We report a study on collaborative learning with Algorithm Visualizations (AV). We have previously confirmed the hypothesis that students' higher engagement has a positive effect on learning outcomes when they learn collaboratively. Thus, in this paper the analysis is targeted on students' collaborative learning process in order to find phenomena that explain the learning improvements. In the video and audio analysis of the learning sessions, we have identified that the amount of collaboration and discussion increase when the level of engagement increases. Furthermore, the groups that used visualizations on higher level of engagement, discussed the learned topic on different levels of abstraction whereas groups that used visualizations on lower levels of engagement tended to concentrate more on only one aspect of the topic. Therefore, one of our conclusions is that the level of engagement indicates, not only the learning performance, but also the amount of on-topic discussions in collaboration. Furthermore, based on previous literature, we claim that the amount and quality of discussions explain the learning performance differences when students use visualizations in collaboration on different levels of engagement.