Temporal diversity patterns have traditionally been analyzed by counting the number of families or genera present over a series of time periods. This approach has been criticized on the grounds that paraphyletic taxa might introduce artifacts. Sepkoski and Kendrick (1993) simulated phylogenetic trees and different classifications of those trees and concluded that paraphyletic taxa need not be rejected. We have reimplemented their model, extended it, and carried out statistical analyses under a variety of experimental conditions. Our results show that the focus on monophyly vs. paraphyly is misplaced. Instead, it appears that the number of groups in the classification and the distribution of the sizes of those groups have dramatic effects on the recovery of diversity information. Furthermore, the influence of these factors depends on whether the fossil record represents a low- or high-frequency sampling of lineages. When sampling is good, the best results are achieved by classifications with large numbers of small taxa. When sampling is poor, however, the best results are achieved by classifications that include some large and medium-sized groups as well as many smaller groups. This suggests that the best estimates of underlying diversity will be achieved by counting (in the same study) taxa assigned to different ranks, so as to best match the inferred quality of the paleontological sample. In practice this will mean abandoning the commitment to counting taxa at a single rank.