What happened

In December 2025, Mateusz Makosiewicz at Ahrefs ran an experiment. He invented a brand called Xarumei, built it a site with a proper FAQ, seeded the web with several mutually contradictory accounts of it, and then put 56 questions to the major AI engines.

The results diverged sharply:

  • Perplexity got roughly 40% of the questions wrong, confused Xarumei with Xiaomi, and insisted it made smartphones
  • ChatGPT-4 and ChatGPT-5 got 53 to 54 of 56 right, and cited the brand’s official FAQ in 84% of their answers
  • Claude produced no hallucinations, but barely grounded itself either — it declined to commit to much of anything
  • Copilot handled neutral questions reasonably and fell apart on leading ones
  • Gemini started out sceptical and was eventually carried along by the planted material (the write-up gives it no specific error rate)

One correction worth making: secondary coverage of this experiment often reports the 40% figure as applying to “Gemini and Perplexity.” The original gives that number for Perplexity only; the Gemini finding is descriptive.

Why it matters to you

What the experiment actually measured isn’t which engine is least capable. It’s this: given a choice between a vague truth and a specific fiction, AI engines frequently take the specific fiction.

A confident, detailed, number-laden fabrication beats a real account that hedges. ChatGPT held up not because it reasoned better but because it went and read the official FAQ — a page where every basic question had been answered definitively, leaving no gap for anyone else to fill.

Which puts the question back on you. Does your brand have that page? How pricing is calculated, what you do and don’t cover, when the company was founded, who runs it, how you relate to the similarly named business down the road. If you don’t write these down, someone else eventually will — and when an AI needs the answer, that other version is the only one on the table.

None of this requires a bad actor. An old price list you never took down, a founding year one publication got wrong, a blog post that used you as an example and described you backwards — that’s enough. The three contradictory sources in the experiment were planted deliberately. Yours grew on their own. The effect is identical.

What to do about it

What needs doing is a genuinely dull page: write your brand’s basic facts down, specifically enough that nothing is left open.

Specific means figures, dates, and clear statements of what is and isn’t true — not “we offer a range of flexible solutions,” which says nothing an AI can use. It can’t cite a sentence like that, so it turns around and finds someone else’s more concrete version. If you’re going to sit down and do it, our piece on FAQ and Q&A readiness covers the writing in full.

The difficulty isn’t the writing. It’s that you first have to know how the AI engines currently describe you — which questions they get wrong, where the wrong version came from, and which page would actually override it. That means asking, recording and comparing, round after round, and because AI answers carry real randomness, what you see in a single attempt doesn’t count as evidence. Writing one FAQ page and calling it finished is usually just a new way of staying wrong.

If you want to know how the AI engines describe your brand right now, we can run a round of questions first.

Further reading