Exposure to high-ability peers entails positive learning externalities, but it also decreases students’ academic self-concept because of lower ordinal ability rank. We show that, as a result, the linear-in-means parameter identifies a composite (i.e., reduced form) effect. We illustrate the empirical relevance of this issue using data from two experiments that randomly assign students to groups. We find that the structural effect of mean peer ability estimated by a model that includes rank is much larger than the reduced form effect obtained when rank is omitted. This finding also holds in non-linear and heterogeneous peer effects models and helps clarify the mechanisms behind the effects of ability tracking policies.