Spatial autocorrelation in ecological systems is a critical issue for monitoring (and a general understanding of ecological dynamics) yet there are very few data available, especially for riverine systems. Here, we report here on assemblage-level autocorrelation in the benthic-invertebrate assemblages of riffles in two adjacent, relatively pristine rivers in south-eastern Victoria, Australia (40-km reaches of the Wellington [surveys in summers of 1996 and 1997] and Wonnangatta Rivers [survey in summer of 1996 only], with 16 sites in each river). We found that analyses were similar if the data were resolved to family or to species level. Spatial autocorrelation was assessed by using Mantel-tests for the data partitioned into different sets of spatial separations of survey sites (e.g. 0-6 km, 6-12 km, etc.). We found strong small-scale (≤6 km) autocorrelation in the Wellington River, which is consistent with known dispersal abilities of many aquatic invertebrates. Surprisingly, there were strong negative correlations at longer distance classes for the Wellington River in one of the two summers (20-40 km) and the Wonnangatta River (12-20 km). That two largely unimpacted, adjacent rivers should have such different autocorrelation patterns suggests that impact assessment cannot assume dependence or independence of sites a priori. We discuss the implications of these results for use of "reference" sites to assess impacts at nominally affected sites. © Springer Science+Business Media, Inc. 2006.