Three rock reference samples. W‐2, DNC‐1, and BIR‐1, were analysed five times over several months by classical chemical methods using a Youden Square as the experimental design. A significant variation for the data classified by months was found for SiO2, CaO, H2O‐, and TiO2among the 15 constituents determined. Nevertheless, most standard deviations for the five data for each oxide in any sample, as well as the error standard deviations from the analyses of variance, are less than or approximately equal to the allowable difference between duplicate determinations traditionally used as limits of error. We also determined nine constituents in three portions each of two samples, W‐2 and DNC‐1, the six portions being analysed concurrently. F ratios for 16 comparisons of the variances of the five data for an oxide obtained over several months and of the three data determined as a group were significant for only one oxide, total iron as Fe2O3in W‐2. We conclude that, if a chemist analyses rock samples carefully, the variation of the determinations will be the same for portions analysed as a group, or analysed over a