Investigations of search within realistic scenes have identified both bottom-up and top-down influences on performance. Here, I describe two types of top-down expectations that might guide observers looking for objects. Initially, likely locations can be predicted based only on the target identity but without any visual information from the scene (“Eyes closed”). When a visual preview becomes available, a more refined prediction can be made based on scene layout (“Eyes open”). In two experiments participants guessed the location of a target with or without a brief preview of the scene. Responses were consistent between observers and were used to predict the eye movements of new observers in a third experiment. The results confirm that participants use both types of top-down cues during search, and provide a simple method for estimating these expectations in predictive models.