Often, manually and numerically derived phytosociological classifications yield different results. Hitherto, a twostep procedure has been suggested in which numerical analysis of the data is followed by the revision of the resulting table (c.f. van der Maarel 1982). In this paper a new methodology is presented which makes manual refinements superfluous. Objectives are derived from phytosociological paradigms and conclusions drawn for the analytical process. The problems to be solved are: data transformation, detection of outliers, selection of clustering methods, checking within-group diversity, analysis of the resulting group structure, rearrangement of relevés and species within the groups, and finally the selection of differential species. The method has been derived using the well known example of Ellenberg (Mueller-Dombois & Ellenberg 1974). The results almost perfectly reproduce the intuitively widely accepted manual refinements in structure and presentation. A test with plant sociological data from Swiss forests (Ellenberg & Klötzli 1972) proves that the method can also classify complex gradient- and group systems and that the numerical result matches Landolt's (1977) system of indicator values. Since the solutions can be exactly reproduced, it is no longer necessary to combine numerical analysis with additional editing.