Previous work investigating relational categories has explored the potential of within-category comparison to promote relational category learning and transfer. This follows from predictions of structure mapping theory (Gentner, 1983, 2003) that alignment leads to highlighting and abstraction of common relational structure. However, straightforward merging of traditional classification learning with comparison has not been effective. We explore the hypothesis that classification and comparison have an unforeseen incompatibility. In a 3x2 between-subjects design we tested three presentation conditions (unconstrained item pairs, category-matched items pairs, single items) in two supervised learning modes: classification and observation. Results demonstrate an interaction driven by highly accurate categorization for observational learners with same-category pairs. The observational mode yielded the predicted, elusive result of an advantage for within-category pairs over twice as many single-item trials. We conclude within-category comparison can be an effective means to promote relational category learning and discuss the apparent impediment of the guess-and-correct cycle.