Watershed models have been widely used for creating the scientific basis for management decisions regarding nonpoint source pollution. In this study, we evaluated the current state of watershed scale, spatially distributed, process-based, water quality modeling of nutrient pollution. Beginning from 1992, the year when Beven and Binley published their seminal paper on uncertainty analysis in hydrological modeling, and ending in 2010, we selected 257 scientific publications which (i) employed spatially distributed modeling approaches at a watershed scale; (ii) provided predictions of flow, nutrient/sediment concentrations or loads; and (iii) reported fit to measured data. Most "best practices" (optimization, validation, sensitivity, and uncertainty analysis) are not consistently employed during model development. There are no statistically significant differences in model performance among land uses. Studies which used more than one point in space to evaluate their distributed models had significantly lower median values of the Nash-Sutcliffe Efficiency (0.70 vs 0.56, p<0.005, nonparametric Mann-Whitney test), and r2 (p<0.005). This finding suggests that model calibration only to the basin outlet may mask compensation of positive and negative errors of source and transportation processes. We conclude by advocating a number of new directions for distributed watershed modeling, including in-depth uncertainty analysis and the use of additional information, not necessarily related to model end points, to constrain parameter estimation.