It is becoming increasingly apparent that the species abundance distribution of many ecological communities contains multiple modes, a phenomenon that has been largely overlooked. Here, we test for multiple modes in the species abundance distribution using a combination of one, two and three mode Poisson lognormal distributions and an extensive arthropod dataset from the Azores. We consider the abundance distribution of twelve native laurisilva forest fragments and the combination of fragments within five islands, allowing us to detect whether patterns are consistent across scales. An information theoretic approach is employed to determine the best model in each case. To explore the processes driving multimodal abundance distributions we tested various potential mechanisms. We classified species as core if they are present in over half the fragments in our study, and as satellite if they are sampled in fewer than half the fragments. Furthermore, species are classified based on body size, whether they are indigenous (i.e. endemic or native non‐endemic) or introduced to the Azores, abundance in land uses other than native forest, and dispersal ability. We find that models incorporating multiple modes perform best for most fragments and islands. A large number of communities are bimodal, comprising a mode of very rare species and a mode of relatively common species. Deconstructing the full assemblages into their constituent subsets reveals that the combination of ecologically different groups of species into a single sample underpins the multimodal pattern. Specifically, the rarer mode prevailingly contains a higher proportion of satellite taxa, introduced species and species that are more adapted to anthropogenic land uses that surround the native forest.