In current practice of probabilistic seismic‐hazard analysis (PSHA), the difference between ground‐motion prediction equations (GMPEs), which are in principle equally valid related to their quality and applicability, is attributed to epistemic uncertainty. The standard practice is to include this uncertainty through logic trees. Based on probability concepts, we present a method to assist during the selection and weighting of GMPEs to be included in different branches of a logic tree. We find that in regions with abundant recorded data, only those models with large likelihood should be considered. Although the presented method is not the only option to define the weighting of GMPEs for PSHA, it offers an ordered way to combine different sources of knowledge, such as recorded data and prior information.