In the comparison of treatments in soil science experiments, a well-chosen experimental design is of crucial importance for making appropriate (optimal) statistical inference at low (lowest) costs. In the very common case that several observations from each plot are obtained, one should be aware that measurement units are not equivalent to so-called statistical experimental units. We illustrate the severe consequences of misinterpretations in some very commonly used experimental designs. Appropriate methodology is presented, which entails the aggregation of observations at the plot level based on the tools of spatial statistics, followed by a classical statistical analysis on the plot aggregated data. The model-based approach, in contrast to the design-based approach, appears to be a natural setting for comparison of treatments in soil-based experiments. Consequences for the planning of the experimental design are discussed, and recommendations are given.