AI recommendation

Classify a recommendation from the answer’s directive or comparative language, preserving the task and alternatives that frame it. Do not infer endorsement from a bare name or from a link displayed nearby.

A recommendation is stronger than a mention because it assigns suitability for a task, but it is still distinct from a citation or a clickable link. Record the wording that makes the advice actionable.

Record the user task, candidate set, recommended object, rationale, alternatives, query, surface, locale, region, timestamp, and link state. These fields show whether the advice was conditional or presented as an unqualified choice.

Define recommendation cues before reviewing results, then test the rule against neutral mentions and sponsored-looking language. Preserve borderline cases rather than forcing every positive-sounding phrase into the recommendation category.

A counted recommendation means that the sampled answer advised an entity for a stated task. It does not establish that the entity is objectively best, that the advice was followed, or that a user converted.

Recommendations can reflect incomplete candidates, prompt framing, personalization, commercial relationships, or hidden policies. External observation cannot usually separate those causes, so recommendation counts are signals of observed advice—not independent quality ratings.