In this paper we present an experimental evaluation of a generic model of cognitive process allowing virtual agents to exhibit anticipatory abilities. This model is included in an agent architecture for agent-based simulations of virtual cities. With user experiments, we show that this mechanism brings about an improvement in the efficiency of the behavior, and check that external observers are able to perceive it. We also confirm that this improvement in efficiency leads, up to a point, to an improvement in believability as judged by human observers. Beyond this level of efficiency, believability reaches a plateau and can even decrease.