The North American Land Data Assimilation System (NLDAS) phase 2 has generated a 31-year (1 January 1979 to 31 December 2009) record of water fluxes, energy fluxes and state variables from four state-of-the-art land models (NCEP/Noah, NASA/Mosaic, OHD/SAC, Princeton and Washington/VIC). It should be noted that SAC model does produce land skin temperature as it is a hydrological model. However, till now, the simulated land skin temperature (LST) has not been comprehensively evaluated. This study uses GOES LST retrievals for the continental United States to validate the NLDAS LST monthly mean diurnal cycle for the period from 1997 to 2009. The results show that the largest errors are in the southwestern continental United States (CONUS) and the smallest errors in the southeastern CONUS for all models. Overall there are larger errors in the western CONUS than in the eastern CONUS. Comparison of the models suggests that VIC has larger errors than Noah and Mosaic, in particular for the simulation of nighttime LST. The comparison analysis between GOES and the model product is used to identify possible reasons that lead to these errors. In addition, this study also uses in situ observations from 22 ARM/CART sites to validate the NLDAS LST monthly mean diurnal cycles for January 1997 to December 1999. The comparison analysis of observed and simulated energy components is used to further investigate what factors affect the model LST simulations. Finally, we suggest future directions to improve NLDAS-based LST simulations from both the model and atmospheric forcing perspective.