Described is a technology by which topics corresponding to web pages are used in relevance ranking of those pages. Topics are extracted from each web page of a set of web pages that are found via a query. For example, text such as nouns may be extracted from the title, anchor texts and URL of a page, and used as the topics. The extracted topics from a page are used to compute a relevance score for that page based on an evaluation of that page's topics against the query. The pages are then ranked relative to one another based at least in part on the relevance score computed for each page, such as by determining a matching level for each page, ranking pages by each level, and ranking pages within each level. Also described is training a model to perform the relevance scoring and/or ranking.