We propose a model of job search with reference-dependent preferences, where the reference point is given by recent income. Newly unemployed individuals search hard given that they are at a loss, but over time they get used to lower income, and thus search less. They search harder again in anticipation of a benefit cut, only to ultimately get used to it. The model fits the typical shape of the exit from unemployment, including the spike at the UI exhaustion point. The model also makes unique predictions on the response to benefit changes. We provide evidence using a reform in the unemployment system in Hungary. In November 2005, Hungary switched from a single-step UI system to a two-step system, with unchanged overall generosity. We show that the system generated increased hazard rates in anticipation of, and especially following, benefit cuts in ways the standard model has a hard time fitting, even when allowing for unobserved heterogeneity. We structurally estimate the model and estimate a weight on gain-loss utility comparable to the weight on the standard utility term, and a speed of adjustment of the reference point in the order of eight months. The results suggest that a revenue-neutral shift to multiple-step UI systems can speed exit from unemployment. Joint with Stefano DellaVigna, Attila Lindner, and Balázs Reizer.