The cost of high-quality aerodynamics simulations for realistic automotive configurations makes comprehensive design studies unfeasible. Data-driven surrogates (learning from data) are an appealing alternative, and there is no shortage of approaches that target shape-to-aerodynamics predictions. However, there is a fundamental limitation (data insufficiency problem) in this context: owing to the proprietary nature of commercial automotive designs, training datasets are limited to a few freely-available geometries. In a previous work the authors, a strategy to construct datasets for training surrogates was introduced. It enables controlled generation of an arbitrary number of samples, by convex interpolation between a small number of basis geometries. In this work, we extend this strategy by introducing three features, namely size, density, and diversity that characterize more general datasets. These are important to assess how useful is a dataset for a specific prediction task (data for learning). A formal measure of diversity is developed and then, datasets of successively increasing diversity but constant size are constructed. We show that the dataset diversity has an impact on the predictive accuracy of machine learning surrogates. A power-law scaling, $ \varepsilon \hskip0.5em \propto \hskip0.5em {M}^{1/2}\hskip0.1em {m}^{-1/6} $, where $ \varepsilon $ is the prediction error, $ M $ is the diversity, and $ m $ the dataset size, collapses 23 controlled experiments onto a single curve, revealing that diversity dominates size in determining prediction error. The proposed framework allows for more rigorous a priori evaluation of models than is currently possible and can be applied readily to other shape optimization problems.