Two experiments were conducted to explore the role of configural representations in contextual cuing of visual search. Repeating patterns of distractors (contexts) were trained incidentally as predictive of the target location. Training participants with repeating contexts of consistent configurations led to stronger contextual cuing than when participants were trained with contexts of inconsistent configurations. Computational simulations with an elemental associative learning model of contextual cuing demonstrated that purely elemental representations could not account for the results. However, a configural model of associative learning was able to simulate the ordinal pattern of data.