Over the past decade, the use of production efficiency models (PEMs) to quantify terrestrial carbon exchange at regional to global scales has been on the rise. This has mainly been due to increased availability of remote sensing data to parameterise these models. However, these models are still subject to large uncertainties. Diagnosis of these uncertainties is necessary to correctly interpret their output and to suggest areas of improvement. In this study, three PEM models (i.e. Carnegie-CASA, C-Fix and MOD17 models) were run in their native format and their capacity to predict gross primary productivity (GPP) at five major biomes across conterminous USA was evaluated against eddy covariance flux tower GPP measurements. The influence of input datasets in the models output was also evaluated. Apart from the cropland biome, the Carnegie-CASA and C-Fix models predicted GPP which were slightly higher than in situ measurements in most of the evaluated biomes (i.e. the needle-leaf evergreen forests, deciduous broadleaf forests, Mediterranean savanna woodlands, and temperate grasslands). The MOD17 model on the other hand predicted lower GPP in most of the evaluated biomes. The overestimation of in situ GPP by the models was attributed to error propagation from the key vegetation biophysical used to drive the models (i.e. the FAPAR product). On the other hand, the low maximum light use efficiency (LUE) term prescribed by the models for particular biomes was responsible for most of the GPP underestimation by the models. Finally, it was noted that the differences in the models structural formulation also resulted in variation of their GPP predictions (e.g. the models which did not account for soil moisture performed poorly in predicting GPP in rain-driven biomes).