According to a widely cited finding by Ellis and Stark (1978 Perception 7 575-581), the duration of eye fixations is longer at the instant of perceptual reversal of an ambiguous figure than before or after the reversal. However, long fixations are more likely to include samples of an independent random event than are short fixations. This sampling bias would produce the pattern of results also when no correlation exists between fixation duration and perceptual reversals. When an appropriate correction is applied to the measurement of fixation durations, the effect disappears. In fact, there are fewer actual button-presses during the long intervals than would be expected by chance. Moving-window analyses performed on eye-fixation data reveal that no unique eye event is associated with switching behaviour. However, several indicators, such as blink frequency, saccade frequency, and the direction of the saccade, are each differentially sensitive to perceptual and response-related aspects of the switching process. The time course of these indicators depicts switching behaviour as a process of cascaded stages.