A method is presented for pruning a peer-to-peer network such that a peer-to-peer search within the peer-to-peer network is initiated beginning with those nodes on which relevant content has previously been found for particular search keywords. For each completed search, a peer node gathers client rating results and then forwards the rating information to a rating server; these results reflect the degree to which searches successfully located content for keywords in a search using particular peer nodes. A rating server maintains a database of results received from nodes. Prior to initiating a new search at a peer node, the peer node consults one or more rating databases to retrieve a set of initial nodes to which the node should initiate the new search in order to maximize the speed and success of finding relevant content. The node then performs the new search using the set of initial nodes.