Computer system designers often use caches to solve performance problems. Caching in the World Wide Web has been both the subject of extensive research and the basis of a large and growing industry. Traditional Web caches store HTTP responses, in anticipation of a subsequent reference to the URL of a cached response. Unfortunately, experience with real Web users shows that there are limits to the performance of this simple caching model, because many responses are useful only once. Researchers have proposed a variety of more complex ways in which HTTP caches can exploit locality in real reference streams. This article surveys several techniques, and reports the results of trace-based studies of a proposal based on automatic recognition of duplicated content.