Category labels are known to guide inductive generalizations by modifying the representation of stimuli (i.e., the labeling effect); yet, the mechanisms of this effect remain unclear. One view suggests that shared category labels increase overall similarity between items as shared physical features do. The other view suggests that category labels are qualitatively different from category features, as category labels help integrate prior knowledge. The present study examined these two competing views with respect to two types of background knowledge – domains of categories (living things vs. man-made objects) and the amount of knowledge (the number of listed exemplars). The results from two experiments suggest that category labels are likely to be used as features for man-made objects, while category labels help create a naïve assumption for living things, implicating that similarity-based statistical processes and knowledge-based structured processes are employed interactively to cope with different domains of knowledge.