Moral licensing is an effect where someone who was initially moral subsequently behaves immorally. The effect is controversial, with some studies finding licensing effects and others consistency effects. Two prior meta-analyses attempted to address these discrepancies, finding a small effect (d = .31). However, a considerable amount of heterogeneity is still unexplained. We theorize that methodological differences between studies moderate the effect. Applying reputation-based theory, we predicted that (1) individuals who were observed during licensing would be more moral and produce a larger licensing effect, and (2) dependent measures that make it easy to infer reputation (good/bad) would have smaller licensing effects. We provide a brand-new meta-analysis of 116 studies (k = 157, N > 10,000) that tests moderators. We found a larger licensing effect when participants were explicitly observed (d = .56) than when they were not (d = .12). There was a larger licensing effect when dependent measures were highly morally ambiguous (d = .35) than less ambiguous (ds < .20), which supported our predictions. Overall, the licensing effect was much smaller than previously reported (d = .17).