The latitudinal biodiversity gradient (LBG), in which species richness decreases from tropical to polar regions, is a pervasive pattern of the modern biosphere. Although the distribution of fossil occurrences suggests this pattern has varied through deep time, the recognition of palaeobiogeographic patterns is hampered by geological and anthropogenic biases. In particular, spatial sampling heterogeneity has the capacity to impact upon the reconstruction of deep time LBGs. Here we use a simulation framework to test the detectability of three different types of LBG (flat, unimodal and bimodal) over the last 300 Myr. We show that heterogeneity in spatial sampling significantly impacts upon the detectability of genuine LBGs, with known biodiversity patterns regularly obscured after applying the spatial sampling window of fossil collections. Sampling-standardization aids the reconstruction of relative biodiversity gradients, but cannot account for artefactual absences introduced by geological and anthropogenic biases. Therefore, we argue that some previous studies might have failed to recover the ‘true’ LBG type owing to incomplete and heterogeneous sampling, particularly between 200 and 20 Ma. Furthermore, these issues also have the potential to bias global estimates of past biodiversity, as well as inhibit the recognition of extinction and radiation events.