We present GazeWheels: a series of visual feedback methods for dwell-based gaze input in the form of a wheel that is filled gradually until target selection. We evaluate three variations: Resetting, Pause & Resume and Infinite GazeWheel, and study how dwell duration and visual feedback position (co-located vs remote) impact performance. Findings from a user study (N = 19) show that Infinite and Pause & Resume GazeWheels are error prone but significantly faster than Resetting GazeWheel even when including error correction time. We conclude with five design recommendations.