Summary1. Disturbance is an important source of variability in species composition and diversity, but application of disturbance models is contingent upon a very good understanding of the spatial scales and frequencies of disturbance. Such information is particularly lacking from streams. In this study, we measured the disturbance levels of rocks (defined here as the proportion of the original sample disturbed after 6 months) of differing sizes and positions within the stream bed and looked at the variation between and within three upland streams.2. Rocks were blazed with distinctive marks in situ and mapped using simple trigonometry and permanently marked points on the banks. Forty rocks were selected and marked completely at random, and a further sixty marked from random selections within three size classes (small, medium and large) crossed with two bed-packing classes (on top of the bed or packed into the bed). This sampling design was used at each of two sites (an upper, order 3 location and a lower, order 4 location separated by ≈ 8–16 km) on each of three rivers and in two periods during the year (a dry summer period and a wetter winter period) for a sample size of 1200 rocks in all. During summer, on-top rocks were removed from below the study sites and placed at random locations through the riffle after marking, to test whether human-placed rocks can provide estimates of natural disturbance levels.3. Rocks were relocated and classified as disturbed (moved or buried) or not disturbed (found at the same location) after ≈ 6 months. Log-linear modelling revealed that human-placed rocks moved half as often as on-top rocks marked in situ. Overall, small rocks disappeared more frequently than medium ones, which disappeared more often than large ones, and rocks lying loosely on top of the bed were disturbed more often than those packed into the bed. There was no interaction between rock size and bed packing in their effects on disturbance and each of these factors affected disturbance levels in the same way at all six sites and in both seasons. During the summer, there were no differences between upper and lower sites, but disturbance was still relatively frequent and patchy in occurrence, with five of six sites showing significant spatial clumping of disturbed rocks. Disturbance levels were higher in the wetter, winter season than during the drier, summer season, but this was caused by a doubling of disturbance rates at all three lower sites, which also showed reduced (but, in two cases, still significant) levels of spatial aggregation. Disturbance levels at upper sites in winter were similar to summer rates, and the level of aggregation of disturbed rocks differed between sites.4. The results obtained by this study suggest that disturbance levels should not be assessed using methods where rocks are placed in riffles. Disturbance models applied to rocky upland streams may need to heed differences seen at small scales (i.e. between individual rocks), as differences seen at these scales were a crucial source of variability and potentially as significant as variation between sites. However, small-scale differences in disturbance were expressed similarly in different locations. Potentially, the same disturbance model could be applied to all sites, with each of them sitting in different locations along common disturbance continua.