Search Session Analysis

Because real information needs are usually satisfied across several queries, not one. Users issue an initial query, scan, reformulate, click, and try again, so judging only a single query misses where the value actually comes from and can mislabel a perfectly normal opening query as a failure. Session analysis takes the episode as the unit, capturing reformulation, effort, and eventual success or abandonment. This matches how people search and lets systems be evaluated for guiding users to their goal over the whole interaction, which single-query measures like precision or reciprocal rank cannot express.

Through heuristics applied to logged activity, since sessions are rarely marked explicitly. The standard method combines an inactivity timeout — a gap beyond roughly 30 minutes ends a session — with topic-shift detection, because one sitting can hold several distinct needs. Both choices are imperfect: long tasks with pauses can be split, and back-to-back unrelated tasks can be merged. Because session-level metrics depend on this segmentation, the timeout and topic-change criteria are reported, and analyses often test sensitivity to them or use richer signals (query similarity, click patterns) to refine the boundaries.

By combining outcome signals with session-level effectiveness measures. Success can be indicated by a satisfying click, a completed task or conversion, or an explicit relevance judgment, while effectiveness across the episode is captured by measures that accumulate and discount the gain a user receives over successive queries, such as session-based discounted cumulative gain. The challenge is that behavioral signals are ambiguous — a short session may mean instant success or rapid frustration — so robust session evaluation triangulates log-based success proxies with task outcomes or user feedback rather than relying on any single signal.