Implicit feedback is an evaluative signal inferred from what users do rather than from a rating they submit. It can supply training data for search and help reveal whether a deployed AI assistant is useful. Each action is a clue, not a definitive judgment. section 4.2.6 section 10.4.2
Useful signals depend on the task:
- A click after a search creates a query–listing pair that suggests relevance and can train an embedding model. section 4.2.6
- Session abandonment and immediate query refinement can indicate that an AI response did not help. section 10.4.2
- Copy-paste behavior can indicate a response worth saving or sharing; return use and session frequency can indicate continuing trust. section 10.4.2
A search click is not a relevance label
A user searches and clicks a listing. That interaction supplies a fresh training pair without asking the user to rate the result. But listings near the top are more likely to receive clicks regardless of relevance. Training directly on those clicks risks teaching the embedding model position bias rather than semantic relevance, so the click data needs debiasing. section 4.2.6
Interpret response behavior with similar care. An immediate sequence of follow-up questions may signal confusion, while an abandoned session may signal frustration. These patterns help monitor usefulness even when users provide no explicit rating. section 10.4.2
