We describe here a method to identify potential binding sites in ensembles of protein structures as obtained by molecular dynamics simulations. This is a highly important task in the context of structure-based drug discovery, and many methods exist for the much simpler case of static structures. However, during molecular dynamics, the cavities and grooves that are used to define binding sites merge, split, appear, and disappear, and cover a large volume. Combined with the large number of sites (∼105 and more), these characteristics hamper a consistent and comprehensive definition of binding sites. Our method is based on the calculation of instantaneous cavities and of the pockets delineating them. Classification of the pockets over the structure ensemble generates consensus pockets, which define sites. Sites are reported as lists of atoms or residues. This avoids the pitfalls of the classification of cavities by spatial overlap, used in most existing methods, which is bound to fail on nonordered or unaligned ensembles or as soon as significant molecular motions are involved. To achieve a robust and consistent classification, we thoroughly optimized and benchmarked the method. For this, we assembled from the literature a set of reference sites on systems involving significant functional molecular motions. We tested different descriptors, metrics, and clustering methods. The resulting method is able to perform a global analysis of potential sites efficiently. Tests on examples show that our approach can make predictions of potential sites on the whole surface of a protein and identify novel sites absent from static structures.