“ChatGPT Named Us!” And Then Nobody Read The Rest
Somebody drops a screenshot in the team channel. There’s your company, sitting inside a Gemini answer as a source.
I get the feeling. Two years ago being quoted by a model was a party trick; now a PR lead puts it in the monthly report, and fairly so.
But watch what happens next: we confirm that we were cited, and then we move on. Almost nobody reads what the citation actually said.
So how often is the sentence next to your name even right?
Somebody finally checked. A group of newsrooms sat down and read thousands of those answers on purpose, with a scoring sheet.
What Happens When Journalists Grade 3,000 AI Answers
The study came out of the European Broadcasting Union (EBU) and the BBC, who pulled in 22 public service broadcasters across 18 countries and 14 languages (the report).
Journalists from those newsrooms went through more than 3,000 answers from ChatGPT, Copilot, Gemini and Perplexity, scoring each one on four things: accuracy, sourcing, telling fact from opinion, and context. (So this isn’t vibes from a weekend spot-check — these are people who verify attributions for a living, reading answers one at a time.) Here’s what came back:
- 45% of answers had at least one significant issue
- 31% had serious sourcing problems (missing, wrong, or misleading attribution)
- 20% had significant accuracy problems, like hallucinations or out-of-date information
That middle number is the one to sit with. A means the who said this part is broken — a source invented outright, or another outlet’s statement handed over as yours. (And a , if the word is new to you, is the model saying something untrue with total confidence; that plus stale information covers one answer in five.)
The Models Aren’t Equally Bad At This
The trouble isn’t spread evenly, either, which is the bit I didn’t expect.
Google’s Gemini came out worst by a distance: roughly 72–76% of its answers carried a significant sourcing problem. Call it three times in four where the attribution is off somehow.
ChatGPT, Copilot and Perplexity kept their sourcing problems generally under 25%.
So does “we got cited by AI” mean one thing? Obviously not. Which model cited you moves the odds that the citation is even true.
That’s easy to lose if you’re only counting mentions. A counter tells you your name appeared; it says nothing about the sentence sitting next to it.
Why A Bad Citation Lands On You, Not On The Model
Here’s the part that costs you something.
Who actually audits a claim that arrives with a trusted name in front of it? Nobody, mostly. “According to the BBC”, “per their announcement” — the name does the believing on our behalf.
That’s exactly what EBU and the BBC flag: when an AI puts a trusted brand behind a claim, users tend to believe the claim even when it’s wrong, and the reputational damage falls on the cited news organisation or brand.
Think of the guy in every office who repeats things with “the director said…” attached. Nobody checks. The prefix does the work. AI citations run on the same wiring, except the prefix can be your company, and nobody asked you first.
Getting named is nice. Getting made the guarantor of something false is a different transaction entirely.
Meanwhile, People Keep Leaning On AI For News Anyway
The awkward part is that this foundation is wobbling while the traffic on top of it grows.
The Reuters Institute at Oxford has the share of people who’ve ever used generative AI jumping from 40% to 61%, with weekly use nearly doubling from 18% to 34% (Digital News Report 2025).
And in the same survey, people say they find AI-generated news less trustworthy than the other kind. Nervous about it, using it more. Both at once.
So the number of people who might meet a broken citation about you keeps going up, while roughly a third of the sourcing underneath stays shaky.
Conclusion: Watch How You Get Quoted, Not Just Whether
The takeaway is that being cited isn’t the finish line. It’s the point where brand management picks up a new job: checking whether the quote is right.
And one check won’t hold. Model answers wobble on the same question, and the models clearly have different habits — so “I looked once and it was fine” ages badly.
Start here, this week: take your two or three main topics, ask ChatGPT and Gemini about them the way a customer would, and read what gets attributed to you. Not whether you appear. What you supposedly said. Then make it a standing check rather than a one-off, so a wrong attribution gets caught while it’s still one answer and not a pattern.
Maybe I’m overweighting a single study here — it’s news content, graded by news people, and your category may behave better. But 31% is a big enough number that I’d rather find out on purpose than by accident.
Sources
- EBU & BBC (2025), “News Integrity in AI Assistants — An International PSM Study”, ebu.ch
- Reuters Institute for the Study of Journalism (University of Oxford, 2025), “Digital News Report 2025”, reutersinstitute.politics.ox.ac.uk
