In order to study further the genetic structure of the pearl oyster Pinctada margaritifera in French Polynesia with a special consideration for the sampling scale, we analyzed or re-analyzed sets of data based on nuclear DNA markers obtained at different spatial scales. At a large scale (several 1,000 km), the remote Marquesas Islands were confirmed to be significantly differentiated from Tuamotu-Gambier and Society archipelagos, with a marked difference however for the two main islands that are different from each other. At a medium scale (several 10 to several 100 km), overall homogeneity was observed within and between these two archipelagos, with some exceptions. This could be attributed both to large-scale larval dispersal and to human-driven spat translocations due to pearl oyster cultivation. These results contrast with those observed (1) at a small scale (less than 10 km) in a lagoon heavily impacted by translocation and cultural practices, where significant genetic differentiation was detected among three laying beds, and (2) at a micro scale where we detected an important variability of the genetic composition of young spat recruited on artificial collectors. Such patterns could result from a high variance in the number of genitors at the origin of each cohort, or from pre- or post-settlement selection on linked loci. Altogether, our data support the hypothesis that under certain conditions populations of bivalves may exhibit patterns of chaotic genetic patchiness at local scale, in line with the increasing report of such patchiness in marine benthic organisms. This underlines the importance of sampling scale that should be rigorously defined depending on the questions to be answered. Nevertheless, a survey of about 80 articles dealing with population genetics of marine invertebrates showed that only 35% of those studies disclosed details about the sampling strategy (particularly the area explored). These results emphasize the need for cautious interpretation of patterns of genetic structure at medium scale when rigorous sampling strategies are not deployed.