Among other objectives, forest inventories are aimed to identify ecological communities and to correlate community composition with environmental variables. The identification of different communities would require several forest inventories, each covering small sampling areas with relatively homogeneous environmental conditions. The multiple plot sampling method, traditionally used in local inventories of tropical forests, cannot assure such homogeneity, since even small sampling areas would show environmental heterogeneity influencing vegetation. In this paper we assessed the consequences of this heterogeneity for sampling by quantifying the variability of species abundance ranks for species sampled with 10 or more individuals in a set of plots covering a small sampling area. The species reference abundance ranks were obtained from a sample of 100 plots of 10 m × 10 m each randomly set in a sampling area of 6.5 ha in a tropical forest fragment (Southeastern Brazil). For each species we used resamplings (30 trials) to obtain the species abundance ranks in sub-samples, considering different sampling intensities (n = 25, 50 and 75 plots), and compared these ranks with the species reference rank (n = 100 plots). Then, we compared the species ranks in sub-samples of 50 plots (10.000 trials) with the reference rank and assessed the frequency and extent of rank displacements. Species rank was highly variable across resampling trials for the sampling intensities of n = 25 and n = 50, but decreased considerable with a sampling intensity of n = 75 plots. The mean rank variability and especially the maximum displacement raised significantly from the seventh most abundant species on, and some species occupied quite discrepant abundance ranks in up to 10% of the 10.000 resampling trials. This high internal variability of forest samples may impair the search for floristic patterns as scale lessens, say, to the meso-scale (1–100 km2). We discussed some possible ways to increase internal homogeneity of tropical forest samples with the multiple plot sampling method. Among these, objective entitation, based on an ancient phytosociological procedure, is suggested as the most appropriate for use on the hilly relieves of the Atlantic forest biome.