Hallucination

Hallucination is the failure mode of generative LLMs: producing content that reads as confident and well-formed but is wrong. The phenomenon is intrinsic to how LLMs work: they generate the statistically likely next token given context, which sometimes lands on a plausible-sounding statement that has no basis in fact.

In answer engines, hallucination is partly mitigated by RAG: the engine grounds against retrieved sources, citing them, and is supposed to stay within what the sources support. But hallucination still happens in three failure modes:

1. Misattribution. The engine cites your page for a claim your page doesn't actually make. The reader trusts your domain; the claim is wrong; the reputational cost lands on you.
2. Out-of-context synthesis. Multiple sources are combined in ways that produce a claim none of them individually made. Citations appear, but the synthesis is fabricated.
3. Citation-free fabrication. Engines that allow generation without grounding (or where grounding is weak) sometimes produce uncited statements that are entirely invented.

For publishers, hallucination management has two sides:

• Reduce being misattributed. Write with specific, citable claims; date your content; use clear definitional language; structure for chunk-level retrieval. Pages with quotable, attributed sentences are misquoted less often than pages of hedged generality.
• Monitor for misattribution. Periodically query the major engines on topics you cover; check whether citations to your domain are accurate. The measurement guide covers attribution-quality auditing.

The publisher's leverage on hallucination is limited (you can't fix the underlying model), but you can be the source that's hard to misquote and easy to verify against.