Large-scale, anonymous online markets solve cooperation problems through reputation systems, which allow traders to readily share and obtain information about exchange partners’ past behavior. Beyond relying on standardized scores in reputation systems, these traders engage in gossip-like production of reputation information in informal communities. We extend the theory of learning and control through social embeddedness to understand how traders and market administrators use different information sources to promote order in markets. We complement market transaction data from two online markets for illegal goods with reputation system scores and community forum texts. First, we find that reputational information in gossip-like discussions complements standardized reputation scores in supporting bottom-up social control by traders themselves. Second, standardized scores and community generated information alike help market administrators exert top-down control by excluding untrustworthy traders from the market. Apart from showing how bottom-up and topdown social control mechanisms interact in promoting cooperation in extra-legal contexts, we demonstrate how manual text coding and deep learning language models can help address sociologically relevant questions.