Many methods have been proposed through the last decades for the experimental estimation of site effects. However, these estimates have largely focused on the amplitude of motion, and less attention has been given to duration. The present study is an attempt to fill this gap through a comparison of the phase of the Fourier spectrum at two nearby sites. The method makes use of the frequency-dependent mean group delay as proposed by Sawada (1998). Its application on weak and moderate motion recordings from the European test site near Volvi (Greece) yields stable and physically satisfactory results: the soft sites inside the graben exhibit a consistent increase of the mean group delay, which is maximum around the fundamental frequency of amplification, while rock sites on either side of the graben have comparable spectra, in terms of both modulus and phase. A further comparison with results of 2D finite-difference simulations provides qualitatively similar results. The mean group delays are found, however, to be much smaller on numerical synthetics than on observed records, although the increase of duration for simple input signals is obvious by visual inspection. This may offer a way to discriminate the existence of 2D or 3D effects at a given site. Finally, this technique also allows construction of site-specific time domain synthetics accounting for both the amplitude and phase modifications associated with site conditions. The method requires, however, a good signal-to-noise ratio.