We develop a model of satisficing with evaluation errors that incorporates complexity at the level of individual alternatives. We test the model predictions in a novel data set with information on hundreds of millions of chess moves by experienced players. Consistent with the theory, complex optimal moves are chosen less frequently than simpler ones. Choice frequencies of suboptimal moves follow the opposite pattern. The former finding distinguishes satisficing from a large class of maximization-based models. We further document that skill and time moderate the adverse effect of complexity, and that they complement each other in doing so. Finally, we provide evidence that suboptimal behavior also hinges on the composition of the choice set but not its size. Our findings help to shed some of the first light on the importance of complexity outside of the laboratory.