ChatGPT Is Saying Something Wrong About Your Company. Now What?

A client tells you that ChatGPT quoted them a price you have never charged. Or listed a service you do not offer. Or described a qualification one of your people does not hold.

This is more common than most businesses realise, and it is a genuinely awkward category of problem. A bad review can be answered. An inaccurate search result can be reported. An assistant confidently stating something false about your business has no reply button and no obvious owner.

It is also not, in most cases, unfixable. But the fix works differently from what people expect.

**Why it happens**

Two mechanisms, and they need different responses.

The first is training data. The model learned something during training — from an old page of yours, a directory entry, a news article, an aggregator that guessed — and repeats it. You cannot edit training data. What you can do is change what the model finds when it retrieves current information, which increasingly outweighs stale training for factual questions.

The second is retrieval. The model is fetching live sources and summarising them. If those sources carry a wrong figure, the model faithfully reports the wrong figure. This is the more tractable case and, in our experience, the more common one for specific claims like prices.

There is a third possibility worth naming: the model may have inferred rather than read. If your site never states a price and the model has seen typical prices for your category, it may produce a plausible number that is simply invented. Silence is not neutral here — an absence of stated facts is an invitation to fill the gap.

**What actually works**

*Publish the correct fact prominently and unambiguously on your own site.* This sounds too simple to be the answer and it is frequently the largest part of it. If your pricing page states a range clearly, in text, with what is included, that is a strong, retrievable, authoritative source. If your pricing exists only in a PDF or behind an enquiry form, you have left the field open.

*Find and correct the upstream sources.* Work out where the wrong claim actually lives. Directory listings, aggregators, old press coverage, a review site with stale data, a partner's website describing you incorrectly. Correcting at source is slower than editing your own site and considerably more durable.

*Add structured data stating the fact.* Schema markup gives the correct value in a form that is unambiguous and machine-readable. Where prose can be misread, structured data is explicit.

*Re-test on a schedule.* Corrections propagate unevenly. Retrieval-based answers can update within days once the underlying source changes. Training-based answers may persist until a model is retrained. Check weekly rather than assuming the fix landed.

**What does not work**

Contacting the AI provider to request a correction is not, at present, a reliable route for an ordinary business. Some platforms have feedback mechanisms; none offer anything resembling a dependable correction process at the speed a business needs.

Publishing a rebuttal page — "contrary to what you may have read, we do not charge X" — tends to backfire. It creates a page that associates your business with the wrong figure, which is the opposite of what you want retrieved.

Ignoring it does not work either. Unlike a bad review, which ages out of relevance, an incorrect fact in a model's answer is repeated identically to every person who asks, indefinitely, until the sources change.

**Why this matters more in some Dubai sectors than others**

For a restaurant, a wrong price is an irritation. For a clinic quoting procedure costs to medical tourists, for a developer whose payment schedule is being misstated to overseas investors, or for a regulated firm whose licensing is being described inaccurately, it is materially damaging — and in the regulated cases potentially a compliance issue in its own right.

Those are also the sectors where prospects are most likely to be researching remotely, at length, through exactly these tools, before any human contact. The wrong answer does its damage before you know there is a conversation happening.

**Treat it as monitoring, not as an incident**

The businesses that handle this well do not treat it as something to fix once. They check what assistants say about them on a regular cycle — prices, services, credentials, regulatory standing — the same way they would monitor reviews.

The difference is that nobody sends you a notification. If you are not checking deliberately, the first you will hear of it is a client mentioning it in passing, and by then it will have been told to everyone who asked.

Prêt à développer votre entreprise à Dubai ?

Obtenez un audit gratuit et découvrez comment nous pouvons vous aider à dominer votre marché local.