Predicting semantic similarity judgments is often modeled as a three-step process: collecting feature ratings along multiple dimensions (e.g., size, shape, color), computing similarities along each dimension, and combining the latter into an aggregate measure (Nosofsky, 1985). However, such models fail to account for over half of the variance in similarity judgments pertaining to complex, real-world objects (e.g., elephant and bear), even when taking into account their description along dozens of dimensions. To help explain this prediction gap, we propose a two-fold approach. First, we provide the first empirical evidence of a mismatch between similarity predicted by feature ratings and that reported by participants directly along individual dimensions. Second, we show that, surprisingly, separate sub-domains within directly reported dimension-specific similarities carry different amounts of information for predicting object-level similarity judgments. Accordingly, we show that differentially weighting directly reported dimension-specific similarity sub-domains significantly improves prediction of free (i.e., unconstrained) semantic similarity judgments.