Islands have been used as model systems to study ecological and evolutionary processes, and they provide an ideal set‐up for validating new biodiversity monitoring methods. The application of environmental DNA metabarcoding for monitoring marine biodiversity requires an understanding of the spatial scale of the eDNA signal, which is best tested in island systems. Here, we investigated the variation in Actinopterygii and Elasmobranchii species composition recovered from eDNA metabarcoding along a gradient of distance‐to‐reef in four of the five French Scattered Islands in the Western Indian Ocean. We collected surface water samples at an increasing distance from reefs (0 m, 250 m, 500 m, 750 m). We used a metabarcoding protocol based on the ‘teleo’ primers to target marine reef fishes and classified taxa according to their habitat types (benthic or pelagic). We investigated the effect of distance‐to‐reef on β diversity variation using generalised linear mixed models and estimated species‐specific distance‐to‐reef effects using a model‐based approach for community data. Environmental DNA metabarcoding analyses recovered distinct fish species compositions across the four inventoried islands and variations along the distance‐to‐reef gradient. The analysis of β‐diversity variation showed significant taxa turnover between the eDNA samples on and away from the reefs. In agreement with a spatially localised signal from eDNA, benthic species were distributed closer to the reef than pelagic ones. Our findings demonstrate that the combination of eDNA inventories and spatial modelling can provide insights into species habitat preferences related to distance‐to‐reef gradients at a small scale. As such, eDNA can not only recover large compositional differences among islands but also help understand habitat selection and distribution of marine species at a finer spatial scale.