In this paper we provide a full-scale evaluation of a cluster-based architecture for P2P IR, focusing on retrieval effectiveness. We observe that there is a significant difference in performance between the architecture we examine and a centralised index. After inspecting our experimental methodology and our results, we provide evidence that suggests that this discrepancy is due to the information clustering algorithms employed throughout. The construction errors of the resource descriptions as well as the failure of the clustering mechanisms to discover the structure of the smallest of peer-collections lead to erroneous query routing. We proceed further to show experimentally how content replication and relevance-feedback mechanisms can help to alleviate the problem.