New concepts and methods for modeling the natural variability of soil properties are presented and illustrated. The proposed technique of modeling the statistical character of soil profiles serves a dual function: (1)It provides a format for quantifying the information gathered during site investigation and testing, about the subsurface conditions at a site; and (2)it provides the basis for predicting performance and for quantifying the reliability of performance predictions. Probabilistic soil profiles are characterized, first, by best estimates of layer depths and of pertinent engineering properties; and secondly, by the coefficient of variation and the correlation scales for the profile parameters of interest. Methodology is developed for dealing with problems that can be formulated in terms of extremes of averages of soil properties. The problems of limit equilibrium slope stability and differential settlement prediction fall into this category.