How do unemployed workers form beliefs about their labor-market prospects, and how do these beliefs affect behavior, policy effectiveness, and welfare? We develop and estimate a structural job-search model with quasi-Bayesian updating, combining Swiss panel data on beliefs and job-search behavior with a randomized counseling-and-coaching intervention. Job seekers enter unemployment with overly optimistic beliefs about interview arrivals, but gradually become pessimistic as past signals fade from memory. The estimated updating process combines features consistent with both motivated beliefs and recency-weighted information processing. The model replicates the observed dynamics of applications, interviews, wages, and job finding, while decomposing treatment effects into changes in search efficiency, search costs, and beliefs. We find that belief formation is the principal channel through which the intervention accelerates transitions into employment. However, despite increasing job finding, the intervention reduces participant welfare because it operates primarily through lower reservation wages.