Gait variability may be especially important to measure in people with Alzheimer's disease (AD) as it is related to risk of falling and may reflect the cognitive demand of walking. Its usefulness as an outcome measure in people with AD is currently limited by the lack of published evaluation of its reproducibility. Therefore measures of temporal and spatial gait variability were recorded using an instrumented mat on two occasions, one week apart in 16 community-dwelling people with mild to moderate probable AD. Data were combined in three ways for analysis: all available strides; all available strides from walks with mean velocity within 10cm/s of each other; and the first 12 strides from the second method. Measures of velocity, stride length and cadence variability were all found to have good reliability using an average of 64 strides from velocity-matched walks (ICC3,1 0.77-0.90) however only stride length variability reached acceptable reliability for a clinical test (ICC3,1 0.9). Estimates of the number of strides required to reach an ICC of 0.9 for velocity, cadence and stride width variability were between 169 and 212. Poor to moderate reliability of gait variability measures was obtained using 12 strides. Minimal detectable change values, calculated to reflect absolute agreement, appear to be feasible and may assist with evaluation of interventions to improve gait. Further research should examine the effects on reproducibility of gait variability measures, of systematic cueing aimed at producing consistent, optimal walking in larger groups with a range of dementia type and severity.