Abstract Random observational errors for sea surface temperature (SST) are estimated using merchant ship reports from the International Comprehensive Ocean–Atmosphere Data Set (ICOADS) for the period of 1970–97. A statistical technique, semivariogram analysis, is used to isolate the variance resulting from the observational error from that resulting from the spatial variability in a dataset of the differences of paired SST reports. The method is largely successful, although there is some evidence that in high-variability regions the separation of random and spatial error is not complete, which may have led to an overestimate of the random observational error in these regions. The error estimates are robust to changes in the details of the regression method used to estimate the spatial variability. The resulting error estimates are shown to vary with region, time, the quality control applied, the method of measurement, the recruiting country, and the source of the data. SST data measured using buckets typically contain smaller random errors than those measured using an engine-intake thermometer. Errors are larger in the 1970s, probably because of problems with data transmission in the early days of the Global Telecommunications System. The best estimate of the global average random error in ICOADS ship SST for the period of 1970–97 is 1.2°C if the estimates are weighted by ocean area and 1.3°C if the estimates are weighted by the number of observations.