Neighbourhood effects on health are partly determined by the way the neighbourhoods are defined (the modifiable areal unit problem), but few studies of place effects have incorporated alternative sets of areal units. This study compared computer-generated zones with areal units identified subjectively by local government officers as communities in the city of Bristol, UK. Automated zone design came close to replicating the subjective communities when the balance of objectives and boundary constraints was adjusted. The set of subjective community areas was compared with automated zone designs, which maximized the homogeneity of a social factor (deprivation) and an environmental factor (housing type), at three different geographical scales, with average populations of 2500, 3700 and 7500. All sets of areas were then matched against the neighbourhood perceptions and social behaviour reported by residents, measured as part of the Avon Longitudinal Study of Parents and Children (ALSPAC). Neighbourhood perceptions and social behaviour varied mostly between individuals, but there were significant small differences between all sets of areas. The neighbourhood perceptions of residents were found to match the areas identified by automated zone design as well as they matched the subjectively defined communities, suggesting that the neighbourhoods identified by experts were not more real to residents than synthetic areas. Differences in perceptions could be explained by variations in social and housing conditions at the very local scale of enumeration districts, with populations of about 500. The neighbourhoods with meaning for residents therefore appeared to be much smaller areas than those typically investigated in geographical studies of health.