Variation in sampling intensity of the geological record has long been suspected to distort our view of the history of life. When the taxonomic diversity of the same widespread group of marine nannoplankton (coccolithophorids) is estimated conventionally and separately from published land and deep-sea fossil records, the two curves are very different: they track sampling intensities from their respective rock records more closely than they do each other. However, when sampling intensity is corrected for, using two recently developed independent techniques, a common underlying signal emerges from the two records. This shows diversity rising steadily throughout the Mesozoic with a marked drop after the Cretaceous (64–58 Ma), and a long-term decline beginning in the Late Eocene and extending to the Early Miocene. We conclude that the observable fossil record is strongly shaped by sampling bias, and that this is a significant confounding factor for biodiversity analysis. Furthermore, based on currently available sources, we find that the raw deep-sea record is a worse estimator of diversity than the land-based record.