Dozens of land surface models (LSMs) now exist, some highly sophisticated, yet their results have not yet begun to converge. Users may be overwhelmed by the huge number of possible combinations of input land characteristics, forcings, and physics packages available. Many studies, including this one, show that the LSMs themselves are the most important factor governing output. Their ranges of output are diverse enough that, for any particular LSM, location, and time period, it may not be possible to reproduce observed states and fluxes, short of tuning the model parameters, no matter what combination of input datasets is chosen. Simulated states and fluxes also may be sensitive to precipitation and radiation forcings, and soil and land cover maps. The observations and atmospheric models which contribute to the datasets available for forcing LSMs have not been consistent enough over time to determine a reliable climatology. What is the effect of this lack of consistency on long term LSM simulations and the value of the resulting output for climate studies? This and other questions relating to LSM sensitivity are explored using NASA's Global Land Data Assimilation System (GLDAS), which includes a massive archive of forcing and land characteristics data, as well as 28 years of output from each of three LSMs.