We reviewed global P export and its controlling factors from 685 world rivers. We used available continuous (runoff, rainfall, catchment area, % land use, and population density) and discrete (runoff type, soil type, biome, dominant land use, dominant type of forest, occurrence of stagnant water bodies in catchment, and Gross Product per Capita [GPC]) variables to predict export of P fractions. P export (kg P km⁻² y⁻¹) spanned 6 orders of magnitude worldwide. The distribution of all fractions of P export (total P [TP], soluble reactive P [SRP], and nonSRP [dissolved organic and particle-bound P]) was right skewed. Export of nonSRP had the highest coefficient of variability, and nonSRP was the dominant part of export. The available environmental variables predicted global P export fairly well (R²  =  0.73) if total N export was included in calculations. The unexplained variance in P export might be attributed to noise in the data set, inaccuracy of measurements of environmental variables at fine scales, lack of quantitative data on anthropogenic P sources, insufficient knowledge of P behavior in catchment soils, and nonlinearity of controlling processes. P exports were highly variable among catchment types, and runoff and population density were the predictors shared by most models. P export appeared to be controlled by different sets of environmental variables in different types of catchments. Quasi-empirical, mechanistic models of P export performed better than did empirical models. Our mechanistic understanding of P export could be improved by refining current analytical methods to obtain fast and reliable values of all P fractions in aquatic ecosystems and by incorporating better and more detailed data on catchment features, anthropogenic sources of P, and instream variables in a mechanistic modelling framework.