The present computation of sea surface temperature (SST) from infrared satellite measurements requires a coincident sample of in situ (drifting buoy and/or ship) SST measurements, to compute by regression the algorithmic coefficients for the infrared data. Ignoring the fundamental difference between satellite-measured "skin SST" and buoy/ship measured "bulk SST," we analyze past buoy and ship SST data to better evaluate the errors involved in the routine computation of SST from operational satellite data. We use buoy and ship SST data for 2 years (1990 and 1996) from the Comprehensive Ocean-Atmosphere Data Set as well as 2 years of previously cloud-cleared satellite radiances with matching drifting/moored buoy SST data from the NASA Pathfinder SST data set. We examine the in situ SST data for geographic distribution, accuracy, and self-consistency. We find that there are large geographic regions that are frequently not sampled by the present drifting buoy network, a natural consequence of the fact that most buoys are not deployed to measure in situ SST for satellite infrared SST calibration. Comparisons between drifting buoy SSTs suggest an error of ∼0.4°C for nearly coincident buoy SSTs. Comparing moored with adjacent drifting buoy SSTs, we find that drifting and moored buoy SSTs are samples from the same population. Ship SSTs are noisier and have a significant warm bias relative to drifting buoy SSTs. We explore the SST measurement accuracy changes that occur with variations in sampling coverage used for the SST algorithm regression. We both vary the total amount of points and restrict the regression data to regional sampling biases. Surprisingly the total number of calibration SST values can be quite small if they cover all latitudes. We conclude that buoy SSTs can have residual bias errors of ∼0.15°C with RMS errors closer to 0.5°C.