In geometric morphometrics, the extent of variation attributable to non-biological causes (i.e. measurement error) is sometimes overlooked. The effects of this variation on downstream statistical analyses are also largely unknown. In particular, it is unclear whether specimen preservation induces substantial variation in shape and whether such variation affects downstream statistical inference. Using a combination of empirical fish body shape data and realistic simulations, we show that preservation introduces substantial artefactual variation and significant non-random error (i.e. bias). Most changes in shape occur when fresh fish are frozen and thawed, whereas a smaller change in shape is observed when frozen and thawed fish are fixed in formalin and transferred to ethanol. Surprisingly, we also show that, in our case, preservation produces only minor effects on three downstream analyses of shape variation: classification using canonical variate analysis, permutation tests of differences in means and computations of differences in mean shape between groups. Even mixing of differently preserved specimens has a relatively small effect on downstream analyses. However, we suggest that mixing fish with different preservation should still be avoided and discuss the conditions in which this practice might be justified.