What to do when ChatGPT gets a fact about you wrong

Find out whether ChatGPT fetched the wrong fact from a page or is recalling it from training, then fix the source if there is one and make the correct fact easy to find if there is not. A fetched error can clear within weeks. A remembered one takes much longer and nobody can promise a date.

What ChatGPT does differently

Because it sometimes searches and sometimes does not, the same wrong fact can have two different causes in the same week. With citations, you can see the page it read. Without citations, the fact came from the model and there is no page to correct.

Asking it where the fact came from is not reliable. It may produce a plausible sounding source that was not actually used. Trust the citation list, which is evidence, rather than the explanation, which is generated.

There is a feedback control on answers, and using it costs nothing. Do not build a plan around it. It is not a correction mechanism you can depend on or track.

What that means for whether you are named

A wrong fact is worse than absence in one specific way. If an assistant believes your pricing is higher than it is, or that you do not serve a segment you serve, it will actively route buyers away from you while appearing to be helpful. So a wrong fact usually deserves attention ahead of a general visibility push.

Most wrong facts come from four places: an old page of your own, a stale third party listing, confusion with a similarly named company, or a gap the model filled because almost nothing was written about you. The last one is fixed by supply rather than correction.

The six wrong facts worth checking for by name

Do not wait to be told. These are the ones that recur, and each one has a specific place to look.

One, an old price. Look at your own old pricing page, any old blog post that quoted the price, and every directory profile.

Two, a product you no longer sell. Old feature pages, old comparison articles, app store listings.

Three, the wrong customer segment. Usually a third party description written from your largest plan, or a directory field left at its default.

Four, the wrong location or coverage. A closed office, a city you have added since, or a coverage count that has grown.

Five, a previous company or product name. Anything published before a rename, including your own old pages and any press coverage.

Six, facts belonging to a similarly named company. Check by asking the question and looking for details that are true about somebody else.

A worked example

On 6 August 2026 ChatGPT reported a contradiction about clawlaw.in, our sister company, in the middle of an answer. Its website and its app store listing carried different plan names and different prices for the same product, and ChatGPT said so.

Both of those pages had been published by the company. That is the useful part. The wrong fact had no third party to blame and no training run to wait for. It was a fetched fact, sitting in public, and correcting it was entirely inside the company's control.

Against the six wrong facts listed above this is number one, an old price, in its commonest form: a price corrected in one place and left standing in another. The app store listing is the place that gets forgotten, because nobody on the marketing side edits it.

The fix here is the short list on this page. Find every place the price is stated, including the listings you do not think of as your site, correct each one in place with the date of the change, and keep one page that says plainly what the price is now. Then re-ask in both modes, and watch the searched answers first.

What to change

• Search your own site for the outdated value and fix every instance, including old blog posts and PDFs.
• Publish one page that states the correct fact plainly, with the date it became true. Keep it updated rather than replacing it.
• Correct old pages in place instead of deleting them. A deleted page leaves the wrong version in circulation with nothing to contradict it.
• Email the third party pages carrying the old value. A short factual note with the correct figure works more often than people expect.
• Make every profile and listing use the same wording. Where sources disagree, the assistant gets to choose, and it may not choose the current one.

What you cannot make happen

Four limits, because the promises made in this area are usually the parts nobody controls.

You cannot edit an answer. There is no correction console and no setting that pins your version of a fact.

You cannot rely on the feedback control. Use it, and do not report it as an action with an expected outcome.

You cannot make the model forget. A remembered error can persist until a future training run whose date is not announced.

You cannot compel a third party. A factual, polite request often works. Some sites will not reply, and then your only move is to make the correct version easier to find than the wrong one.

For genuinely harmful claims or personal data there are formal channels and, in some places, legal routes. Those are slow and out of proportion for an out of date price, so keep them for real harm.

Common questions

How long until the correction shows?

The searched answers can move in weeks, because they depend on pages being re-crawled. The unsearched ones have no honest estimate. Report the two separately so the wait is expected rather than treated as failure.

Should I delete the page that had the wrong fact?

Usually no. Correct it in place and date the change. Deleting removes your corrected version while copies of the old value survive elsewhere.

Can I add a note to my site telling AI the right answer?

Write for readers and state the fact plainly with a date. There is no instruction you can place on a page that an assistant is obliged to follow, and text addressed to a machine reads badly to the humans who arrive.

It is confusing us with another company. What actually helps?

Distinguishing detail used consistently: full legal name, city, sector, and what you do, in the same words on your site and every profile. Where the other company is much better known, add the clarifying detail to your own pages rather than arguing with theirs.

Is it worth telling my sales team?

Yes, and ask them to report what prospects quote back at them. Sales often hears a wrong fact months before anybody measures it, and that is free monitoring you already have.

Does one correct answer mean it is fixed?

No. Answers vary between runs, so check across several readings in both modes before you call it done.

What to do first

Re-ask the same question monthly in both modes, search off and search on, and log the answer text. Watch the searched version first: that is where a correction lands soonest, because it depends on pages rather than on a future training run.

Today, do three things. Copy the wrong sentence into a log with the date and the mode. Open every citation and fix whatever you control. Send one short factual note to whichever third party page still carries the old value.

If the searched answer is now right and the unsearched one is still wrong, the fix worked and you are waiting on the slower clock. That is the normal shape of this, and it is worth telling whoever asked you to fix it.