It is shown that, for mineral soils, it is not the total amount of organic carbon (or organic matter) that controls soil physical behaviour but the amount of complexed organic carbon (COC). We assume that this complex is formed by the association of unit mass (i.e. 1 g) of organic carbon with n grams of clay. Analysis of data from two French and two Polish databases shows that, for these soils, n = 10. A consequence of this is that in soils with small contents of organic carbon (OC), such as arable soils, COC is proportional to OC. However, in soils with large contents of organic carbon, such as pasture soils, COC is proportional to the clay content. This explains why we find that soil bulk density is significantly correlated with OC in French arable soils but with the clay content in French pasture soils. The use of COC instead of OC enables the arable and pasture soils to be considered on the same scale. Water retention data were fitted to a double-exponential equation which allows both the matrix and structural porosities to be estimated. It is shown that in soils with low contents of organic carbon, the carbon content is positively correlated with the matrix porosity. In contrast, in soils with high contents of organic carbon, the matrix porosity is constant at its maximum value and the structural porosity is not significantly correlated with either the total organic carbon or the non-complexed organic carbon (NCOC). It is suggested that the complexed organic carbon can be considered as being sequestered. The soil clay content can similarly be partitioned between clay that is complexed with organic carbon and clay that is not complexed. It is shown that non-complexed clay is more easily dispersed in water than clay that is complexed with organic carbon. These findings indicate how improved pedo-transfer functions for the prediction of soil physical properties may be produced. Such functions need to use the values of complexed and non-complexed organic carbon and clay which must be determined by algorithms. The values produced by the algorithms may then be used in the improved pedo-transfer functions.