Abstract A database of numerical ground response analyses has been generated to assess the effectiveness of site classification schemes adopted in seismic building codes. 1-D Ground models have been generated with a stochastic approach based on a Monte Carlo procedure and the corresponding seismic response has been evaluated by the equivalent linear elastic approach for a set of natural strong motion records. The study is intended to contribute to the discussion of optimal subsoil classification schemes in terms of inter-class dispersion and reliability of the predicted amplification factors. Moreover some specific analyses have been performed to show the relevance of stiff layer inclusions and of nonlinear soil behavior on the performances of classification schemes. The variability observed for each class in the different systems indicates that simple classification schemes should be preferred according to Occam's razor statement.