In this research, we use a new approach to address an old problem of hydrologic modeling: what is the appropriate level of spatial disaggregation to model the rainfall-runoff relationship at the watershed scale? We show that, for rainfall-runoff models, the overwhelming part of the improvement brought by spatial distribution is linked to the possibility to account of rainfall variability. In practical terms, this means that efforts to improve the reliability of hydrological models through increasing spatial resolution should be directed primarily towards better spatial distribution of rainfall data and not so much towards the spatialization of watersheds (land-surface) parameters. To reach our objective, we consider the most basic sort of semi-distribution--a watershed that can be split into two sub-watersheds. We introduce the concept of "chimera watershed", where two actual watersheds of similar size are associated, even if they do not belong to the same hydrological region. This results in a large number of quite contrasted hydrological situations, which certainly exist in reality, but are seldom observed by existing stream gauge networks. We based our study on more than 300 French watersheds, which we used to produce 4,500 chimeras of all sizes. On these chimera watersheds, we systematically compared the efficiency of a rainfall-runoff model, which was either fully or partially semi-distributed, or lumped. We believe that our conclusions can be useful to all those who work on improving rainfall-runoff models for operational purposes. For streamflow forecasting, for example, it is important to know how to discretize space, as more subwatersheds mean a need for more gauging stations. We also think that the concept of chimera watersheds may have further hydrological applications, either to study theoretical problems (spatial disaggregation, parameter combination) or applied ones (such as improvement of lumped engineering models to improve their account of spatial variability).