Being Cited by AI Doesn't Mean It Told the Truth
It’s common to see someone celebrate a screenshot: “look, Perplexity cited me, it listed my site as a source.”
Almost no one stops to read what the AI actually said based on that citation. The assumption is: if there’s a source, what’s being said must be true.
That exact assumption — “if it’s cited, it’s true” — is the one a group of researchers decided to put to the test. And the result isn’t reassuring.
The experiment
Pranav Narayanan Venkit, Philippe Laban, Yilun Zhou, Yixin Mao, and Chien-Sheng Wu (a team with backgrounds including Salesforce AI Research) designed two studies, published at ACM FAccT 2025 (one of the most respected conferences on fair and accountable AI):
- Human study: 24 people — with strong academic backgrounds (around 62% PhD candidates or postdoctoral researchers, the rest industry experts) — used You.com, Bing Copilot, and Perplexity for 90 minutes each, comparing the experience to traditional Google.
- Automated evaluation: 303 real questions, assessed across 8 distinct accuracy and grounding metrics.
Venkit et al., “Search Engines in an AI Era”, ACM FAccT 2025 — the goal was simple: measure whether the citations these engines display actually support what the AI claims, or whether they’re just decoration that gives an appearance of rigor.
The uncomfortable number
Here’s the rate of citations that actually support what was said, by engine:
| Engine | % of citations that genuinely support the claim |
|---|---|
| You.com | 68.3% |
| Bing Copilot | 65.8% |
| Perplexity | 49.0% |
Perplexity — the engine much of the market treats as the most rigorous, the one that “always shows its sources” — was the worst of the three. Essentially a coin flip.
And it doesn’t stop there: 31.6% of Perplexity’s claims and 30.8% of You.com’s had no real source behind them at all — nearly 1 in 3. Bing Copilot fared somewhat better on this specific point (23.1%), but that’s still close to 1 in 4 unsupported claims — even though all three engines run on the same technology (RAG, retrieval-augmented generation) that’s supposed to solve exactly this problem.
What’s most surprising about the study
Two findings almost no one mentions when discussing this paper:
- Bing Copilot shows sources it doesn’t even use: 36.2% of the links it lists are never actually cited in the generated text. They’re there purely to create a sense of backing.
- Perplexity gave one-sided answers (showing only one side) on 83.4% of debate-style questions — and did so with 81.6% overconfidence, even when that single side was wrong.
And perhaps the most revealing data point from both studies combined: people verify AI answers far less than they verify Google results. On average, they open just 2 sources to confirm an AI answer, versus 12 when using traditional Google.
One participant, when asked whether it felt like the AI was using citations to legitimize what it generated, put it this way:
“You see a citation, you assume it’s a valid source… I’ll just see that there is a source. That’s it. I don’t verify it.”
What to do with this, in practice
- Don’t celebrate a citation without reading what it says. Before using an AI mention as social proof, confirm exactly what claim it’s making about your brand.
- Assume the user won’t verify. If people only perform 2 of the 12 checks they’d run on Google when using AI instead, what the AI says about your business is, in practice, the final word for a lot of people.
- Treat Perplexity with more caution than its reputation suggests. It’s the engine the market trusts most — and, according to this study, the one with the widest margin of error.
- Make the correct facts about your brand easy to find and hard to get wrong (structured data, consistent information across your site) — not because it guarantees the AI will never make a mistake, but because it reduces the odds that it does.
An honest limitation
This is a study with a relatively small human sample (24) focused on three specific engines — it doesn’t directly cover ChatGPT or Gemini. But the automated test (303 questions, 8 metrics) backs up the same conclusion through a different method, and the paper went through peer review at a serious conference. It’s solid evidence, not the final verdict on every AI engine.
Do you know what AI is saying about your brand right now?
At Tentacron IA we run that diagnosis — free, within 48 hours, no commitment. We check what ChatGPT, Perplexity, and Gemini are claiming about your business today, and flag anything that’s wrong, misunderstood, or simply missing.
Request your free diagnosis → tentacronia.com
Source: Pranav Narayanan Venkit, Philippe Laban, Yilun Zhou, Yixin Mao & Chien-Sheng Wu, “Search Engines in an AI Era: A Qualitative Understanding to the False Promise of Factual and Verifiable Source-Cited Responses in LLM-based Search”, ACM Conference on Fairness, Accountability, and Transparency (FAccT ‘25), 2025.