When a Model Gets Your Brand Wrong: A Correction Playbook

The call usually arrives in the same shape. Somebody senior has asked an assistant about the company, been told something that is not true, and wants it taken down by the end of the day.

There is no taking it down. What there is instead is a sequence that reliably works, and a first step that is almost always skipped.

Reproduce it before you react

Half the escalations that arrive as emergencies are not reproducible, and knowing that in the first ten minutes changes everything about the response.

Open a temporary chat or a logged-out session. Run the exact prompt the person used, word for word. Then run it several more times.

Three outcomes, and they call for different responses.

It appears every time. A persistent error, probably sitting in the model’s memory rather than retrieved from a page. Serious, and slow to shift.

It appears sometimes. The claim is being pulled from a specific source that the model reaches for on some runs and not others. This is the most tractable version, because the source is findable and the clock is weeks rather than years.

It does not appear at all. Very common. Personalisation in the original account, a one-off generation, or a prompt that steered the model towards the answer it gave. Worth checking exactly how the question was phrased before anybody spends a quarter on it.

Whatever happens, capture it now: exact prompt, engine and model, date and time, the full answer as text, and every source named. Answers are non-deterministic and they drift. Evidence you did not capture is frequently not reproducible a week later, and that matters enormously if this ever becomes a legal matter.

Triage by what it would actually cost

Not every error deserves a project. Sort by consequence, not by irritation.

Critical. Claims of illegality, safety failures, fraud, regulatory action, insolvency, or anything about a named individual. Also anything materially false about a listed company. Escalate immediately, preserve evidence, involve legal.

Commercial. Wrong pricing model, capabilities you do not have, capabilities you do have described as missing, a rival named as the better choice on a false premise. This is where most real money is lost, and it is the tier that gets least attention because nothing about it feels like a crisis.

Factual. Stale leadership, old funding, retired products, wrong headquarters, incorrect ownership. Genuinely worth fixing, rarely urgent, and usually the easiest work on this list.

Cosmetic. Tone, emphasis, a description you find unflattering but which is defensible. Log it and move on. A programme that treats every unflattering adjective as an incident will be exhausted before it reaches anything that matters.

The most expensive errors are almost never the ones that feel most offensive. A wrong pricing model quietly loses deals for a year. A slightly rude adjective loses nothing at all.

Classify the error, because the fix differs

Stale facts. Something was true and is not any more. The correction is straightforward: publish the current fact clearly and get it repeated somewhere independent. The old source usually does not need removing, only outweighing by newer and clearer material.

Blended identity. You have been mixed with a company that shares part of your name or occupies an adjacent category. The fix is disambiguation rather than correction. Be described consistently everywhere, use your full legal name alongside your trading name in reference material, and make sure your category and location are stated explicitly in the places models read. This error is badly under-diagnosed because the text reads as plausible until somebody who knows the company looks closely.

Fabrication. No source, no basis, the model produced a confident sentence out of pattern rather than fact. There is nothing to trace, which is frustrating, and the only durable answer is to make the correct fact abundant and easy to find so the model has something better to reach for. These are also the errors most worth reporting through the provider’s own channels.

Inherited framing. Every individual claim is sourced and true, and the overall impression is wrong, usually because one critical period is over-represented in what exists about you. Nothing here can be corrected, because nothing is false. This is a publishing problem: the record needs more recent, substantive material to summarise, and that takes quarters rather than weeks.

Trace the source

For everything except fabrication, the claim came from somewhere.

If the engine cited, you already have the list. If it did not, ask it directly which sources support the claim, and then search the claim in the model’s own phrasing, because the wording usually echoes whatever it read closely enough to find the original.

Then read those pages properly. You are looking for the origin rather than the copies: one wrong sentence in a widely syndicated article can be the root of everything downstream, and correcting the original is worth more than chasing twenty reprints.

Fix at the source

Publish the accurate version where models retrieve from, starting with your own site. One question per page, the correct fact stated plainly in the opening lines, dated, and specific enough to be repeated without hedging. A correction buried in paragraph nine of a general About page will not do the job.

Get it corroborated off your own domain. A correction that exists only on your site is a claim with an obvious motive attached. The same fact in trade press, a reference source or an analyst note is weighted very differently. This is ordinary media relations, which is genuinely the main tool here.

Approach the original publisher where the source is a specific article. A factual correction request to a journal that got something wrong is a normal, unremarkable request, and it is far more effective than anything you can do on your own domain.

The provider channels, honestly assessed

Every major assistant has a feedback mechanism, and most have a formal route for reporting inaccurate information about a person or organisation. Some jurisdictions also give individuals data-correction rights that can apply to model outputs, though how those work in practice is still being established.

Use them for the critical tier. Document that you used them and when.

But set expectations internally: these channels are not a support queue with a turnaround time, and they rarely produce a targeted fix on a schedule you can plan around. Filing a report is a sensible and sometimes legally useful step. It is not the plan.

When to involve legal

Sooner than feels proportionate, if the claim alleges illegality, harm, regulatory trouble, or anything about a named individual.

Whether any of this amounts to defamation is a genuinely open question that varies by jurisdiction, by the specific claim and by who saw it. That makes it a question for your lawyers, not for your comms team, and not for an article. The useful operational test is simpler: if you would have sued a publisher for printing it, treat it with the same seriousness here, starting with evidence preservation while the output is still reproducible.

Verify, and keep a register

The step everyone skips, and the reason most programmes cannot tell whether any of this worked.

Re-run the exact prompt weekly, in a clean session, several times per run. Record what came back. Retrieval-based errors can improve within weeks of a corrected source going live. Memory-based errors will not move until the model does.

Keep a register with a row per issue: the claim, the date found, the severity, the error type, the sources traced, what you published, the date published, and the state of the answer at each re-test.

That register is what turns a scramble into a programme. It is the only thing that lets you say which corrections worked, it is what a board or a regulator will ask for, and it is the difference between fixing an error and merely reacting to one.

The wider discipline this sits inside is covered in AI reputation management, and the testing method it depends on is in how to check what ChatGPT says about your brand. Running the re-tests automatically, across every engine, is what AI reputation tracking does.

Common questions

How do I correct false information an AI gives about my company?

You cannot edit the output, so the work happens on the sources. Reproduce the error in a clean session to confirm it is real, trace which pages support the claim, publish a clear dated correction on your own site, get it corroborated independently, then re-test the same prompt weekly and record the result.

Can I make ChatGPT or Gemini delete something false about my brand?

There is no delete button, and any vendor offering one is describing something that does not exist. Providers do offer feedback and reporting channels, and they are worth using for serious factual errors, but they are slow and rarely produce a targeted fix. Source work is what reliably moves an answer.

Is an AI hallucination about a company defamation?

That depends entirely on jurisdiction, on the specific claim and on who saw it, so it is a question for your lawyers rather than your comms team. The practical trigger for escalating is straightforward: if the claim alleges illegality, harm, or anything you would sue a publisher over, involve legal early and start preserving evidence immediately.

How do I prove an AI said something false about my brand?

Capture it properly at the moment you find it. Record the exact prompt, the engine and model, the date and time, the full answer text rather than a screenshot alone, and every source it cited. Answers are non-deterministic and change over time, so evidence that was not captured is often not reproducible later.

How long does it take to fix an AI error about a company?

Errors coming from live retrieval can improve within weeks of a corrected source being published and corroborated. Errors sitting in a model's training data move on the far slower clock of model releases. Running the same prompt several times tells you which one you are dealing with before you promise anyone a timeline.