A driver steering a car on a twisting road has two distinct tasks: to match the road curvature, and to keep a proper distance from the lane edges. Both are achieved by turning the steering wheel, but it is not clear which part or parts of the road ahead supply the visual information needed, or how it is used. Current models of the behaviour of real drivers or 'co-driver' simulators vary greatly in their implementation of these tasks, but all agree that successful steering requires the driver to monitor the angular deviation of the road from the vehicle's present heading at some 'preview' distance ahead, typically about 1 s into the future. Eye movement recordings generally support this view. Here we have used a simple road simulator, in which only certain parts of the road are displayed, to show that at moderate to high speeds accurate driving requires that both a distant and a near region of the road are visible. The former is used to estimate road curvature and the latter to provide position-in-lane feedback. At lower speeds only the near region is necessary. These results support a two-stage model of driver behaviour.