An AI engine can hand you an accurate fact wrapped in a citation that never said it — or never existed. The Fabricated-Source Rate measures how often the evidence an engine points to fails to support the claim it is attached to.
A citation to a source that does not exist — the classic hallucination, dressed as a reference.
A real source that says nothing of the kind, attached to lend a claim borrowed authority.
An accurate fact with fabricated provenance — the answer reads as sourced while the chain is fiction.
The dangerous case is not the obvious hallucination — it is the answer that is correct and confidently mis-attributed. An engine surfaces a real fact, sometimes one it absorbed from a verified record, then assembles a citation that did not produce it. The reader sees a sourced, trustworthy answer; the evidence chain underneath is invented.
Because Entidex tracks every claim against a verified record, it catches a laundered citation even when the fact itself checks out — the source resolves to nothing, or to something that never made the claim. That is the gap a plain accuracy score cannot see, and the one that carries the most exposure: a German court has already treated AI-Overview-style fabricated attribution as the operator's own liability.
Read the citation-laundering case study · see the methodology.
The unified Truth Gap surface — per-engine accuracy, claim ladder, and citation provenance for a live entity:
It is the share of an AI engine’s citations that do not support the claim they are attached to. That includes sources that were invented outright, real sources that say nothing of the kind, and real facts dressed in a citation the engine assembled after the fact. It is a measure of broken provenance, not of whether the underlying fact happens to be right.
This is the failure mode that matters most. An engine can surface an accurate claim — sometimes one it learned from a verified record — and then attach a citation that did not produce it, or one that does not exist. The answer reads as sourced and trustworthy while the chain of evidence is fiction. We call this citation laundering: accurate facts with fabricated provenance.
Entidex resolves each cited source and checks whether it actually supports the claim it is attached to, across ChatGPT, Claude, Gemini, Perplexity and Grok. A citation that resolves and supports the claim counts as grounded; one that cannot be found, or that does not contain the claim, counts against the rate. Because every claim is tracked against a verified record, a laundered citation — right fact, wrong or invented source — is caught even when the fact itself checks out.
Because provenance is what makes an answer auditable, and because the next claim from the same broken chain may not be right. A German court has already described AI-Overview-style fabricated attribution as the operator’s own liability. For any brand, a confidently mis-sourced statement is a reputational and legal exposure regardless of whether that particular fact landed.
Run a free scan. Entidex shows which claims the AI engines make about your brand, which carry real supporting sources, and which are riding on citations that do not hold up.