Species distribution models (SDM) have become one of the most popular predictive tools in ecology. With the advent of new computation and remote sensing technology, high-resolution environmental data sets are becoming more and more common predictors in these modelling efforts. Understanding how scaling affects their outputs is therefore fundamental to understand their applicability. Here, we develop a theoretical basis to understand the consequences of aggregating occurrence and environmental data at different resolutions. We provide a theoretical framework, along with numerical simulations and a real-world case study, to show how these scaling rules influence predictive outputs. We show that the properties of the environment–occurrence relationships change when the data are aggregated: the mean probability of occurrence and species prevalence increases, the optimal environmental values shift and classification rates increase at coarser resolutions up to a certain level. Furthermore, and contrary to the widespread expectation that high-resolution data would produce better predictions, we show here that model performance may increase using coarser resolution data sets rather than the inverse. Finally, we also show that model performance depends not only on the environment–occurrence relationship but also on the interaction between this and the geography and distribution of the available environment. This theoretical framework helps understanding previously incoherent results regarding SDM upscaling and model performance, and illustrates how theoretical and empirical results can provide important feedbacks to advance in understanding scaling issues in macroecology. The interaction between the shape of the environment–occurrence relationship and the rates of change of the environment is fundamental to understand the effects of upscaling in model performance, and may explain why some models are more difficult to transfer to different regions. Most importantly, we argue