AbstractTimescales are of fundamental importance to evolutionary biology as they facilitate hypothesis tests of historical evolutionary processes. Through the incorporation of fossil occurrence data, the fossilized birth–death (FBD) process provides a framework for estimating divergence times using more paleontological data than traditional node calibration approaches have allowed. The inclusion of more data can refine evolutionary timescale estimates, but for many taxonomic groups it is computationally infeasible to include all available fossil occurrence data. Here, we utilize both empirical data and a simulation framework to identify approaches to subsampling fossil occurrence data that result in the most accurate estimates of divergence times. To achieve this we assess the performance of the FBD-Skyline model when implementing multiple approaches to incorporating subsampled fossil occurrence data. Our results demonstrate that it is necessary to account for all available fossil occurrence data to achieve the most accurate estimates of clade age. We show that this can be achieved if an empirical Bayes approach, accounting for fossil sampling through time, is applied to the FBD process. Random subsampling of occurrence data can lead to estimates of clade age that are incompatible with fossil evidence if no control over the affinities of fossil occurrences is enforced. Our results call into question the accuracy of previous divergence time studies incorporating the FBD process that have used only a subsample of all available fossil occurrence data.