In this paper, we evaluate the performance of 19 coupled general circulation models (CGCMs) in twentieth‐century simulations of the Sahel during the 1970s to 1990s. Correlation, regression, and cluster analyses are applied to observations and model outputs including Sahel monthly precipitation, evaporation, soil moisture, and sea surface temperature (SST). We find that only eight CGCMs (hit models) produce a reasonable Sahel drought signal, while seven CGCMs (miss models) produce excessive rainfall over the Sahel during the observed drought period. Even the model with the highest prediction skill of the Sahel drought could only predict the increasing trend of severe drought events but not the beginning and duration of the events. From analyses of the statistical characteristics of the hit and miss models, we conclude that a good simulation of the Sahel drought requires (1) a strong coupling between Sahel rainfall and Indian Ocean SST, with warm (cold) SST identified with Sahel drought (flood), (2) a significant coupling between Sahel rainfall and the Atlantic Ocean SST, with a warm equatorial Atlantic and cold extratropical North Atlantic coexisting with Sahel drought, and vice versa, and (3) a robust land surface feedback with strong sensitivity of precipitation and land evaporation to soil moisture. These three characteristics constitute sufficient conditions for a good simulation of Sahel drought in CGCMs.