Why does ChatGPT describe my company incorrectly?

Because the model is answering confidently from incomplete evidence. Language models do not look you up in a registry; they assemble a description from training data that may be years old, from whatever live sources they retrieve, and from patterns that fill the gaps. When the evidence about you is thin, the gap-filling takes over, and it does not announce itself as a guess.

We have watched this happen to ourselves. Asked about a product of ours, one assistant confidently described a different company’s product that happens to share the name: wrong platform, wrong purpose, delivered fluently. Nothing about the answer signalled uncertainty. That is the failure mode you are up against: not hostility, but confident conflation.

The usual causes

• Stale training data. You pivoted, renamed, repriced or moved, and the model learned the old version. It will repeat the old version until stronger, fresher evidence displaces it.
• Name collisions. If a company, product, app or even a common word shares your name, the model may merge you. Distinctive naming context on your site (“Acme, the Bristol structural engineering practice”) gives it the handle to keep you separate.
• A vague website. If your own pages never state plainly what you do, who for, and where, the model has no authoritative version of the facts to prefer. Abstract brand copy reads as atmosphere, and atmosphere cannot correct an error.
• Third-party noise. Old directory entries, an abandoned Crunchbase profile, a review of a product you killed years ago. Retrieval-based assistants can surface these as if current.

How to correct the record

You cannot email ChatGPT a correction, but you can change the evidence it works from. State the facts in plain text on your own site: what you are, who you serve, where you operate, what you cost. Add JSON-LD structured data so the facts exist in machine format, not just prose. Publish an llms.txt summarising the company for AI crawlers. Update or kill stale third-party listings. Where a name collision exists, consistently pair your name with disambiguating context everywhere you appear.

Then re-check. Retrieval-based assistants can pick up corrected facts quickly; descriptions rooted in training data change more slowly, on model release cycles. The pattern across several questions and providers tells you whether the record is healing.

The first step is knowing precisely what each assistant currently says, because you cannot correct an error you have not seen. We built Sourcive to capture that: it asks the real model APIs about your company, classifies whether their descriptions are accurate, outdated or wrong, and generates the plain-text facts, structured data and llms.txt that put the correct version on the record.