How do I correct something an AI keeps getting wrong about me?

Can I just tell the AI it is wrong?

No, and the time spent trying is the most common way this goes wrong.

Correcting a model inside a chat window changes that conversation and nothing else. There is no support desk that edits what an engine believes, no submission form, and no relationship to invoke. Practitioner guidance on this is blunt: attempting to correct the engine directly does not work and is not worth the time.

What can be changed is the material the answer was assembled from. Every wrong statement about you is downstream of something: a stale directory entry, an old press mention, a competitor's comparison page, a page of your own you forgot to update. The error is a symptom. The source is the thing you can actually reach.

So the whole job is: find the source, correct it there, and give the engine something better to find next time.

How do I find where the error came from?

Ask the question that produces the error, then ask what it was based on.

Most assistants will cite or link when asked directly. Those citations are the shortlist. Where nothing is cited, search the wrong claim itself as a phrase, in quotes, and see which pages carry it; a fabricated fact usually has a real ancestor somewhere, and a widely repeated one has a single origin further back.

Test across more than one engine before concluding anything. Each has different training data and different retrieval, so an error on one is not evidence of an error everywhere, and the sources they draw on differ. Write down what each one says, verbatim, with the date. That record is what lets you tell a genuine change from a lucky reroll later.

Three patterns account for most of what people find:

• Something true that expired. Old pricing, a service you stopped offering, a location you left, a role someone no longer holds.
• Something inferred from a gap. Where sources conflict or say nothing, models fill in with what is statistically plausible, which is how a confident invention appears with no ancestor at all.
• Something about a different entity. Two organisations with similar names, merged in the summary.

What do I actually change?

Whatever you can reach, in order of how much control you have.

• Your own pages first. If the wrong fact is on a property you own, this is the whole fix and it is available today. Start with the pages that get crawled most: the home page, the about page, the main service pages.
• Structured data. Organisation markup stating your name, location, founding date, services and identifiers gives a machine an unambiguous reading of facts it would otherwise infer from layout.
• Listings and directories you control. Claim them and correct them. Guidance in 2026 singles out Bing Places as directly influencing what ChatGPT says, which is not obvious.
• Third-party pages you do not own. Ask. A dated, polite correction to a publisher with the accurate fact attached succeeds more often than people expect, particularly where the error is demonstrable.

The consistent finding underneath all of this is that weak or inconsistent signals across the web are what produce wrong answers in the first place. Saying the same true thing, in the same words, everywhere you appear, is doing more work than any single correction.

Why publish a correction of my own?

Because removing a wrong source is often impossible, and out-weighing it is not.

Write the page that answers the question directly. If an engine says you offer something you do not, publish a page that states plainly what you do and do not provide. Give it a question-shaped heading, put the correct fact in the first screen in a sentence that survives being quoted alone, and date it.

This is the part most people skip, because it feels indirect. It is the part that lasts. A correction sent to a third party fixes one page; an authoritative page of your own is a source that keeps being available every time the question is asked.

The same logic applies to a presence that answers on your behalf. When a persona here answers, it draws on approved material and names what it used, so a wrong sentence can be traced to the document that produced it and corrected at that document rather than argued with.

How will I know it worked?

By asking the same questions again, on a fixed schedule, and writing down what comes back.

Keep the prompt set small and stable, ten to twenty questions a customer would actually ask, and run it monthly across the engines you care about. Record the answers verbatim with the date. Changes are gradual and uneven: one engine may update while another does not, and an answer can improve and then regress.

Be careful about what counts as evidence. Asking twice in one session and getting different answers is normal variance, not progress. A change you can see across a month and across sessions is the real thing.

Set the expectation early with anyone waiting on this. Your own surfaces move quickly because you control them. What third parties say, and what engines have already absorbed, moves slowly. Anything promising a decisive change within a fortnight is describing a dashboard rather than reality.

What if I cannot change the source?

Then the honest remedy is slower than anyone wants, and it is still the one that works.

Publish the correct fact clearly on a property you own, make it easy to retrieve, keep it current, and let corroboration accumulate. Where an error is defamatory or legally actionable that is a different route with different advice, and worth real legal input rather than a marketing process.

What is worth avoiding is the vendor promising to make it disappear. There is no delete button on a model, nobody can guarantee what an engine will retrieve tomorrow, and an offer built on that is selling certainty that does not exist.

One consolation from the shape of the problem: the same work that corrects an error is the work that makes you retrievable in general. A clear, current, well-structured account of what you do is not a remedy you throw away afterwards.