End-member modelling algorithms extract valuable information from grain-size distribution datasets, such as provenance, transport and sedimentation processes, which can subsequently be used to reconstruct past climate variability. Twenty years ago, Weltje (1997) published the algorithm EMMA, first applied to unmixing of grain-size distribution datasets in Prins and Weltje (1999a). In recent years, a range of new non-parametric algorithms developed specifically for unmixing of these datasets has become available. We evaluate whether the new algorithms as well as the original EMMA produce genetically meaningful results. We use artificial datasets and a published dataset from a geological case study. Experiments with the artificial datasets demonstrate that some algorithms are more appropriate for the unmixing of grain-size distribution data than others. Furthermore, the “geologically feasible” number of end members determined by the algorithms depends on the criteria for a statistical goodness of fit and does not necessarily correspond to the true number of end members in a dataset. Our study indicates that when an incorrect number of end members are modelled the shape and modal position of the end member grain-size distributions may deviate. Thus, end-member modelling should be used cautiously and in tandem with geological background information.