This paper attempts to assess the impact of improved estimates of areal potential evapotranspiration (PE) on the results of two rainfall-runoff models. A network of 42 PE stations was used for a sample of 62 watersheds and two watershed models of different complexity (the four-parameter GR4J model and an eight-parameter modified version of TOPMODEL), to test how sensitive rainfall-runoff models were to watershed PE estimated with the Penman equation. First, Penman PE estimates were regionalized in the Massif Central highlands of France, a mountainous area where PE is known to vary greatly with elevation, latitude, and longitude. The two watershed models were then used to assess changes in model efficiency with the improved PE input. Finally, the behavior of one of the model's parameters was analyzed, to understand how watershed models cope with systematic errors in the estimated PE input. In terms of model efficiency, in both models it was found that very simple assumptions on watershed PE input (the same average input for all watersheds) yield the same results as more accurate input obtained from regionalization. The detailed evaluation of the GR4J model calibrated with different PE input scenarios showed that the model is clearly sensitive to PE input, but that it uses its two production parameters to adapt to the various PE scenarios.