20 prompts and columns for identity, relationship, category, audience, offer, limitation, proof, and alternative. Add exact passages, source URLs, severity, and reviewer confidence.

The matrix can include a few adjacent observations without becoming a glossary. Record whether the answer displays a source for the claim, how it frames the relevant attribute, and whether it recommends the company for the stated persona. Keep qualified referrals or conversions in analytics, where they can be tied to observed activity. Teams that need every layer from access through business value can use the canonical AI brand perception audit.

Use a materiality rule. Wrong legal identity, safety limits, pricing eligibility, or product availability may deserve urgent correction. A missing slogan usually does not. Have a second reviewer check high-severity claims.

Compare products and sources

Run the same panel in each named product you care about, but report experiences separately. Browsing and non-browsing modes are not equivalent. Interface labels and source displays can change, so retain complete captures.

Open visible citations and map them to claims. Compare owned sources, documentation, independent profiles, directories, and outdated pages. A repeated third-party source may be important, but it does not reveal internal model weighting.

OpenAI states that outputs can be inaccurate in its ChatGPT accuracy guidance, and it documents separate crawler controls in its official bot documentation. Those sources support caution and access checks. This test therefore reports what the product says under recorded conditions rather than claiming access to internal knowledge.

Protect confidential information

Limit this test to facts approved for public disclosure. Trade secrets, personal data, credentials, private customer information, unreleased financials, and privileged material belong in systems and workflows approved by the relevant security and privacy owners.

OpenAI’s Data Controls FAQ explains controls for chat history, model improvement, data export, and account deletion in ChatGPT. Product plans and policies can differ, so review the current terms and your organization’s rules before using business data. If the test requires private systems, involve security, privacy, and legal owners.

Build a remediation backlog

For every material error or omission, record the canonical fact, conflicting sources, technical access, proposed owner, action, and retest date. Repair owned contradictions first. Request external corrections through normal editorial channels. Use platform feedback where available.

The dedicated protocol for wrong company information in ChatGPT covers incident severity and retesting. A correction can improve source clarity without changing a generated answer immediately. Report before-and-after chronology without claiming causation.

If the test reveals broad entity, access, evidence, and measurement gaps, Leaf’s SEO and AEO audit can scope the controlled work and retest plan.