Accurate and precise estimates of relationships between stressors and environmental responses can inform management decisions most usefully when models can be easily interpreted. Here, we describe an approach for classifying lakes and reservoirs that can improve estimates of the relationships between total P (TP) and chlorophyll a (chl a) concentration, while preserving a model that can be readily interpreted by environmental managers and stakeholders. We selected classification variables statistically with a classification and regression tree in which relationships between TP and chl a were the terminal nodes of the tree. We developed a set of classification trees from bootstrapped replicates of the calibration data to explore a broader range of possible trees. We chose a final tree based on its predictive performance with a validation data set. The total N:TP mass ratio was the classification variable selected most frequently from a broad array of biological, chemical, and physical candidate classification variables. Relationships between TP and chl a in the resulting lake classes provided predictions that were substantially more accurate than predictions computed using nutrient ecoregions based on aggregations of Omernik Level III ecoregions, but predictions from a random forest model that averaged an ensemble of trees were even more accurate. Thus, the classification approach presented here sacrifices a small amount of predictive accuracy to retain a tree structure that is readily interpretable.