Conserving biodiversity often requires assessment of which sites should be prioritised for protection. Sites are often selected based on area or connectivity, with the assumption that a site’s long-term conservation value, as defined by the number of regional species extinctions its removal causes, is smallest for small, disconnected sites. In a simulation study of a mechanistic metacommunity model we find across the parameter range studied that site area is a good predictor of biomass loss following site removal but an insufficient predictor of the long-term species losses incurred as a result. We show that, out of five conceptually distinct predictors tested, including biodiversity, area and connectivity measures, the strongest predictor of long-term species loss (conservation value) is compositional distinctness (average between-site Bray-Curtis dissimilarity) of the impacted community. In extreme cases, small sites located in highly distinct habitats can lead to more species loss when removed than large sites located in more common habitats. Fitting our model to observation data on Andean diatoms and Brazilian lichen-fungi, we show that compositional distinctness exceeds area (total biomass) as a predictor of long-term species losses in the empirically relevant parameter range. Since conservation is primarily concerned with maintaining biodiversity, as opposed to undifferentiated biomass, our results robustly demonstrate that site area alone is not sufficient to gauge conservation priorities; comparative assessment of the community composition of sites is essential.