Mapping the thermal and compositional structure of the upper mantle requires a combined interpretation of geophysical and petrological observations. Based on current knowledge of material properties, we interpret available global seismic models for temperature assuming end-member compositional structures. In particular, we test the effects of modelling a depleted lithosphere, which accounts for petrological constraints on continents. Differences between seismic models translate into large temperature and density variations, respectively, up to 400 K and 0.06 g cm−3 at 150 km depth. Introducing lateral compositional variations does not change significantly the thermal interpretation of seismic models, but gives a more realistic density structure. Modelling a petrological lithosphere gives cratonic temperatures at 150 km depth that are only 100 K hotter than those obtained assuming pyrolite, but density is ∼0.1 g cm−3 lower. We determined the geoid and topography associated with the density distributions by computing the instantaneous flow with an existing code of mantle convection, STAG-YY. Models with and without lateral variations in viscosity have been tested. We found that the differences between seismic models in the deeper part of the upper mantle significantly affect the global geoid, even at harmonic degree 2. The range of variance reduction for geoid due to differences in the transition zone structure (i.e. from 410 to 660 km) is comparable with the range due to differences in the whole mantle seismic structure. Since geoid is dominated by very long wavelengths (the lowest five harmonic degrees account for more than 90 per cent of the signal power), the lithospheric density contrasts do not strongly affect its overall pattern. Models that include a petrological lithosphere, however, fit the geoid and topography better. Most of the long-wavelength contribution that helps to improve the fit comes from the oceanic lithosphere. The signature of continental lithosphere worsens the fit, even in simulations that assume an extremely viscous lithosphere. Therefore, a less depleted, and thus less buoyant, continental lithosphere is required to explain gravity data. None of the seismic tomography models we analyse is able to reproduce accurately the thermal structure of the oceanic lithosphere. All of them show their lowest seismic velocities at ∼100 km depth beneath mid-oceanic ridges and have much higher velocities at shallower depths compared to what is predicted with standard cooling models. Despite the limited resolution of global seismic models, this seems to suggest the presence of an additional compositional complexity in the lithosphere.