The number of search queries that are associated with geographical locations, either explicitly or implicitly, has been quadrupled in recent years. For such geo-sensitive queries, the ability to accurately infer users' geographical preference greatly enhances their search experience. By mining past user clicks and constructing a geographical click probability distribution model, we address two important issues in spatial Web search: how do we determine whether a search query is geo-sensitive, and how do we detect, disambiguate, and visualize the associated geographical location(s). We present our empirical study on a large-scale dataset with about 9,000 unique queries randomly drawn from the logs of a popular commercial search engine Yahoo! Search, and about 430 million user clicks on 1.6M unique Web pages over an eight-month period. Our classification method achieved recall of 0.98 and precision of 0.75 in identifying geo-sensitive search queries. We also present our preliminary findings in using geographical click probability distributions to cluster search results for queries with geographical ambiguities.