Environmental filtering and dispersal history shape global biodiversity, yet how their relative importance shifts across macroevolutionary timescales remains largely unknown in marine ecosystems. Analyzing phylogenetic and taxonomic beta diversity turnover offers a powerful framework to evaluate the historical biogeographical processes structuring contemporary assemblages. Here, we examined the relative impacts of environmental dissimilarity and geographical distance on marine fish turnover for 5,659 ray-finned fish species globally, reconstructing paleogeographical barriers, linear distances, and least-cost pathways across evolutionary timescales. Environmental conditions and geography jointly explained a substantial proportion of the deviance in taxonomic (60.43%) and phylogenetic (49.61%) turnover. Tracking these dynamics forward from the Late Cretaceous (100 Ma) to the present day (0 Ma) revealed a pronounced evolutionary shift in community assembly rules. At 100 Ma, ancestral lineage distribution was overwhelmingly governed by environmental filtering, while the unique effect of dispersal history was negligible (4.97%). Over the ensuing 100 million years, the signature of dispersal limitation steadily accumulated across the fish tree of life, rising sharply to dominate shallow-time phylogenetic turnover at the present day (33.15%). While the capacity of the environment to filter lineages remained a highly consistent and persistent constraint across the entire macroevolutionary timeline, the specific drivers evolved. Deep-time filtering was primarily dictated by non-thermal constraints such as salinity and dissolved oxygen, whereas contemporary turnover has shifted to joint regulation by modern sea surface temperature and geographic least-cost barriers. Our findings demonstrate that over deep evolutionary time, marine fish lineages eventually disperse widely across ocean basins to match their physiological niches. Consequently, the contemporary dominance of dispersal history represents a recent accumulation of spatial constraints that overlays a persistent, deep-time foundation of environmental filtering.