In semantic categorization, masked primes that are category-congruent with the target (e.g., Planets: mars-VENUS) facilitate responses relative to category-incongruent primes (e.g., tree-VENUS). The present study investigated why this category congruence effect is more consistently found with narrow categories (e.g., Numbers larger/smaller than 5, Planets) than with broad categories (e.g., Animals). Experiments 1 and 2 used broad and narrow categories, respectively, and both experiments showed robust frequency effects for non-exemplar targets. This was in contradiction to Forster’s [Forster, K. I. (2004). Category size effects revisited: Frequency and masked priming effects in semantic categorization. Brain & Language, 90(1–3), 276–286.] category search hypothesis, which suggests that frequency-sensitive lexical access is by-passed for narrow categories. These experiments also showed that for non-exemplar targets, semantic feature overlap of non-exemplar primes (e.g., pistol-RIFLE vs. boots-RIFLE) facilitated NO decisions for a broad category but not narrow categories. Experiments 3 and 4 showed that non-exemplar primes which share features with members of narrow categories (e.g., Planets: moon), which we call impostors, facilitated responses to exemplar targets and interfered with responses to non-exemplar targets, relative to unrelated non-exemplar primes. These results suggest that category decisions are made by selectively monitoring semantic features relevant to, but not restricted to those that define, category membership, and they inform the debate regarding the extent to which masked primes are processed semantically.