Deep coral reefs (that is, mesophotic coral ecosystems) can act as refuges against major disturbances affecting shallow reefs. It has been proposed that, through the provision of coral propagules, such deep refuges may aid in shallow reef recovery; however, this “reseeding” hypothesis remains largely untested. We conducted a genome-wide assessment of two scleractinian coral species with contrasting reproductive modes, to assess the potential for connectivity between mesophotic (40 m) and shallow (12 m) depths on an isolated reef system in the Western Atlantic (Bermuda). To overcome the pervasive issue of endosymbiont contamination associated with de novo sequencing of corals, we used a novel subtraction reference approach. We have demonstrated that strong depth-associated selection has led to genome-wide divergence in the brooding species Agaricia fragilis (with divergence by depth exceeding divergence by location). Despite introgression from shallow into deep populations, a lack of first-generation migrants indicates that effective connectivity over ecological time scales is extremely limited for this species and thus precludes reseeding of shallow reefs from deep refuges. In contrast, no genetic structuring between depths (or locations) was observed for the broadcasting species Stephanocoenia intersepta, indicating substantial potential for vertical connectivity. Our findings demonstrate that vertical connectivity within the same reef system can differ greatly between species and that the reseeding potential of deep reefs in Bermuda may apply to only a small number of scleractinian species. Overall, we argue that the “deep reef refuge hypothesis” holds for individual coral species during episodic disturbances but should not be assumed as a broader ecosystem-wide phenomenon.