Under neutral theory, the level of polymorphism in an equilibrium population is expected to increase with population size. However, observed levels of diversity across metazoans vary only two orders of magnitude, while census population sizes (Nc) are expected to vary over several. This unexpectedly narrow range of diversity is a longstanding enigma in evolutionary genetics known as Lewontin’s Paradox of Variation (1974). Since Lewontin’s observation, it has been argued that selection constrains diversity across species, yet tests of this hypothesis seem to fall short of explaining the orders-of-magnitude reduction in diversity observed in nature. In this work, I revisit Lewontin’s Paradox and assess whether current models of linked selection are likely to constrain diversity to this extent. To quantify the discrepancy between pairwise diversity and census population sizes across species, I combine genetic data from 172 metazoan taxa with estimates of census sizes from geographic occurrence data and population densities estimated from body mass. Next, I fit the relationship between previously-published estimates of genomic diversity and these approximate census sizes to quantify Lewontin’s Paradox. While previous across-taxa population genetic studies have avoided accounting for phylogenetic non-independence, I use phylogenetic comparative methods to investigate the diversity census size relationship, estimate phylogenetic signal, and explore how diversity changes along the phylogeny. I consider whether the reduction in diversity predicted by models of recurrent hitchhiking and background selection could explain the observed pattern of diversity across species. Since the impact of linked selection is mediated by recombination map length, I also investigate how map lengths vary with census sizes. I find species with large census sizes have shorter map lengths, leading these species to experience greater reductions in diversity due to linked selection. Even after using high estimates of the strength of sweeps and background selection, I find linked selection likely cannot explain the shortfall between predicted and observed diversity levels across metazoan species. Furthermore, the predicted diversity under linked selection does not fit the observed diversity–census-size relationship, implying that processes other than background selection and recurrent hitchhiking must be limiting diversity.