A critical assumption underlying the use of functional localiser scans is that the voxels identified as the functional region-of-interest (fROI) are essentially the same as those activated by the main experimental manipulation. Intra-subject variability in the location of the fROI violates this assumption, reducing the sensitivity of the analysis and biasing the results. Here we investigated consistency and variability in fROIs in a set of 45 volunteers. They performed two functional localiser scans to identify word- and object-sensitive regions of ventral and lateral occipito-temporal cortex, respectively. In the main analyses, fROIs were defined as the category-selective voxels in each region and consistency was measured as the spatial overlap between scans. Consistency was greatest when minimally selective thresholds were used to define "active" voxels (p<0.05 uncorrected), revealing that approximately 65% of the voxels were commonly activated by both scans. In contrast, highly selective thresholds (p<10(-4) to 10(-6)) yielded the lowest consistency values with less than 25% overlap of the voxels active in both scans. In other words, intra-subject variability was surprisingly high, with between one third and three quarters of the voxels in a given fROI not corresponding to those activated in the main task. This level of variability stands in striking contrast to the consistency seen in retinotopically-defined areas and has important implications for designing robust but efficient functional localiser scans.