Although scaling relationships that characterize fractal species distributions offer an exciting potential for unification in biogeography, empirical support for fractal theory remains the subject of debate. We synthesize and test multiple predictions of two interrelated fractal models and a null model of random placement using Californian serpentine grassland data describing the spatial location of over 37 000 individually identified plants. The endemics–area relationship and species-abundance distribution recently derived from a community-level fractal property performed poorly because of an inaccurate assumption of homogeneity among species. In contrast, a species-level fractal model that incorporates species-level differences predicted abundances well, but systematically overestimated endemism and predicted a species–area relationship that violated the observed power law. These findings indicate that in order to make predictions based on the existence of a power-law species–area relationship, ecologists need a unifying theory of how the community-level fractal property arises in the presence of species-level distributional differences.