The within-site variability of soil characteristics on sites with different soil types remains poorly quantified, although this information is crucial for the success of research on soil properties, and especially for monitoring soil properties over time. We used coefficients of variation and multilevel variance component models to examine the within-site variation of soil (0–30 cm) mineral nutrient concentrations (P, K, Ca, Mg, Fe, mg g–1; Mn, Zn, mg g–1) and bulk density (kg m–3) on boreal deep-peat sites. We then evaluated the reliability of the site-level estimates (sample means) obtained using different sampling intensities (numbers of samples per site). Our 11 sites represented a single original site type within the oligotrophic nutrient level. Two of the sites were undrained while the rest had been drained for forestry at different points in time. Overall, P concentrations showed the smallest and Mn concentrations the largest within-site variation. The sampling depth contributed more than 50% of the total variance in all other characteristics except the concentrations of P and Fe, and bulk density. The variance proportions of peatland basin, site (within basin), and sampling location (within site) varied by sampling depth for most soil characteristics. The estimates obtained when using a certain number of samples per site were always more reliable for the 0–30 cm layer’s composite samples than for any single 10-cm layer at any depth sampled. On average, it was found that between 4 (P) and some 200 (Mn) samples per site would be needed for the estimates to have a theoretical 10% maximum deviation.