There is an increasing need for ecosystem-level distribution models (EDMs) and a better understanding of which factors affect their quality. We investigated how the performance and transferability of EDMs are influenced by (1) the choice of predictors, and (2) model complexity. We modelled the distribution of 15 pre-classified ecosystem types in Norway using 252 predictors gridded to 100 m × 100 m resolution. The ecosystem types are major types in the "Nature in Norway" system mainly defined by rule-based criteria such as whether soil or specific functional groups (e.g., trees) are present. The predictors were categorised into four groups, of which three represented proxies for natural, anthropogenic, or terrain processes (‘ecological predictors’) and one represented spectral and structural characteristics of the surface observable from above (’surface predictors’). Models were generated for five levels of model complexity. Model performance and transferability were evaluated with data collected independently of the training data. We found that (1) models trained with surface predictors only, performed considerably better and were more transferable than models trained with ecological predictors, and (2) model performance increased with model complexity, levelling off from around 10 parameters and reaching a peak around 20 parameters, while model transferability decreased with model complexity. Our findings support that surface predictors enhance EDM performance and transferability, most likely because they represent discernible surface characteristics of the ecosystem types. A poor match between the rule-based criteria that define the ecosystem types and the ecological predictors, which represent ecological processes, is a plausible explanation for why surface predictors better predict the distribution of ecosystem types. Our results indicate that, in most cases, the same models are not well suited focontrasting purposes, such as predicting where ecosystems are and explaining why they are there.