Why it happens
Assistants answer from two places: what they retrieve live, and what they absorbed during training. When a query has strong grounding, the model leans on the retrieved source. When grounding is thin or missing, it falls back on parametric priors — the statistical patterns of its training — and produces something that fits the shape of a real answer without a source behind it.
Nothing in the output marks the difference. A hallucinated price is formatted like a real one; an invented company name follows the same naming conventions as a real one. The model isn't lying in any deliberate sense — it's completing a pattern, and the completion happens to be untethered from fact.
The three shapes it takes
| Form | What it looks like |
|---|---|
| Invented option | A named business, product, or feature that doesn't exist but is offered as a real choice. |
| Wrong attribute | A real company pinned to a service, price, or fact it never had. |
| Plausible-but-false | A detail that fits the category and sounds verified, with no source supporting it. |
Why measurement catches it
A single query can't tell you whether an answer is grounded or invented — one confident reply looks the same either way. Reading the actual answers at scale is what surfaces the fabrication: run the query across a panel of engines, sample each engine several times, and the invented names and attributes stop hiding.
The AI Citation Institute reads the real answers from a frozen five-engine panel — ChatGPT, Claude, Gemini, Perplexity, and Google AI Overviews — sampling each query k times per engine. Hallucinated details rarely reproduce cleanly across samples and engines, so sampling exposes the variance that a single lucky query would mask. A wrong attribute pinned to a real business shows up when you compare the claim against what any real source supports.
The point isn't to score confidence — a fabrication is fully confident. It's to read what was actually said, at enough samples that the ungrounded claims separate from the grounded ones.
Common questions
Is a hallucinated recommendation the same as a lie?+
No. There's no intent behind it. The model completes a pattern from its training when live grounding is missing, and the completion happens to be untethered from fact. It reads as confident because the model has no internal signal that it's unsupported.
Can you tell a hallucination from a real answer just by reading one reply?+
Usually not. A fabricated price, name, or feature is formatted exactly like a grounded one. The only reliable tell is comparing the claim against a real source, and checking whether it reproduces across repeated samples and multiple engines.
Why does sampling the same query several times help?+
Grounded facts reproduce cleanly; ungrounded ones tend to vary or shift across samples and engines. Sampling each query k times per engine on a fixed panel makes that variance visible, so hallucinated names and attributes separate from the details a real source supports.
keep reading
See what AI says about your brand right now — free, on a live engine.
Scan your domain