Hierarchical menu trees can be generated by either top-down sorting, i.e. successive divisions of groups of commands into smaller subgroups, or bottom-up sorting, i.e. successive aggregation of small groups of commands into larger superordinate categorie s. Previous researchers have used a hybrid sorting technique employing both top-down and bottom-up procedures together. The present paper examined the effects of top-down and bottom-up sorting tasks separately. Native English speakers, 14 males and 36 females, sorted 25 Automated Teller Machine commands twice. Although sorting time for top-down and bottom-up did not differ, both tasks were performed faster with practice. Top-down and bottom-up sorting tasks were expected to result in different hierar chical trees and mental models: top-down should emphasize differences between commands, while bottom-up should favor similarities among commands being sorted. As predicted, top-down sorting created trees with a greater number of terminal nodes, while bo ttom-up sorting produced trees with a larger mean number of commands per terminal node. In addition, bottom-up sorting generated trees with greater breadth. Moreover, mean breadth increased for both top-down and bottom-up from the first sort to the seco nd sort while depth decreased. Surprisingly, depth was not a factor determining the differences between top-down and bottom-up conditions. Multi-dimensional scaling solutions revealed 3 underlying polarities: 1) object vs. action oriented commands, 2) c hange vs. no change in account balances, and 3) possession of money. Unexpectedly, top-down and bottom-up generated mental models did not converge toward a common model from the first sort to the second sort, but tended to diverge. The results suggest t hat bottom-up trees are superior to top-down. These findings indicate that interface designers should utilize bottom-up sorting in hierarchical menu tree construction.