What is the difference between outdated, mixed up and invented?
The fix. In an answer they look identical — a sentence about you that is not true — and they call for three different actions in three different places. Anyone lumping them together as "hallucination" is left with one word where three assignments could have stood.
| kind of error | what is going on | where the fix sits |
|---|---|---|
| outdated | A real source says it, and it was true once. An old address, a previous owner, a service you no longer offer. | Correct that source. The address is in the finding, so it is a letter or a form and not a search. |
| mixed up | True, but about someone else. A similar name, a namesake in another city, a parent company that swallows the subsidiary. | Untangle the entity: your own identifiers, your own record, and on every shared source say who is who. |
| invented | No source says it; the model filled the gap. An opening hour that is written nowhere, a certificate that does not exist. | There is no source to correct; the fix is making the right fact machine-readable enough that there is nothing left to guess. |
How do you know whether AI invented it or read it somewhere?
From what the engine did just before answering. If it fetched sources, something is written down wrong somewhere and the address is in your finding. If it fetched none, it answered from memory and there is nothing to correct. That is an indication and not a verdict — we do not read those sources, we count whether they were there — and it says so.
| what the engine did | the indication | what you do with it |
|---|---|---|
| sources fetched | The engine fetched sources; the error probably comes from one of them. | The list comes with it. You start at the source that weighs heaviest, not at the one that annoys you most. |
| no sources fetched | The engine fetched no sources. This looks invented rather than picked up. | There is nothing to correct, because nothing was written down wrong. What helps is putting the right fact somewhere an engine will fetch next time. |
Can you have a wrong ChatGPT answer removed?
No, and nobody can — we cannot either. Those models belong to OpenAI, Google, Anthropic and Perplexity, and there is no button with which a company lifts a sentence out of their answers. Anyone promising you that is selling something they have no access to.
Who is liable for what AI says about our company?
That is a legal question and we are not lawyers, so what follows is what we know rather than what we hope. In practice the damage lands with you and the text comes from someone else: the model provider wrote the sentence, you lose the customer. What we can contribute is the evidence a lawyer can work from — the question as it was asked, the full answer, the engine, the date and the sources underneath.
What does it cost to get a wrong AI claim put right?
It depends on which of the three it is, and that is the most useful sentence on this page. Correcting an outdated source is often a form at a directory and half an hour of your time. Untangling a mix-up takes longest, because it sits in the sources of both companies and not in one place. An invented fact is in between: nothing to correct, only something to add, and that means one page that is right and machine-readable.
Frequently asked questions
How often does AI say something wrong about a company?
We only know that for your company after a measurement, and we do not add a general percentage. Figures on this circulate widely and are almost never comparable: they differ per model, per market, per kind of question and per month. What you get from us is your own number, with the questions and answers it comes from.
Do you publish counter-content to displace a wrong story?
Yes, and always under your name. That is the hard condition under everything we place outside your own site: it says who is writing. A correction posing as an independent voice is astroturfing — banned in the EU, and ruled out with us even when someone asks for it.
What if the error comes from a review I am not allowed to have removed?
Then the source stays and the weight changes. An engine rarely reads one source; it weighs what it finds. An incorrect claim appearing once against a correct fact standing in four independent places loses — the same movement as with an invented fact, only with a dissenting voice in it.
Do you also measure sentiment, or only errors?
Both, and the sharpest signal is neither. Per answer we record three values plus the caveat the engine adds itself — "good but expensive", "strong but far". That is not a negative answer and it costs you the customer anyway; a sentiment figure is exactly what makes it disappear.
Can I see for myself which source causes the error?
Yes. Every finding carries the question, the engine, the date and the list of sources the engine fetched for that answer. We do not point at one culprit — that would be a judgement on texts we have not read — but the list is complete, sorted by how often a source sits under your findings.
Read on
Free, no account and no card.