In 2002 I asked a question that sounded almost naive at the time. What does a buyer actually see of the hundred biggest Hungarian companies, when they look them up online? We measured it, a journal even cited it, and then for a long time I thought that was a closed chapter.
This autumn I asked it again. Same question. The difference is who answers it now, because the buyer doesn’t really search anymore, they ask, and the answer comes from an AI.
So we asked. Eight AI surfaces, the questions a normal buyer would ask about 65 companies selling in Hungary: what does delivery cost, how long is the guarantee, can I send it back, which number do I call. ChatGPT, Claude, Gemini, Perplexity with web search, the ChatGPT and Gemini apps, and Google’s AI Overview and AI Mode. Then we sat down and checked every single answer against the company’s own website.
Before any of the errors, one number stopped me. For 95% of our questions, Google put an AI-written summary on top of the page. Think about that for a second. The first sentence a buyer reads about your brand is very often not written by you. And nobody in your company approved it.
Where I was wrong
I have to be honest about what I expected, because I think most of you expect the same.
I expected fabrication. Every week there is a story about a chatbot inventing a court case or a refund policy, so I was sure the big risk for a company is that the AI just makes something up.
It doesn’t, mostly. Completely invented facts showed up in roughly one answer out of fifty or sixty. About 1–2%.
But here is the big “but”. In about one answer out of twelve, around 8%, there was a fact that did not match the company’s own site, and it was the kind of fact that makes a buyer decide wrong. The wrong delivery fee. An old deadline. Another phone number. Every surface did it, from 2.6% for ChatGPT up to 11.9% for Gemini, and the AI Overview on top of Google was in the middle, roughly one in twelve.
And the AI almost never warned anybody. 84% of the wrong claims came out as plain fact. In one case out of twenty-five it said maybe check the official site. One out of twenty-five.
Eight percent does not sound like much. Until you remember that a buyer usually needs four or five of these facts before they order anything.
The part that actually changed my mind
So then we went back to every material error and asked a boring question: where did this wrong value come from? Was it on the pages the AI cited, or not?
30% of them came from the company’s own website. An old page, or two pages saying different things. Another 8% from a foreign sister company or the international group. 17% from some third party, and 36% were not on any cited page at all, those the AI added itself.
Put the first two together and you get four errors out of ten that the company could fix on its own, today, without asking anybody.
And then the finding I still keep thinking about. At more than half of the companies, their own pages disagree with each other on at least one fact. Different guarantee, different cancellation period, different delivery fee, different phone number.
In those cases the AI is not wrong. It is choosing.
Nobody did anything stupid here, I want to be clear about that. Legal writes the terms. Sales writes the product pages. Customer service writes the FAQ and the delivery page, and where the error came from the company’s own site, five or six times out of ten it started exactly there, on a customer service page. Marketing writes the home page. Head office writes the group page. Everybody did their job. Nobody’s job was to check if all of it says the same thing.
In October I was in London at the Semrush Spotlight conference, and I heard the same thing from the other side of the table. Andrew Warden, VP of Marketing at Adobe, said the worst outcome is when a brand is invisible in AI answers, and that this happens when the company’s own data does not match what others say about it. He named product documentation and the legal team, which I did not expect from a marketing stage. The same day LinkedIn’s VP of Marketing told us they built a joint group from content, product marketing, digital and events, because no channel can fix this alone.
I sat there and thought, okay, our data says the crack starts one step earlier. Before your data disagrees with the world, it already disagrees with itself.
Two things that are not Hungarian at all
First, language. We asked the same 50 questions in English too. In English the AIs gave foreign facts about the Hungarian company five times as often, 26% of answers against 5.5%. Nearly half of those were the international group’s data, presented like it applies here. So if your partner, a foreign buyer, or an AI agent asks in English, they simply meet a different company. If you run a company in more than one country, I would check this one first.
Second, agents. More and more, it is not the buyer who looks around but an agent doing it for them. We gave Claude Code and Codex buying tasks. With one short rule, something like “check the official site first, flag what you can’t confirm, invent nothing”, the decisions built on a wrong fact went from 9% and 16% of tasks down to 0% and 4%. Small sample, not statistically proven, I know. But you can see where it goes: the careful agent stops more often and asks somebody to check.
Then the part I did not see coming. The same Claude, with the same 130 tasks, could not open a single page of the company 32 times when it ran from a user’s computer, and 13 times when it ran from Anthropic’s own servers. Whether the buyer’s agent gets in is not decided in your robots file. It is decided in your security settings. And from the inside you mostly can’t even see it.
So who answers for it?
Most of the advice going around is technical. Add this file, add that markup, optimise the page. In our data, the technical readiness of a site made no visible difference to how accurate the answers were. What made the difference was much less exciting: does the company say one thing about itself, everywhere, or not.
That is not a technical question. It is the old question I keep asking every leadership team I work with. Who decides, based on what, and who answers for it?
If I had one month, I would do four things. I would write down ten to thirty questions a buyer really asks about us, put them into Google’s AI Overview and at least two chatbots, and save the answers with the date. I would trace every wrong one back to its source, our page, the group’s page, somebody else’s, or nowhere. I would give the whole thing one owner, together with brand, digital, customer service, and yes, legal. And I would ask the same questions again every month.
You don’t need a new tool for any of this. You need one person whose job it is.
In 2002 the question was what buyers see. Now it is what the AI tells them. More than twenty years, and the question did not move. Only the place where the buyer looks for the answer did.
What changed for me is where I start the conversation. I don’t start with what the AI says anymore. I start with who in the company is responsible for saying the same thing everywhere.
If you checked what AI says about your company this week, I would really like to hear what you found.
Summary, charts and method: https://arcanian.com/research/what-ai-says/2026-11. We publish companies without names. The research continues every month.


