Plot sizes recommended for phytosociological sampling vary by several orders of magnitude. Even within a vegetation type, differences between plot sizes used by phytosociologists are often considerable. Still, the effect of plot size on the classification of vegetation remains obscure. We tested the effect of plot size and various transformations of plant cover values on results of cluster analyses in species-rich grasslands in southern Moravia (Czech Republic). Data from 21 plots, each consisting of five nested subplots ranging from 1 to 49 m2 in size (105 plots in total), were classified using the Ward's method of clustering, separately for each plot size. Significance of the resulting clusters was evaluated by bootstrap resampling developed by Pillar (1999). The clusters were compared across plot sizes using species fidelity. Five significant clusters were found for plots 4 to 49 m2 in size, with the same assignment of relevés to individual clusters among the classifications. By contrast, only two significant clusters were obtained at the smallest scale (1 m2). The number of species with high fidelity values increased with increasing plot size. The number of significant clusters and assignment of individual relevés was more substantially affected by different transformations of plant cover values. If the original ordinal data were transformed to percentages or to presence-absence data, the number of significant clusters decreased to two, independently of plot size. Our results suggest that plot sizes ordinarily used by phytosociologists in grasslands affect the shape of the classification trees only little, unless extremely small plot sizes are combined with larger ones. In contrast, various transformations of plant cover (and presumably also biased estimates of plant cover) may affect results of the classifications considerably.