The environmental DNA (eDNA) approach is an emerging tool for monitoring marine biodiversity. However, the sampling effort needs optimization according to the site characteristics and target taxonomic groups. In this study, we optimized the eDNA sampling effort in terms of sample volume and number of replicates to monitor the diversity of marine vertebrates (mainly fish) in Hong Kong's subtropical waters that show a gradient of estuarine to oceanic waters. To maximize detection, we used three pairs of metabarcoding primers (12S‐v5, MiFish‐U, and MiFish‐E). We compared vertebrate diversity in 78 water samples, ranging from 1 to 10 L, collected from oceanic and estuarine sites. Metabarcoding yielded a total of 140 vertebrate species, of which 18 were unique to the estuarine site, 66 unique to the oceanic site, and 56 shared between both sites. The detected species were predominantly ray‐finned fish (136 species), and the three primer pairs exhibited differential sensitivity toward different taxa, especially cartilaginous fish and cetaceans. Increasing sampling volume per replicate generally increased the total detected species, average species per replicate, and species coverage, and sampling 3 or 4 × 4 L represented the most efficient sampling effort for the estuarine and oceanic sites, respectively. The diversity analysis revealed that sampling >2 L per replicate reduced variability and improved diversity analysis. The results also showed that a larger sampling volume per replicate increased the probability of detecting endangered, indicator, invasive, and elusive species, with 4 L representing the most efficient volume. This study recommended sampling 4 L per replicate and 3 replicates for estuarine and 4 for oceanic sites, respectively for effectively monitoring marine fish in subtropical waters using the eDNA approach.