We use data from two experiments that randomly assign students to groups to show that, so long as ordinal rank has a causal effect on educational achievement, estimates of the effects of peer ability composition obtained from models that omit rank are downward biased. This finding holds both in the standard linear-in-means model as well as in models that allow for non-linear and heterogeneous peer effects, and contributes to explain why previous studies have detected only modest effects of peer ability on achievement. We also illustrate how this finding helps understand the mechanisms behind the effects of ability tracking policies.