Three experiments are described which examine the effects of similarity on face recognition using a new application of hierarchical clustering analysis (HCS). Experiments I and II employed a within-groups design where subjects attempted to recognize targets among decoys from the same and different cluster to the target. Common cluster membership accounted for 72% and 84%, respectively, of all false alarms in the two experiments. Absolute error rates were affected by the number of targets actually present in the array and the style of the instructions given to subjects but were not influenced by whether successive or simultaneous test presentation was employed. Experiment III used a between-subject design where targets were embedded in arrays composed from same or different clusters. False alarm rates were significantly higher for the same cluster condition, but hit rates were unaffected by recognition context. The significance of these findings for theories of face identification is discussed.