While search engines have demonstrated improvements in both speed and accuracy, their response time is prohibitively long for applications that require immediate and accurate responses to search queries. Examples include identification of multimedia resources related to the subject matter of a particular class, as it is in session. This paper begins with a survey of collaborative recommendation and prediction algorithms, each of which applies a different method to predict future search engine usage based on the past history of a search engine user. To address the shortcomings identified in existing techniques, we propose a proactive search approach that identifies resources likely to be of interest to the user without requiring a query. The approach is contingent on accurate determination of similarity, which we achieve with local alignment and output-based refinement of similarity neighborhoods. We demonstrate our proposed approach with trials on real-world search engine data. The results support our hypothesis that a majority of users exhibit search engine usage behavior that is predictable, allowing a proactive search engine to bypass the common query-response model and immediately deliver a list of resources of interest to the user.