I present evidence from CEO turnover decisions suggesting that boards update their beliefs about CEO ability more in industry crises than in booms. Consistent with predictions from an extended learning model that allows for increased productivity of CEO ability in crises, I find that the turnover-performance relation is weaker the more often the board has observed the CEO in past crises, and crisis performance reduces future dismissal risks more than boom performance. These effects persist after controlling for CEO entrenchment and firm effects, and they are stronger for more severe and recent crises. Employing a proxy of CEO ability, I also find that the dismissal risk of weak-ability CEOs is highest in crises. The results help refine our models of how boards learn about CEO ability and, in particular, help explain the turnover puzzle, i.e., why boards are more likely to dismiss CEOs in industry downturns: rather than misattributing poor industry conditions to CEOs, boards view performance in crises as a more informative signal of CEO ability than performance in booms.