Research has shown that integration of visual and verbal information sources during learning promotes successful student outcomes. However, it is unclear whether it is better to provide students with integrated visual-verbal representations, or to require them to build such integrated representations themselves. In a classroom study, three conditions were used to explore the impact of integrated visual-verbal representations that emphasized rule-diagram mappings in geometry. Students viewed highlighted rule-diagram mappings during learning, generated these mappings themselves, or saw only numerical information embedded in diagrams (control). Students problem-solving knowledge was measured at posttest and delayed posttest. Overall, students who generated rule-diagram mappings during intelligent tutoring demonstrated better long-term understanding of geometry principles, but effects were only visible at delayed posttest. Results show that integrated visual-verbal representations best support deep learning when they help the learner make connections between features of a visual representation and relevant domain information, and student interactions can be an effective method to scaffold these connections.