Geographic features in web search retrieval

We conduct large-scale search engine relevance experiments, using the 12% of queries that contain placenames, matching the placenames to places in the documents, and examining the impact of geographic features on web retrieval relevance. Specifically we examine distance between query and document place-names mentioned, noting that when a document has multiple places (which we observe in 82% of documents) we must choose a function over those multiple places. We find that the minimum distance between the document locations and query location is the strongest geographical predictor of document relevance, and that combining geographic features with text features gives us a 5% improvement in relevance over using text features alone.