Many Web sites provide consumers with product recommendations, which are typically presented by a sequence of verbal reviews and numerical ratings. In three experiments, we demonstrate that when participants switch between formats (e.g., from verbal to numerical), they are more prone to preference inconsistencies than when they aggregate the recommendations within the same format (e.g., verbal). When evaluating recommendations, participants rely primarily on central-location measures (e.g., mean) and less on other distribution characteristics (e.g., variance). We explain our findings within the theoretical framework of stimulus-response compatibility and we make practical recommendations for the design of recommendation systems and Web portals.