When ChatGPT Gives Wrong Information About Your Business, Where Did It Come From?

A customer forwards you a screenshot. They asked about your product, and ChatGPT gives wrong information back: a price you retired last year, a feature you never shipped, an integration you dropped. There is no correction form, no support queue, and no obvious place to file a complaint.

The instinct is to find the sentence somewhere on the internet that caused it and delete that sentence. Sometimes that works. Often there is no single sentence, and chasing one costs you a week.

Reproduce it before you spend anything on it

Run the exact prompt yourself, several times, in a fresh session with no memory or custom instructions carrying over. Then run it on the other engines your buyers actually use.

The question is not really whether ChatGPT gives wrong information once. It is whether it does so reliably, because only one of those is a problem you can act on. Treat the pattern as diagnostic rather than causal: none of these outcomes identifies the source on its own.

• Several engines repeat it. The problem is not isolated to one conversation. That can come from overlapping training data, a widely repeated piece of misinformation, a common public source, or retrieval, and you cannot yet tell which.
• Only one engine repeats it. The behaviour may be specific to that engine, its search provider or its index state. That narrows where to look without telling you what to blame.
• You cannot reproduce it. The original answer may have depended on generation variability, retrieval state, conversation context, personalisation or a configuration you did not replicate. Log it, watch it, and do not reorganise your week around it.

That third case is more common than people expect, and it is the reason the screenshot alone is not enough to act on. A single screenshot captures one draw from a process that varies between runs.

Ask the engine for its sources, then treat the answer as a lead

Ask directly where the claim came from. When an engine retrieved live and shows citations, those links are real and checkable, and this step can end the investigation in five minutes.

When the answer carries no verifiable citation or retrieval trace, treat the model's account of where the claim came from as a hypothesis rather than provenance. A model can produce a plausible explanation that does not faithfully describe how the answer was actually generated. And even when citations are present, open each one and confirm it exists and genuinely contains what it was cited for.

This is where people lose the most time. A fabricated citation is very convincing and sends you looking for a page that was never there.

Search the way a retrieval system would

If there are no citations, reconstruct what a retrieval step would surface. Search the wrong claim itself, in quotes, alongside your brand name. Then work through the places that outrank your own site for questions about you.

• Your own stale pages. Old pricing pages, superseded documentation, a changelog entry, a landing page for a discontinued plan that still resolves.
• Directories and marketplaces. Listings created years ago by someone who has left, still carrying the old description.
• Press and funding databases. Company profiles built once and never revised.
• Review sites and comparison pages. Including competitor comparison pages, which describe you in whatever terms suit them and are frequently out of date.
• Structured knowledge bases. Wikidata and similar sources often carry facts about your business, products or leadership. They are collaboratively maintained and can go stale, though some items are edit-protected. Do not assume one caused the answer without evidence linking it.

In practice the culprit is usually mundane. A page you forgot you controlled, or a third-party listing nobody has looked at since it was created.

Can you make ChatGPT forget something?

Not in the sense you want. You cannot reliably erase a public business fact from a deployed model's learned knowledge by telling it to forget. Deleting your own conversations or saved memories is a separate mechanism and does nothing about what the model knows in general.

What you can change is what is retrievable and what is unambiguous. Correct the sources you own, get the third-party ones updated where someone will take your request, and make the accurate version easy to find and easy to parse. Then the next retrieval has a better chance of landing on the right thing.

If the claim is coming from learned model knowledge rather than retrieval, editing your page will not rewrite weights that are already trained. The correction has to arrive through retrieval or through a later model or provider update. Retrieval may override stale learned knowledge, but it is not guaranteed to.

Fix the pages you control first, because they are the only guaranteed win

Everything you own can be corrected today. Everything you do not own depends on somebody else answering an email.

Start with the page that should be authoritative on the disputed point and make the correct claim explicit and current. Vague copy leaves the gap open. If your pricing page only says "flexible plans" while a three-year-old directory listing names a specific dollar figure, that stale listing is a concrete conflicting fact sitting there available to be surfaced. Specificity alone does not decide which source gets chosen, but vagueness guarantees you are not the one offering the answer.

Retire or redirect genuinely dead pages rather than leaving them to resolve quietly. A discontinued plan page that still returns a 200 stays publicly accessible, so crawlers and retrieval systems can keep discovering it. A successful response does not prove any particular system has indexed or used it, but it does leave the contradiction reachable.

Make your identity declarable rather than inferable. If a machine has to work out what your company is and does from prose and navigation, it will guess, and its guess may be several years old. Organization schema is how you state it without leaving room for interpretation, and structuring content for AI search covers the rest of making a page legible.

How long until the correction shows up?

It depends on which layer the wrong claim lives in, and you usually cannot tell from the outside.

A retrieval-based answer can change once the relevant crawler, index or search provider refreshes your page, but the timing is not predictable and no provider offers a refresh guarantee for a specific URL. Indexed and cached layers add lag you cannot observe from outside. If the information is coming only from learned model knowledge, editing the page will not update it at all.

Because of that spread, do not check once and conclude you failed. Re-run the same prompt on a schedule for a few weeks. This is also the point where a single check stops being enough, because a generated answer varies between runs and one bad draw after your fix does not mean the fix did not work.

Watch it instead of rechecking it by hand

When ChatGPT gives wrong information about a company for months on end, the reason is almost always that nobody was looking. It surfaces when a customer happens to mention it, which means it has been shaping buying conversations for a while before you hear about it.

Decide which claims about you actually matter commercially, write them as prompts a buyer would genuinely type, and check them on a schedule across the engines your buyers use. You can see what the engines currently say about you with our free AI visibility check, and AI visibility tracking reruns your configured prompts across supported engines on a schedule, then alerts you when your visibility drops or you lose a citation you previously held.

One honest limit worth stating plainly: from outside a black-box assistant you generally cannot prove that one particular public source caused one particular sentence. Citations and retrieval traces show what the system consulted or attached to its answer, which is useful, but they are not a one-to-one causal explanation of every sentence. What monitoring gives you is the thing that actually matters operationally, which is knowing what is being said about you and when it changed, rather than finding out from a customer six months later.