The removal, alteration and fragmentation of habitat are key causes of biodiversity decline worldwide. In Australia, temperate woodlands have been disproportionately cleared following European settlement. Biodiversity decline in such systems may be reversed by restoration of native vegetation on agricultural land. However, rebuilding functioning habitat will require understanding the determinants of species distributions in existing habitat. We used logistic regression of bird occurrence data from 240 sites across northern Victoria, to determine the probability of occurrence of 29 woodland-dependent bird species. We modelled occurrence as a function of habitat variables that characterise both the extent (amount) and composition of native vegetation surrounding sites. Our specific goal was to determine whether the predictive performance of models is improved by accounting for both extent and composition of native vegetation compared with models that characterise native vegetation by extent alone. For nearly all species, accounting for vegetation composition in addition to extent and weighting habitat variables by distance improved the explanatory power of models, explaining on average 5. 4 % (range 0-27. 6 %) of the residual uncertainty in models that accounted for extent alone. Models that incorporate variation in vegetation composition can not only provide more accurate predictions of species occurrence, but also guide more appropriate restoration. They highlight the need for restoration to incorporate sites with fertile soils that support productive vegetation types. These models of woodland birds will be used to inform a spatially-explicit optimisation model for restoring native vegetation cover on agricultural land in this region, with the goal of achieving biodiversity gains while minimizing loss to production. © 2013 Springer Science+Business Media Dordrecht.