Hallucination

What it means in practice

For brand monitoring the important cases are invented products, wrong pricing, and attributed quotes that were never said. These are visible only by reading the answer text, which is why storing the raw answer behind every score matters: an aggregate that says "mentioned" cannot distinguish an accurate recommendation from a fabricated one.

How refd treats it

This is category vocabulary, not a refd metric. refd measures what AI answers actually said across repeated runs, and publishes the raw answer behind every number. Where a term above carries a claim about what causes visibility, refd treats that claim as a hypothesis to test rather than a setting to configure.

This definition is editorial. It describes how the term is used across the field rather than something the software computes, so unlike the metric pages it is not generated from the product.