We simulate habitat loss and derive species accumulation curves (SAC) and endemics–area relationship curves (EAR) in order to predict expected extinctions. The EAR may have a very different shape depending on the geometry of habitat loss. If area is lost in a spatially random way we may preserve more species than if area is lost in a clustered way, but with a larger extinction debt. If area is lost continuously inwards (‘inward EAR’) then the immediate loss of species can be much greater than if the same area is lost from the core towards its edge (‘outward EAR’). The main reason for these effects is the spatial autocorrelation of species distributions and the definition of endemics. Spatial autocorrelation means that sampling plots that are clustered are occupied by communities with more similar composition. If endemism is defined in relation to the study area, we can observe great species losses at the edge due to the large numbers of ranges that intersect the study area edge, but most of these species persist outside the study area. If instead we examine endemism on a global scale then the pattern of species losses is not influenced by the geometry of habitat loss.