Optimal assignment and matching mechanisms have been the focus of exhaustive analysis. We focus on their dynamic effects, which have received less attention, especially in the empirical literature: Anticipating that assignment is based on prior performance may affect prior performance. We test this hypothesis in a lab experiment. Participants first perform a task individually without monetary incentives; in a second stage, they are paired with another participant according to a pre‐announced assignment policy. The assignment is based on the first‐stage performance, and compensation is determined by average performance. Our results are largely consistent with a theory: Pairing the worst‐performing individuals with the best yields 20% lower first‐stage effort than random matching (RAM) and does not induce truthful revelation of types, which undoes any policy that aims to reallocate types based on performance. Perhaps surprisingly, however, pairing the best with the best yields only 5% higher first‐stage effort than RAM and the difference is not statistically significant.