We sought to identify those in-site habitat characteristics that best predict distributions of woodland birds in the box ironbark region of central Victoria, Australia. Our focus was on comparing and melding outcomes from several forms of ensemble modelling methods, which account for uncertainty in model structure and allow assessments of variable importance. We used boosted regression trees (BRT), Bayesian additive regression trees (BART) and random forests (RF) to model bird occurrences for 47 species using 43 predictor variables measured at 184 2-ha sites. The majority of predictor variables were in-site habitat variables, but vegetation cover in the surrounding landscape (500 m radius) and geographic coordinates were included to account for known effects of habitat fragmentation and of geographic clines. A consensus model also was developed, built from averaged predictions from the three techniques. We subdivided the avifauna into guilds and other categories (e.g. conservation status) to examine whether there were differences among such subdivisions. Based on cross validation, the consensus model and RF performed best, followed by BART and then BRT. Of the in-site habitat variables, the basal area of red-ironbark trees and groundstorey characteristics such as fineand coarse-litter cover and litter depth had greatest influence on bird occurrences. These results can inform on-site restoration actions (what to restore) and, therefore, complement strategic landscape planning (where and when to restore).