When comparing models against observations, researchers often use long-term averages without due regard for the temporal sampling of the underlying data sets. We study the errors introduced by this practice and show they are often larger than observational errors and comparable to model errors. We further analyse what causes these errors and suggest best practices for eliminating them.