The contributions of different sources of error in sampling mixed and unmixed bacterial microcosms were evaluated by using analysis of variance. Culturable heterotrophic bacteria from a turbid freshwater impoundment were sampled from 9-liter tanks that were unagitated or mixed with magnetic stirrers or pumps and from dilution bottles that were unagitated or agitated with a mechanical shaker. Axenic cultures of Enterobacter aerogenes were also sampled from manually shaken test tubes. In both agitated and unagitated tanks and in unagitated dilution bottles, dilutions made from the same sampling pipette were significantly different, showing a clumping of bacteria on the scale of millimeters. Also, microcosms within a single experiment differed from one another by a large margin. Dilution mean squares and tank or bottle mean squares were homogeneous for all types of tanks and unagitated bottles, indicating that the gentle mixing provided by pumps and stir bars did not reduce either millimeter scale or intermicrocosm variability over what prevailed in unagitated microcosms. By contrast, the vigorously shaken bottles and test tubes showed no millimeter scale variability. Intermicrocosm variability was undetectable in test tubes and two orders of magnitude less in shaken bottles than in unshaken bottles. When these facts are coupled with the inherent statistical advantage of replicating large rather than small experimental units, it is concluded that sampling error in the enumeration of aquatic bacteria in microcosms will be reduced by using numerous, small, violently agitated microcosms with a minimum of subsampling per microcosm.