Estimating soil carbon content as the product of mean carbon concentration and bulk density can result in considerable overestimation. Carbon concentration and soil mass need to be measured on the same sample and carbon contents calculated for each individual sample before averaging. The effect of this bias is likely to be smaller (but still greater than zero) when the primary objective is to determine stock changes over time. Variance and mean carbon content are significantly and positively related to each other, although some sites showed much higher variability than predicted by this relationship, as a likely consequence of their particular site history, forest management, and micro-topography. Because of the proportionality between mean and variance, the number of samples required to detect a fixed change in soil carbon stocks varied directly with the site mean carbon content from less than 10 to several thousands across the range of carbon stocks normally encountered in temperate and Boreal forests. This raises important questions about how to derive an optimal sampling strategy across such a varied range of conditions so as to achieve the aims of the Kyoto Protocol. Overall, on carbon-poor forest sites with little or no disturbance to the soil profile, it is possible to detect changes in total soil organic carbon over time of the order of 0.5 kg (C) m(-2) with manageable sample sizes even using simple random sampling (i.e., about 50 samples per sampling point). More efficient strategies will reveal even smaller differences. On disturbed forest sites (ploughed, windthrow) this is no longer possible (required sample sizes are much larger than 100). Soils developed on coarse aeolian sediments (sand dunes), or where buried logs or harvest residues of the previous rotation are present, can also exhibit large spatial variability in soil carbon. Generally, carbon-rich soils will always require larger numbers of samples. On these sites, simple random sampling is unlikely to be the preferred method, because of its inherent inefficiency. More sophisticated approaches, such as paired re-sampling inside relatively small plots (see, for example, Ellert et al., 2001) are likely to reduce sample size significantly and lead to detection of smaller differences in carbon stocks over time. However, it remains to be shown that at these sites the application of efficient sampling designs will result in the detection of differences relevant for the objectives of the Kyoto Protocol (cf., Conant et al., 2003). Finally, it should also be noted that, compared to the accuracy with which changes in atmospheric carbon content can be detected (less than 1 p.p.m. CO2), changes in soil carbon stocks are very uncertain. A release of 0.5 kg (C) from 1 m2 of soil surface is equivalent to an increase in CO, concentration of about 125 p.p.m. in the air column above the same area.