Ecologically meaningful and scientifically defensible nutrient criteria are needed to protect the water quality of USA streams. Criteria based on our best understanding of naturally occurring nutrient concentrations would protect both water quality and aquatic biota. Previous approaches to predicting natural background nutrient concentrations have relied on some form of landscape categorization (e.g., nutrient ecoregions) to account for natural variability among water bodies. However, natural variation within these regions is so high that use of a single criterion can underprotect naturally occurring low-nutrient streams and overprotect naturally occurring high-nutrient steams. We developed Random Forest models to predict how baseflow concentrations of total P (TP) and total N (TN) vary among western USA streams in response to continuous spatial variation in nutrient sources, sinks, or other processes affecting nutrient concentrations. Both models were relatively accurate (root mean square errors <12% of the range of observations for independent validation sites) and made better predictions than previous models of natural nutrient concentrations. However, the models were not very precise (TP model: r²  =  0.46, TN model: r²  =  0.23). An analysis of the sources of variation showed that our models accounted for most of the spatial variation in nutrient concentrations, and much of the imprecision was caused by temporal or measurement variation. We applied 2 methods to determine upper prediction limits that incorporated model error and could be used as site-specific nutrient criteria. These site-specific candidate nutrient criteria better accounted for natural variation among sites than did criteria based on regional average conditions, would increase protection for streams with naturally low nutrient concentrations, and specified more attainable conditions for those streams with naturally higher nutrient concentrations.