Employee referrals are a very common means by which firms hire new workers. Past work suggests that workers hired via referrals often perform better than non-referred workers, but we have little understanding as to why. In this paper, we demonstrate that this is primarily because referrals allow firms to select workers better-suited for particular jobs. To test our model, we use novel and detailed productivity and survey data from nine large firms in three industries: call-centers, trucking, and high-tech (software). Referred workers are 10-30% less likely to quit and have substantially higher performance on rare "high-impact metrics" (e.g. creating patents and avoiding truck accidents), despite having similar characteristics and similar performance on non-rare metrics. To identify the source of these behavioral differences, we develop four new statistical tests, all of which indicate that firms benefit from referrals predominantly by selecting workers with a better fit for the job, as opposed to referrals selecting workers with higher overall quality; to referrals enabling monitoring or coaching; or to it being more enjoyable to work with friends. We document that workers refer others like themselves, not only in characteristics but in behavior (e.g. unsafe workers refer other unsafe workers), suggesting that firms may gain by incentivizing referrals most from their highest quality workers. Referred workers achieve substantially higher profits per worker and the difference is driven by referrals from high productivity workers.