When an AI Says Something Wrong About Your Brand
AI models state wrong facts about companies with total confidence. Here is why it happens to brands specifically, and a way to check if it is happening to yours.
Ask ChatGPT what your company does and it might answer with total confidence, in clean sentences, and be wrong. Not vague, not hedged. Wrong, stated as fact, in the same tone it uses for things it gets right. Maybe it says you sell a product line you discontinued years ago. Maybe it merges you with a company that has a similar name in a different country. Maybe it invents an acquisition that never happened, or puts your headquarters in a city you have never had an office in. This is a hallucination, and if you have not gone looking for one about your own brand, there is a decent chance one exists somewhere in these systems right now.
What a hallucination actually is
A generative model does not look things up by default. It predicts the next most plausible chunk of text given everything it was trained on, one token at a time. Most of the time that prediction lines up with reality, because reality is what most of the training text described. But the model has no internal notion of "I am not sure" the way a person does. It does not check a fact against a source before it writes it down, unless it is actively retrieving from the live web while it answers, the way Perplexity typically does and ChatGPT does when browsing is switched on. Absent that retrieval step, the model is producing the statistically likely continuation of a sentence, and a fluent, confident, wrong continuation looks identical, on the surface, to a fluent, confident, correct one.
That is the uncomfortable part. There is no visual cue, no lower-confidence font, no asterisk. A hallucinated sentence about your brand reads exactly like an accurate one, because it was generated by the exact same process.
Why brands are an easy target
Some subjects are hallucination-prone by nature, and a company is one of them, for a few specific reasons.
- Sparse or contradictory training data. A smaller or newer company simply does not have much written about it on the open web. Where there is little signal, the model fills gaps with whatever is statistically plausible for "a company like this," based on similar businesses it has seen described elsewhere. The result reads like a real description. It just is not a description of you.
- Name collisions. If another company, in another country or industry, shares your name or something close to it, the model can blend the two. This gets worse the more generic the brand name is.
- Stale information with no expiry date. A model's absorbed knowledge comes from a training cutoff. It has no way of knowing that the product you led with is discontinued, that your pricing changed, or that you were acquired. Unless something is actively retrieving current information, the model answers from a snapshot that was already out of date the day it started serving traffic.
- Filling a gap with a plausible neighbor. A model does not say "I don't know" nearly as often as it should. It tends to produce the answer that a company in your category would plausibly have, which can mean inventing a certification you don't hold, a location you don't operate in, or a feature your competitor has and you don't.
None of this is malicious and none of it is really a bug in the sense of being a fixable defect. It is a predictable consequence of how these systems generate text when they are not grounded in a live lookup.
A falsifiable way to check whether this is happening to you
You do not need a tool to find out whether a model is saying something wrong about your company. You need four questions and four assistants.
Ask ChatGPT, Claude, Gemini and Perplexity the same set of plain factual questions about your own business: what does the company do, who owns it, where is it based, and what does its main product or service cost. Write down each answer verbatim. Then compare each one, line by line, against what is actually true.
Do this across all four assistants rather than one, because they draw on different training data and different retrieval behavior, and they do not make the same mistakes. One might get your ownership structure wrong while getting your location right. Another might be accurate on everything except your pricing, absorbed from an old page that was never taken down. A single check on a single platform tells you almost nothing about the other three, and it helps to repeat the question more than once per assistant: an inaccurate detail that shows up once might be a fluke, while one that shows up consistently across several attempts is telling you something real about what the model absorbed.
What you can actually do about it
Here is the honest part, and it is not a comfortable one. There is no submission form for correcting a model's absorbed knowledge, no "report an error" button that reaches into the training data and fixes a sentence. You cannot email a lab and have them patch what a model believes about your headquarters address the way you could request a correction from a directory listing.
The only lever that exists is indirect: making accurate, specific information about your company exist clearly and repeatedly across the web, so that the next training run, or the next live retrieval, finds the correct version instead of a vague or wrong one. That means your own site stating plainly what you do, where you are based and what changed. It means outdated pages actually being updated or removed rather than left live next to the current ones. It means the same facts showing up consistently across your site, your documentation, press mentions and any directories that list you, instead of several slightly different versions of your own story scattered across the web.
This is slow. It is not a fix you apply this week and see resolved next week. An inaccuracy absorbed at one point can persist for a long time regardless of what you publish afterward, and there is no shortcut here. Anyone telling you otherwise is not being straight with you.
Where measurement fits, and where it stops
This is also where we should be precise about what a tool like PSentry does and does not do. PSentry runs your prompt set against ChatGPT, Claude, Gemini and Perplexity and reports back what each one actually said about your brand, in each market and language you operate in. If a scan surfaces a sentence that misstates what you do, invents a product, or gets your ownership wrong, that sentence is sitting right there in the raw scan text, because that is what the model said.
What PSentry does not do is decide, on your behalf, that a given sentence is a hallucination. There is no classifier flagging inaccurate claims, no alert firing when something looks wrong, no dedicated detection layer separating fact from fabrication. The scans measure mentions and visibility; judging whether a statement about your company is true is still a human task, because it requires knowing the actual facts about your own business, which only you reliably know. A tool can put the model's exact words in front of you across every platform and market you care about. It cannot tell you, on your behalf, that those words are wrong.
Frequently Asked Questions
Does PSentry detect hallucinations for me?
No. PSentry surfaces the raw text of what each AI platform says about your brand during a scan. There is no hallucination-detection feature, no automatic flagging and no classification layer. A human still has to read the output and judge it against the facts, the same way you would when checking any other written claim about your company.
Is a hallucination the same thing as a bad opinion about my brand?
No, and the distinction matters. An unfavorable but accurate description is not a hallucination. A hallucination is a factual claim that is simply not true: a wrong founding date, a discontinued product described as current, an office that does not exist. Sentiment and accuracy are two different things to check for, and a scan surfaces both without separating them.
Can retrieval-based tools like Perplexity still hallucinate about my brand?
Yes, though it happens differently. Retrieval reduces the risk because the model is grounding its answer in something it just fetched rather than only what it absorbed during training, but it does not eliminate it. If the source it retrieves is itself outdated, wrong, or a page about a different company with a similar name, the answer inherits that error.
If I fix the wrong information on my own website, how fast does the model's answer change?
There is no fixed timeline. A model answering from training data will not reflect your change until it goes through a new training cycle, on no published schedule you can check. A model that retrieves live sources can pick up a correction faster, but only once your updated page is established enough to outrank the old, wrong version in whatever the model finds.
Should I ask about my competitors too, not just my own brand?
Worth doing at least once. If a model consistently gets basic facts about your close competitors right while getting yours wrong, that points to how little or how confusingly the web talks about your specific company, rather than something inherent to the topic.
Does a hallucination about my brand mean I do not exist as a recognized entity to these models?
Not necessarily. A model can have a clear, consistent notion of your company as a distinct entity and still get specific facts about it wrong. Entity recognition and factual accuracy are related but separate things worth checking.