Choosing the best-performing server for a particular client from a group of replicated proxies is a difficult task. We offer a novel, two-step technique for server selection that chooses a small subset of five servers, and isolates testing to that subset for ten days. We present an empirical evaluation of both our method and previously proposed metrics based on traces to 193 commercial proxies. We show that our technique performs better than any of the other metrics we studied -often one to two seconds better for a one-megabyte file-while requiring considerably less work over time. Metrics such as round-trip time and tests using small files usually select servers that are two to three times worse than the best server. Network-layer metrics such as minimizing router and autonomous system count poorly predict which server provides the best performance. These metrics often select servers with transfer times four to six times that of the best-performing server.