Neuroimaging researchers increasingly take advantage of the known functional properties of brain regions to localize them and probe changes in their activity under different conditions. The utility of this approach depends in part on the reliability of the methods used to define these regions of interest. Two operations may affect the reliability of functionally identified regions: spatially normalizing data to a stereotactic atlas and statistically combining data across participants to form a composite region (as opposed to identifying individual regions for each participant). The effect of these two operations on reliability was evaluated for two functionally identifiable regions: the MT complex and the frontal eye fields. Spatial normalization had almost no effect on within-subject reliability, while grouping across participants negatively affected retest measures of the activation and location of regions defined on separate occasions. We conclude that, for typical sample sizes and numbers of observations per subject, functional localization is most reliable when performed for each individual using data in atlas space.