The Global Land Data Assimilation System (GLDAS) project, based at NASA Goddard Space Flight Center, forces the Mosaic land surface model (LSM) with realistic global fields of precipitation, radiation, and near-surface meteorology in order to generate realistic fields (at least within the context of the LSM) of soil moisture, temperature, and other land surface states. A key motivation for this exercise is the provision of realistic land surface initial conditions for seasonal forecasts. The degree to which improved initial conditions can actually improve such forecasts, however, has never been quantified. We examine this question with side-by-side ensembles of seasonal forecasts performed with the NASA Seasonal-to-Interannual Prediction Project (NSIPP) climate modeling system. For each forecast period, we run one ensemble of simulations that makes use of the improved initial conditions and a second ensemble that does not. The parallel forecasts cover a number of summer seasons (1979-1993), thus providing an adequate statistical basis to identify any improvement in forecast skill.