When an AI describes you wrong

Find out what it says, before anything else

Twenty minutes, no tools, and most companies have never done it.

Open the four assistants people actually use and ask each one about your company by name. Then ask three more things: what do they sell, who are their competitors, and would you recommend them.

Ask the same questions in a fresh conversation, because these systems are influenced by what you said earlier and the second answer is often cleaner than the first.

Write down what is wrong and what kind of wrong it is, because the fix is different for each. Out of date is the easiest. Confused with somebody else is the most common. Invented is the least common and the most alarming, and there is less you can do about it directly.

Where the wrong answer comes from

Almost never from your website, which is the part that catches people out.

For questions about a company, these systems lean heavily on third parties. Encyclopaedic sources, forums, directories, press, review sites, and older versions of the web that are baked into what the model was trained on.

That last one explains most of the out-of-date answers. A model trained on a snapshot of the web knows the version of you that existed when the snapshot was taken. Your new site does not update the snapshot.

It also explains why the four assistants disagree with each other. They read different things: one leans on encyclopaedic and reference sources, another on forums and recent pages, a third takes almost everything it cites from ordinary search results. Studies of citation overlap between them find only around one domain in ten appears in both of two major systems. A correction that works on one may not touch another.

What actually corrects it

In order of effect, and none of it is a button you press.

Fix the third-party records first. Directory listings, professional profiles, industry associations, anywhere your company is described by somebody else. These are the sources being read. Your own site is not the authority here; the corroboration is.

Make the facts machine-legible on your own site. Structured data stating who you are, where you are, what you do and what you offer. It does not force a correction, but it gives a system reading your site something unambiguous rather than something to infer.

Point at the proof. The field linking your site to your external profiles is what turns a page making claims about itself into an entity corroborated elsewhere. If those profiles do not exist yet, that is the prior job.

Publish the correct version somewhere new and dated. For the systems that retrieve live, recency counts, and one of them cites content under thirty days old several times more often than older material.

Use the feedback mechanisms. Most assistants have a way to report a bad answer. It is slow, opaque, and worth doing anyway on anything defamatory or seriously wrong.

The confusion case, which is the common one

Being mixed up with another company is the most frequent complaint, and it has a specific cause.

If your name is close to somebody else's, or you share a name with something in a different industry, a system with thin information about you will fill the gap with what it has about them. It is not malicious and it is not exactly a bug: it is what happens when there is not enough distinguishing information to separate two entities.

The fix is therefore not to argue about the mix-up. It is to become more distinguishable: consistent company name, location and sector everywhere you appear, external profiles that carry those consistently, and content that ties your name to your actual field often enough that the association is unambiguous.

That takes months rather than days. It is also the same work that makes you findable, so it is not a detour.

How long a correction takes

This is where expectations need setting honestly.

For assistants that search the web while answering, a corrected third-party page can start showing up within days of being indexed.

For answers drawn from training data, it depends entirely on the next training cycle, which has no published schedule and no way for you to influence it. A company that changed its name a year ago may still be described by the old one.

So the realistic goal is not "get it corrected this month". It is to make the correct version the easiest thing to find, everywhere, and let the systems converge on it. There is no support line for this.

What not to spend money on

A market has grown here faster than the evidence has, and some of it is priced as certainty.

Tools promising to make an assistant say what you want are, on inspection, monitoring tools. Monitoring is genuinely useful. It is a report, not a lever, and it should be priced as one.

Anyone guaranteeing a specific answer in a specific assistant is selling something nobody can deliver. These systems change without notice and disagree with each other by design.

And the file some people are selling as the way to instruct assistants about your site has been publicly stated by at least one major search provider to have no effect on their systems, with measurements showing the overwhelming majority of sites carrying it receive no requests for it at all. It costs one file to publish, so publish it. Do not buy it as a service.

When the wrong answer does not matter

Worth keeping proportionate.

If the error is trivial, an old team size or a slightly wrong founding year, it is not costing you work. Fix it when you next touch the profiles.

What matters is anything that changes a buying decision: a service you no longer offer, a market you no longer serve, a competitor named as you, or a claim about quality that is not true. Those are worth the months.

And if an assistant says nothing at all about you, that is a different problem with a different fix, and it is the more common of the two.