AI is already talking about your brand. Do you know what it says, and who answers for it?
We asked eight AIs about 65 companies selling in Hungary, the way a buyer would. We looked at what they say, where they get it from, and what leadership can do about it.
Research · · Arcanian
In short: We asked eight AIs about 65 companies selling in Hungary and checked every answer against the company's own website (November 2026). About 8% of answers contained a misleading fact, and about 1–2% contained a completely made-up one. Nearly a third of the wrong facts come from the company's own website: at more than half of the companies, their own pages do not agree with each other.
In 2002 we measured what buyers see of Hungarian companies. By 2026, what AI says about them.
How big is the problem?
about 8%
7.9% (6.6–9.3)
Misleading fact (differs from the company's own website)
about 1–2%
1.6% (1.1–2.3)
Completely invented fact
Someone searches for a shop on Google. Before they click a single result, an AI-written summary of the company sits at the top of the page: delivery charges, guarantee, opening hours, customer service.
For 95% of the questions we asked about companies in Hungary, that summary appeared (730 of 766 questions, 16 November 2026). So the first sentence a buyer reads about a brand often does not come from the brand.
Nobody approved that sentence. The question is whether it is true.
How often does AI get companies wrong?
Every surface we tested made mistakes. Between 4–12% of answers about companies contained at least one wrong fact that would lead a buyer to the wrong decision: not ordering when they could; calling the wrong number; paying more than they think; or being put off the company.
How often does AI get it wrong? Misleading facts in answers, by AI surface
- ChatGPT (developer API) 4.0% (1.8–6.6)
- ChatGPT app 5.9% (3.4–8.5)
- Perplexity 6.2% (3.5–9.2)
- Google AI Overview 8.0% (4.4–11.7)
- Claude 8.8% (5.5–12.3)
- Google AI Mode 9.5% (6.1–13.2)
- Gemini app 9.9% (6.2–14.2)
- Gemini (developer API) 11.7% (7.7–16.0)
Google's AI Overview, which sits at the top of the page for almost every question, was wrong in roughly one answer in twelve.
Across the four sectors we studied, the AIs that answer with web search made this kind of error in about 8% of answers on average (95%: 6.6–9.3%), about as often as Google's AI Overview. By AI: ChatGPT 2.6%, Perplexity 8.3%, Claude 8.7%, Gemini 11.9%. That sounds small. For a buyer who needs four or five facts to make a decision (a charge, a deadline, a guarantee, a returns policy), it is not.
Completely made-up facts are rare: roughly one answer in fifty to sixty (about 1–2%). So AIs seldom invent things, but they often get the details wrong, and those details mostly come from what companies publish themselves.
We found no difference in accuracy between the apps and the APIs. The ChatGPT and Gemini apps made as many mistakes as the developer versions of the same AIs.
Does AI tell you when it is unsure?
Rarely. The AI stated 84% of wrong claims as fact (524 of 625 wrong claims). Only once in twenty-five did it say it was unsure, or suggest checking the company's official website.
So a buyer cannot tell a good answer from a bad one. Both sound equally confident.
The AIs also almost never decline to answer. When they do not know something, they say something anyway.
Where does the wrong fact come from?
For every material error we checked where it could have come from: whether the wrong value appears on the pages the AI cited (347 errors).
Where do the wrong facts come from?
30% of the wrong facts come from the company's own website, 8% from the group's website, 17% from third parties, and 36% were added by the AI.
- The company's own website (outdated or contradicting itself) 30%
- A foreign sister company or the international group 8%
- A third-party site 17%
- Added by the AI (on none of the cited pages) 36%
- Other (no citation / wrong company / unreadable page) 9%
Four errors in ten can be fixed by the company itself or its group.
Think of old terms and conditions, a half-updated delivery page, or a forgotten customer service subpage. Two of the company's pages say different things, and the AI quotes the wrong one.
Of the remaining errors, some come from third parties and some from the AI itself. For these the answer is regular monitoring: you need to know about them before the buyer notices.
What changes when the question is asked in English?
We put the same 50 questions to the four chatbots in English as well. In English, we got foreign facts about the company in Hungary five times as often: in 26% of answers, against 5.5% for the Hungarian-language questions. Material errors also became more frequent: 24.5% instead of 14.5%. (Both differences hold up statistically. On these 50 questions the baseline is higher than in the study as a whole.)
Nearly half of the foreign facts are international group data, and one in five is a foreign sister company's terms, presented as if they applied in Hungary.
When a buyer, a partner or an AI agent asks about the company in English, they meet a different company.
What happens when AI decides for the buyer?
We tested two AI agents that businesses use, Claude Code and Codex. We gave them 25 buying tasks: collect four facts about a company, then make a decision.
There is a visible direction, but the sample is small and the difference is not statistically proven. Without extra rules, the agents based a decision on a wrong fact in 9% and 16% of tasks respectively. When the agent was given rules ("look at the official website first, flag anything you cannot confirm, do not make anything up"), this fell to 0% and 4% respectively.
In return, the agents stopped more often and said: this needs checking. A safe agent does not decide for you. The question is who checks when it stops. That is not a technical question. It is a leadership decision.
What is more, the buyer's agent may not even get in. One of the 65 companies blocks AI agents acting on a user's behalf in its robots.txt file. But more than a quarter of the websites already stop, on the very first request, any automated visitor they do not recognise. Whether the buyer's agent gets in depends on where it comes from and whether the website recognises it. One agent read a page that another could not. The same Claude, running from the user's own computer, got stuck two and a half times as often as when it ran from Anthropic's own servers: in the same 130 tasks, it could not open any of the company's own pages 32 times and 13 times respectively. The robots.txt file does not decide this. The security settings do.
Who visits your website today?
We looked at one week of a Hungarian service company's website (10–16 November), using data from its content delivery network (CDN).
- More and more bots present themselves as AI. In one week, 4250 requests came from bots that said they were AI crawlers, 55% more than the week before. Their names alone cannot tell you how many of them were genuine.
- AI agents acting on a user's behalf are already here, or at least they say they are. The website was also requested by bots that presented themselves as the agents of ChatGPT, Claude, Perplexity or DuckDuckGo, looking for an answer to a person's question.
- But not all of them are who they say they are. We examined the blocked "agent requests": none came from the AI companies' networks. They came from rented cloud servers and were looking for passwords and configuration files. Anyone can write a bot's name into a request, and attackers pose as AI because many websites let AI in.
If the machine cannot read your own website, it answers about you from somewhere else. And your own traffic data does not tell you what the agent sees or can do on your site, or even whether it really was an agent. You can only check this from the outside: ask the public AIs and agents, and look at what they answer and where they get it from.
What we could not confirm
There is plenty of advice in the market about how a brand gets into AI answers. Our own measurements did not support it:
- When asked about companies in Hungary, the AIs cite the company's own website most often. Reddit accounts for less than half a percent of citations.
- Blocking crawlers does not remove a company from AI answers. It may reduce the share of the company's own website, but other people's text takes its place.
- Having an llms.txt file, structured data or an FAQ page showed no link with whether the AI cites the company's website. And the facts the AIs get wrong are not the ones structured data describes.
- The ranking in a single query tells you nothing: the AIs' order changes from run to run.
- Machine readiness on its own did not improve how AI describes a company. In our sample, an llms.txt file, rules for crawlers and similar technical preparation did not go with more accurate AI answers. What mattered was what the company's own pages say, and whether they all say the same thing.
- A company can get into an AI answer even with a self-published ranking list written a week earlier. So a simple count of mentions does not tell you whether the AI is telling the truth about you.
Who is responsible for your company saying the same thing everywhere?
AI does not invent a second company. It repeats what the company's different departments publish. The legal team writes the terms and conditions, sales writes the product pages, customer service writes the FAQ and the delivery page, marketing writes the home page and the "About us" page, and head office writes the group-level page.
- Where the error comes from the company's own website, five or six cases in ten start on a customer service page, two or three on a product page, and one on a legal page.
- At more than half of the companies, their own pages disagree on at least one fact: a different guarantee period, a different cancellation period, a different delivery charge, a different phone number. In those cases the AI is not wrong; it is choosing.
- A third of the errors that come from the company's own website repeat a value that one of the company's pages states and another contradicts.
This is not one department's job. The question is whose job it is to make every page of the company say the same thing.
Others see this too. In October 2026, at the Semrush Spotlight conference in London, Andrew Warden, VP of Marketing at Adobe, called it the worst outcome when a brand is invisible in AI answers. In his view, this happens when a company's own data does not match what outside sources say about it, and he singled out product documentation and the legal team. On the same day, LinkedIn's VP of Marketing described how AI search had led them to form a joint working group from their content, product marketing, digital and events teams, because no single channel can solve this alone.
Our research suggests the problem starts earlier. At more than half of the companies selling in Hungary, the company's own pages do not even agree with each other, long before outside sources come into it.
What can leadership do?
- Measure the baseline. Ask 10–30 questions about your company the way a buyer would, at least in Google's AI Overview and two chatbots. Write down the date and the answer.
- Compare it with your own website. Where the AI is wrong, find out where it gets the information. Nearly a third of errors come from the company's own old or contradictory pages.
- Decide who fixes what. The company fixes its own website; the parent company fixes the group website; third parties fix their own websites. What comes from the AI itself cannot be fixed, only monitored.
- Give it an owner. One person responsible, working with the brand, digital and customer service teams. Decisions about AI do not belong to IT.
- Measure again. With the same questions, every month. One measurement is a snapshot; the accuracy of the same AI can change within a week.
- If you use agents, decide in advance who checks when they stop.
- Have a written policy on which AI crawlers may come in, and give it an owner. If nobody has decided, the default settings have decided. What we recommend:
- On your public brand and fact pages (about us, services, prices, delivery, guarantee, contact details, terms and conditions), let every AI crawler in: search crawlers, agents acting on a user's behalf, and training crawlers too. More than a third of errors come from the model's memory. If the training crawler cannot read your own website, the model learns about you from old or other people's data.
- Exclude paid, closed or protected content (client material, subscription content) from training; decide separately there about search crawlers and agents.
- Do not let them in based on their name. Anyone can fake a bot's name, but a content delivery network can check whether a visitor really is who it claims to be. You can recognise an attacker by what it looks for: passwords, configuration files, system paths. Block those paths.
- If resources are tight, two things are worth considering. State in your robots.txt what AI may use your content for: search, AI answers, training (this is what the Content-Signal line is for). And give machines a clean text version of your pages (Markdown); many content delivery networks offer this with a single switch. Both are cheap good practice. We did not measure whether the AIs respect the signal, or whether they answer more accurately from the text version.
- Review it every year, and measure what actually reaches the machines: the setting and its effect are not the same thing.
How did we measure it?
What did we count as an error? We counted an AI answer as wrong if it stated at least one fact about the company that contradicts the company's own website and would lead a buyer to the wrong decision: believing the wrong price, deadline, contact detail or condition. Inaccurate but harmless details were counted separately, as minor errors. Where the wrong value also appeared on the company's own pages, we counted it separately, as the company's own error. A person decided the disputed cases. Every figure is given with a 95% uncertainty range.
65 companies selling in Hungary, in four sectors: marketplaces, drugstores and pharmacies, home furnishing, consumer electronics. 696 branded and 100 unbranded buyer questions, in Hungarian. Eight AI surfaces: ChatGPT, Claude, Gemini and Perplexity with web search; Google's AI Overview and AI Mode (viewed from Hungary, logged out); the ChatGPT and Gemini apps. Collection: 10–16 November 2026. Benchmark: the companies' own websites; for time-sensitive facts (price, charge, change of ownership), what was valid on the day of collection.
As far as we could find, this is the first study in Hungary to test whether AIs tell the truth about companies using live search, Google's AI surfaces and scoring checked against a human-reviewed sample.
What we did not measure
- An indicative measurement. It shows trends and orders of magnitude; on its own it is not enough for a budget decision.
- One point in time. One run per surface is a snapshot. From the December round onwards we will repeat it monthly, with the same questions.
- Machine scoring, one human reviewer. A machine scored the answers, and one expert decided the disputed cases. On a blind sample of 68 items, the machine was much stricter than the human: what the machine judged misleading, the human mostly considered a minor error. That is why we give two figures. The roughly 8% is misleading facts that differ from the company's own website, by the machine's standard. The roughly 1–2% is completely made-up facts, by the machine standard calibrated to human judgements.
- Not every question can be measured. Where the company's website does not answer the question, or sat behind bot protection, we did not score the answer.
- The traffic example comes from a single website, over one time window; it says nothing about proportions across the market.
- The research looked at AI answers, not at how many people read them, or what a wrong answer costs a company.