Abstract. This paper investigates the variability of the seismic response of structures subjected to various ensembles of ground motions in the framework of nonlinear dynamic analysis. In this kind of analysis, seismic loading, introduced in the form of excitation accelerograms, constitutes the most important source of uncertainty, as it depends on a number of random factors (e.g. distance from the fault source, magnitude, local soil conditions). In order to quantify the effect of seismic loading simulation on the response variability of structures, nonlinear dynamic analysis is carried out on a steel frame, subjected successively to three ground motion ensembles. The ensembles comprise, respectively, real ground motions and artificial accelerograms generated by two different algorithms, one in accordance with the Code provisions and one targeted to a more realistic simulation of the seismic event. Comparison of the three approaches is made by monitoring the statistical properties of the maximum displacement and interstorey drift induced by the assumed ensembles.