How AI Hallucinates About Hungarian Companies: Patterns From 48 Measured Queries

When AI doesn't know the answer, it rarely says "I don't know." It says something instead. A made-up company name, a wrong address for a real clinic, a provider from another city — stated confidently, as if it were fact. I asked four free models about Hungarian dental practices across 48 queries in May 2026. The answers fell into four recurring error patterns. None of them were caused by the practices themselves — and that's exactly why they're dangerous.

I've anonymized every company name here, because my goal isn't to single anyone out or put anyone in the stocks. The pattern is the point, not the individual case. I describe the full set of raw answers and the scoring logic on the methodology page, and anyone can reproduce the measurement in a few minutes.

What kinds of errors did I see?

Reading through the raw answers, four patterns came back again and again. Each is a different face of hallucination: the model doesn't stay silent, it fills the gap — it just fills it wrong.

The first: the pure invention. Several models confidently supplied clinic names together with a web address that, when looked up, led to a non-existent or misspelled URL. One answer recommended a Budapest clinic with a web address in which even the word "dental" was misspelled in the domain — so the model manufactured not just the name, but the address that went with it:

"[a Budapest clinic] — excellent services, including implants." The web address provided pointed to a misspelled, non-existent domain.

The patient would type in that address and end up nowhere. Yet the model stated it with exactly the same confidence as its real data.

The second: the right name, the wrong data. This one is more insidious. The model named a real, existing practice — but with incorrect accompanying data: the wrong district, an invented street address, or a specialty the practice doesn't even offer. One model went so far as to give specific street addresses for downtown clinics, "near Deák Square," "next to Váci Street" — with confident precision, even though no source backed up any of it. The patient trusts the real name and sets off based on the wrong address.

The third: the other city, the wrong audience. When I searched for a dentist in a specific district, the model would more than once list international clinics aimed at tourists and foreign patients — with airport transfers and foreign-language websites. For a local patient looking for the practice around the corner, that's useless. The model pulled the best-known, tourism-optimized names from its "dentistry in Budapest" training data, and failed to notice that the question was about one specific district.

The fourth: the vague deflection instead of a specific answer. Some models named no one at all, offering advice instead: check the reviews, ask for a consultation, compare prices. One answer began like this:

"I'm sorry, but I'm a model that provides general information, and I can't offer real-time updated information or personal recommendations."

Of the four, this is the most honest, because it invents nothing. But the patient still got no answer — only more homework. And in practice the ending was the same: the patient moved on to another tool or to a competitor's name.

On top of that, the four models gave four different lists. It's not that there's one stable winner and it just isn't you — it's that there's no stable winner at all. To the very same question, one model gave a made-up name, another a tourist clinic, the third a deflection. The terrain isn't claimed. It's empty.

Why is this more dangerous than being invisible?

Here's the part most articles leave out. If a business simply doesn't appear in the AI's answer, that's bad — but at least it's a visible, measurable gap. Hallucination is worse than that, because the customer doesn't know they got a hallucination.

Picture it from the patient's side. They ask ChatGPT, get five names with five websites, and the answer sounds confident. How would they know the third name is invented and the fifth belongs to a practice in another city? They wouldn't. They trust it, because it sounds accurate. So they either call a non-existent clinic, show up at the wrong address, or simply end up at a competitor the model happened to rank higher.

With traditional search, at least you can see what position a company holds, and there's something to improve. The AI conversation is closed: the company can't see what's being said about it, and it can't correct a wrong mention with the press of a button. That's why it's not enough to ask, "Do I appear?" You also have to check, "Do I appear correctly, or is the model getting me wrong?"

What can you do about it?

The good news is that hallucination isn't random noise — it has a pattern, and there's something you can do about it. The model invents data when it can't fill the gap from a reliable source. So the work is about filling the gap yourself, before the model does it for you.

The first step is entity consistency. The name, the address, the phone number and the opening hours have to appear in exactly the same form everywhere: on your own website, in your Google Business Profile, in directories, on social media. If these details contradict each other across the web, the model has no single fixed point to rely on — and it would rather invent one. There's a Hungarian wrinkle that matters here too: only 34% of domestic SMEs even have a Google Business Profile, which means most are missing the very structured baseline data the model could reliably work from.

The second is structured data: machine-readable company information on your website, so the model isn't forced to interpret but can read a fact. The third is the freshness of external sources — maintaining your reviews, mentions and directory listings, because the model typically draws on those, not on the text of your website.

And the fourth is regular self-checking. A monthly, dated test shows whether the picture has changed: whether the name now appears, whether accuracy has improved, whether an earlier incorrect mention has gone away. I put together the step-by-step guide in the article what ChatGPT says about your business, and the full logic broken down across seven dimensions is on the methodology page.

It's important, though, to see the limit clearly too. Hallucination can't be eliminated entirely — no method guarantees the model will never again get something wrong about you. That isn't the realistic goal. The realistic goal is that when the model talks about you, it finds primary, fresh and consistent sources — so the chance of error drops to a fraction, and the chance of a correct mention rises many times over. This isn't a promise of perfection. It's the deliberate filling of the gap, so the model doesn't have to guess.

And if you'd like a dated, verifiable measurement of what AI invents about your company today — and where the gaps are that need filling — let me know on the contact page. I'll measure what artificial intelligence says about you, and show you where the work begins. With measurements, not promises.

Kindl Dávid · Published: May 27, 2026 · Updated: June 4, 2026