What AI gets wrong about businesses
We asked the main AI engines the same factual questions about the companies, people and shows we track, and checked every answer against a verified record. This page reports what they got wrong, how often, and which facts go wrong most.
Measured 25 May 2026 – 13 October 2026
The short answer
We checked 67 companies, people and shows across ChatGPT, Claude, Gemini, Perplexity and Grok. 57 of them (85%) had at least one engine state at least one wrong fact. For companies alone, it was 7 of 16, and these are large, well-documented companies.
Which facts about a business AI gets wrong
Share of answers that contradicted the verified record, companies only, most-missed first. Facts with fewer than 15 checked answers are left out.
| Fact | Wrong | Answers checked |
|---|---|---|
| Head office town or city | 20% | 45 (16 companies) |
| Who runs it | 14.7% | 34 (11 companies) |
| Year founded | 11.4% | 35 (16 companies) |
| Registered legal name | 11.1% | 45 (15 companies) |
| Official website | 2.9% | 35 (11 companies) |
| What kind of business it is | 2% | 49 (16 companies) |
| Head office country | 0% | 35 (16 companies) |
The pattern is that facts which change get missed. A company that moves its head office or changes its leadership can keep its old details in AI answers. A country rarely changes, and the engines almost never got it wrong.
The less written about you, the more AI gets wrong
Agreement with the verified record, all engines together: 92% for companies (327 answers), 61% for people (608 answers) and 61% for podcasts and creators (140 answers). The companies are global names with registry filings and plenty of coverage. People and shows have a thinner public record, and the engines get more of it wrong.
A small business usually has less written about it than any of these. That is why we check your own business rather than assume these averages apply to you.
Asking a second AI engine does not settle it
Of 316 facts that two or more engines answered, they split on 22%: one engine matched the record and another, asked the same question, contradicted it. A customer who checks a second engine can get a different answer, and has no way to tell which one is right. More on engine agreement.
Being known is not the same as being recommended
We also ran multi-turn buying conversations with each engine, the way a real buyer would ask. Across 21reference entities, mostly companies, the brand was still being recommended at the end of an open “what would you recommend?” conversation 15% of the time, and was still the pick after the buyer narrowed down to a final choice 6% of the time. Measured 12 July 2026 – 11 October 2026.
When a buyer pushed back on a brand the engine had already named, it usually held its ground (72% of the time). The hard part is being recommended at all, and then surviving the narrowing questions.
How we measured it
- Each engine got the same fixed set of factual questions per entity and answered in a structured format.
- Every answer was compared with a verified record from registries, official sources and Wikidata.
- Wrong means the answer contradicted the record. A refusal or “I don’t know” counts as unknown, not wrong.
- We use the latest answer from each engine for each entity, so old answers that were later fixed do not count against it.
- Sometimes the record is the one that is out of date. When an engine turns out to be right, we correct the record.