Query Log Analysis

A consistent and influential picture of real search behavior. Jansen, Spink, and Saracevic's analysis of Excite queries found that web queries are very short (about two terms on average), that advanced operators and Boolean syntax are used rarely and often incorrectly, and that users typically view only the first page of results and issue few queries per session. Silverstein and colleagues' billion-record AltaVista study reported the same short-query, shallow-viewing pattern and documented Zipfian term-frequency distributions. Together these findings reshaped how search systems are designed, favoring support for short, unstructured queries.

By heuristics, since logs rarely mark sessions explicitly. The standard approach groups a user's consecutive activity into a session and starts a new one when the user is inactive longer than a timeout — commonly around 30 minutes — or when the identifying field (login, cookie, or IP) changes. This is imperfect: a single long task can be split if the user pauses, and unrelated tasks can be merged if they occur close together, while shared IPs can blend multiple users. Because session-level metrics depend on this segmentation, the timeout and identifier choices are reported and sometimes varied as a robustness check.

Only indirectly and unreliably. Logs record observable actions — queries, clicks, pages viewed, dwell signals — but not the user's goal or whether it was met, so satisfaction must be inferred from proxies like clicks, reformulation, or abandonment, all of which are ambiguous (a quick exit can mean success or frustration). This is the core limitation of the method: it answers what users did, not why or how well. To capture intent and satisfaction, query log analysis is paired with surveys, interviews, or controlled user studies that supply the missing why.