Search mode

For a study, treat the mode as the interface state recorded with a saved answer. Describe what the user could select or see, and avoid inferring hidden routing from that label.

Keep four observations separate: a named entity, an exposed link, an attributed passage, and a recommendation. Their co-occurrence describes one answer, not a single combined ranking signal.

Log the selected mode alongside the query, surface, locale, region, timestamp, complete answer, exact fragment, entity, link, and event type. This makes a mode comparison auditable without claiming knowledge of the backend.

Pre-register the mode labels and apply them to a control sample before interpreting outcomes. Preserve ambiguous cases and reviewer disagreements; they show where the interface does not support a confident classification.

You can claim that a result was observed in a named mode at a given time. Rates and trends additionally require a fixed query set, denominator, surface, and mode version; they do not reveal the implementation behind the mode.

The label can hide experiments, personalization, regional routing, and changes to the retrieval stack. A mode comparison therefore describes the tested interface states only; it is not evidence that one internal pipeline consistently outperforms another.