Taphonomic deformation, the distortion of fossils as a result of geological processes, poses problems for the use of geometric morphometrics in addressing paleobiological questions. Signal from biological variation, such as ontogenetic trends and sexual dimorphism, may be lost if variation from deformation is too high. Here, we investigate the effects of taphonomic deformation on geometric morphometric analyses of the abundant, well known Permian therapsid Diictodon feliceps. Distorted Diictodon crania can be categorized into seven typical styles of deformation: lateral compression, dorsoventral compression, anteroposterior compression, ‘saddle-shape’ deformation (localized collapse at cranial mid-length), anterodorsal shear, anteroventral shear, and right/left shear. In simulated morphometric datasets incorporating known ‘biological’ signals and subjected to uniform shear, deformation was typically the main source of variance but accurate ‘biological’ information could be recovered in most cases. However, in empirical datasets, not only was deformation the dominant source of variance, but little structure associated with allometry and sexual dimorphism was apparent, suggesting that the more varied deformation styles suffered by actual fossils overprint biological variation. In a principal component analysis of all anomodont therapsids, deformed Diictodon specimens exhibit significant dispersion around the ‘true’ position of this taxon in morphospace based on undistorted specimens. The overall variance associated with deformation for Anomodontia as a whole is minor, and the major axes of variation in the study sample show a strong phylogenetic signal instead. Although extremely problematic for studying variation in fossil taxa at lower taxonomic levels, the cumulative effects of deformation in this study are shown to be random, and inclusion of deformed specimens in higher-level analyses of morphological disparity are warranted. Mean morphologies of distorted specimens are found to approximate the morphology of undistorted specimens, so we recommend use of species-level means in higher-level analyses when possible.