Ground-based radiation measurements are frequently used for validating the performance of a model in simulating clouds. Such important questions are often raised as: (1) How well do the measurements represent model grid mean values?; (2) How much of model-observation differences can be attributed to inherent sampling errors?; and (3) What scale does modeling need to be performed in order to capture the cloud variation? We attempt to address these questions using surface solar irradiance data retrieved from the Geostationary Operational Environmental Satellite (GOES) and measured at the Atmospheric Radiation Measurement (ARM) Southern Great Plains (SGP) site. The satellite retrievals are used to mimic ground measurements with various spatial densities and temporal frequencies, from which the sampling errors of the ground observations are quantified and characterized. Most of the differences between point-specific measurements and area-mean satellite retrievals originate from ground sampling errors. We quantify these errors for different months, model grid sizes, and integration intervals. In March 2000, for example, the sampling error is 16 W m-2 for instantaneous irradiances averaged over an area of 10 ¿ 10 km2. It increases to 46 and 64 W m-2 if the model grid size is enlarged to 200 ¿ 200 km2 and 400 ¿ 400 km2, respectively. The sampling uncertainties decrease rapidly as the time-averaging interval increases up to 24 hours and then level off to a relatively small and stable value. Averaging over periods greater than 5 days reduces the error to a magnitude of less than 15 W m-2 over all grid sizes. The sampling error also decreases as the number of ground stations increases inside a grid, but the most substantial reduction occurs as the number of ground sites increases from 1 to 2 or 3 for a grid size of 200 ¿ 200 km2. This means that for computing grid-mean surface solar irradiance, there is no need for an overly dense network of observation stations.