The use of plant functional types (PFTs) to describe patterns and processes in plant communities has become essential to study and predict consequences of global change on vegetation and ecosystem processes. A PFT is a group of plants that, irrespective of phylogeny, are similar in a given set of traits and similar in their association to certain variables, which may be factors to which the plants are responding or effects of the plants in the ecosystem. To define PFTs relevant traits must be selected and an appropriate method must be used to classify plants into types. We critically review methods used for the analysis of PFT-based data and describe a new recursive algorithm to numerically search for traits and find optimal PFTs. The algorithm uses three data matrices: describing populations by traits, communities by these populations and community sites by environmental factors or effects. It defines PFTs polythetically by cluster analysis, revealing plant types whose performance in communities is maximally associated to the specified environmental variables. We test the method with data from natural grassland communities of southern Brazil, which were experimentally subjected to combinations of grazing levels and N-fertilizer. The new method is found to be better than similar analytical procedures previously described. Redundancy among traits is discussed and a procedure for comparing alternative solutions is presented based on the similarity in terms of PFT responses between different trait subsets. The concept of PFT response group is illustrated by example.