The unemployment rate remains persistently high after recessions even after job losses subside. Standard search and matching models have difficulty capturing this pattern. In this paper, I provide new evidence that noise shocks, which capture agents’ expectational errors due to the noise in received signals about the persistence of aggregate productivity, generate substantial persistence in the unemployment rate. I first identify these noise shocks using a novel structural VAR and find that unemployment would have recovered to its pre-recession level six quarters earlier in the absence of noise shocks in the 1968-2019 period. To understand these findings, I set up a general equilibrium search and matching model with on-the-job search, endogenous search effort and wage rigidity and consider three shocks: a permanent productivity shock; a transitory productivity shock and a noise shock. The presence of imperfect information and noise shocks allows the model to match the longer unemployment recovery than in a full information setting. The results point to the importance of imperfect information in explaining the slow recoveries, through two channels. First, responses to persistent productivity shocks are more persistent as it takes time for agents to learn whether a shock is persistent or not. Second, noise shocks provide an additional source of persistence, which are amplified through sticky wages, on-the-job search and firms' vacancy posting decisions.