What to do when ChatGPT gets your company wrong

Someone asks ChatGPT about your company. The answer comes back fluent, confident — and wrong. A discontinued product described as current. A regulatory matter mischaracterised. A founder’s history conflated with someone else’s. Worse, the same error turns up in Gemini and in Google’s AI Overviews.

This is now a routine reputation problem, and it has a distinctive feature: there is no “report a correction” button that reaches the model. Fixing it means changing the inputs the engines learn from, not arguing with the output. Here is the sequence we use.

1. Document it precisely

Before doing anything, capture the problem properly. Vague reports (“ChatGPT says bad things about us”) cannot be acted on.

• Record the exact prompts that produce the error, word for word.
• Screenshot the responses, with dates, across each engine where it appears.
• Note whether the answer cites sources — and which ones.

The aim is a clear, repeatable record of what is wrong, where, and how often. That record is also how you will later prove the fix worked, and, just as importantly, how you avoid over-reacting to a one-off oddity. Answer engines vary their phrasing run to run, so a single strange response is not necessarily a pattern. Documenting the same prompts across engines and over a few days separates a genuine, repeatable error from noise, and tells you whether the problem is widespread or confined to one model. That distinction shapes everything that follows: a narrow issue on one engine warrants a lighter touch than the same falsehood echoing across ChatGPT, Gemini and Google’s AI Overviews at once.

2. Find where the error comes from

Inaccurate AI answers almost always trace back to a real source the model has ingested. Common origins:

• Stale or thin pages on your own site that still describe an old structure, product or position.
• Outdated third-party coverage (directories, profiles, old news) that ranks well and reads as authoritative.
• Reference sources and knowledge graphs whose entry for you is incomplete or wrong.
• Conflation with a similarly named company, person or product.

Tracing the chain from the wrong answer back to its likely sources is the part that most determines whether a correction will hold. Treat it as information-environment analysis, not guesswork.

3. Correct the record at the source

You cannot edit the model. You can improve what it reads.

• Fix your own house first. Update, consolidate or retire the pages that carry the outdated facts, and state the correct position plainly and prominently. This is the fastest lever you fully control.
• Strengthen the accurate signals. Make sure the correct facts appear, consistently, across the places an engine trusts: your site, structured data, and credible third-party sources.
• Pursue reference-source accuracy. Where a knowledge-graph or reference entry is the origin of the error, correct it through the proper channels, with evidence.
• Disambiguate. If you are being confused with another entity, add the clarifying detail — full legal name, sector, location — that lets a model tell you apart.

Throughout, keep to verifiable facts. The goal is narrative integrity: an accurate, well-evidenced version of you that is easier for a model to adopt than the wrong one.

A note on what not to do

Two instincts make the problem worse. The first is to demand the platform “take it down”. There is rarely a mechanism for that, and chasing one wastes the time better spent correcting the sources the model actually reads. The second is to flood the web with thin, repetitive content asserting the correct version. Answer engines are increasingly good at discounting low-quality, self-serving material, and a wave of it can read as manipulation rather than correction. The durable fix is fewer, better, genuinely authoritative sources — not volume.

4. Re-measure, then keep watching

Models update on their own schedules, so corrections do not register instantly. Re-run your documented prompts periodically and compare against the baseline you captured in step one.

Because answers drift and new sources appear, this is not a problem you fix once. The organisations that stay accurate in AI answers are the ones that monitor their machine-generated reputation as a standing capability, the same way they would media coverage.

Why this problem is growing

Two things are making AI accuracy a standing concern rather than an occasional nuisance. First, more people now begin their research inside an answer engine rather than a results page, so a wrong answer reaches decision-makers — customers, candidates, counterparties, journalists — before they ever visit your site to check. Second, the engines increasingly cite and quote each other’s sources, so an error that takes hold in one influential place can propagate across several assistants at once.

The practical implication is that accuracy in AI answers is no longer something to address only when a problem surfaces. The organisations least exposed are the ones that already treat their machine-generated reputation as a monitored asset, checking what the major engines say on a defined cadence, and keeping the accurate, well-structured version of themselves discoverable so the engines have something better to reach for. That is the preventative side of generative engine optimisation.

How we run this in practice

When a client brings us an inaccurate AI answer, we work the same sequence above as a managed service rather than a one-off clean-up. It sits inside our reputation management work, because the levers are the same: the sources an engine reads, and the structured signals that tell it who you are.

In practice that means a few concrete things. We baseline the problem across the major engines, capturing the exact prompts, responses and any cited sources, then trace each error back to its origin, whether that is a stale page on your own site, an outdated third-party profile, or a thin knowledge-graph entry.

From there we correct at the source: rewriting and consolidating your own pages, tightening structured data so the accurate facts are machine-readable, and pursuing reference-source and knowledge-graph corrections through the proper channels, with evidence. Then we keep the documented prompts on a standing watch, so drift and new sources are caught before they spread rather than after.

When to bring in help

Many cases are resolvable in-house with patience and the sequence above. It is worth getting specialist support when the error is spreading across multiple engines, when it touches a regulated or contested matter, or when the source is outside your control and needs careful, evidence-led correction.

If ChatGPT — or any answer engine — is stating something untrue about your business and it matters, the worst response is to ignore it and hope the next model update sorts it out. The record is correctable. It just has to be corrected at the source.

This is closely tied to the broader discipline of Generative Engine Optimisation, which is about being visible and accurate inside AI answers in the first place.