Why does ChatGPT say something wrong about my company?

Because AI models answer from what the open web says about you, weighted toward sources they treat as established: encyclopaedias, news archives, review platforms, directories and forums. Where those carry an outdated claim, a dispute that was later resolved, or simply nothing at all, the model fills the gap with the most confident-sounding source it can find.

Where does the claim actually come from?

One of two places, and the difference decides how fast it can be fixed.

Some of it is baked into the model from training. That material is fixed until the model is retrained, and no request changes it.

The rest is retrieved live. When an assistant searches before answering, it is reading today's web and summarising it. That part moves when the web moves, which is the part worth working on.

A wrong answer is usually both: a stale impression from training, confirmed by a stale page the model found when it looked.

Why does it sound so certain when it is wrong?

Because fluency and accuracy are separate properties. A model produces the most plausible continuation of a sentence, and a confident sentence is more plausible than a hedged one.

It is also why the gaps are dangerous. Where the record says nothing about you, the model does not say nothing. It reaches for the nearest confident-sounding source, which is often a directory entry, an old forum thread, or another company with a similar name.

Silence about you is not neutrality. It is an invitation.

Can I just tell the model it is wrong?

No. Correcting an assistant inside a conversation changes that conversation and nothing else. Nobody else sees it, and the next person asking gets the original answer.

There is no edit button, no verified-business appeal, and no registry of corrections. Any service offering to "update your record with ChatGPT" is describing something that does not exist.

What actually changes the answer?

Changing what the sources say. That is the whole mechanism, and it is slower and more durable than it sounds.

Three moves, usually at once. Remove material that breaches a policy or a law, so it stops being available to retrieve. Get factual errors corrected at the publisher, because a corrected article is read as current. And publish substantive, attributable material that answers the same question better, because a model choosing between sources picks the one that looks most like an answer.

Consistency matters more here than in search. A company whose own details differ between its site, its LinkedIn and three directories is giving the model a reason to trust none of them.

Freshness also decays. Vendor analyses put pages left unrevised for a quarter at roughly three times more likely to lose an AI citation, which matches how the live-retrieval engines behave — so a correction is something to maintain rather than to land once.

How long does a correction take to show up?

It varies by engine, and they do not move together.

An assistant that retrieves live can reflect a change within days of the source changing. An assistant reading a search index waits for that index to recrawl, which is weeks. Anything baked into training waits for a retrain, which is not on a schedule you can plan against.

So a correction lands in stages, and checking one engine tells you The spread between engines is wide enough to plan around. One published analysis found citation volumes for the same brand differing by a factor of 615 between AI systems.

almost nothing about the others.

Where does each engine get its answer?

This is also why the search work and the AI work are not separate budgets. Most of what an answer engine reads is what a search engine indexed.

How do I check whether it worked?

Ask them, in the words a buyer would use, and keep the answers.

Query each engine with the questions a client, investor or journalist would actually ask about your category and your name. Record not only what each says but which sources it cites — the citations are the actionable part, because they name the pages doing the damage.

Repeat on a schedule. The answers move as the web moves, and a correction that landed once can be undone by a single new article.