In order to compare model predictions of upper mantle convection with surface observations, we have computed global residual depth and geoid anomalies on a 5¿¿5¿ grid over all oceans, after removal of lithospheric cooling effects and long-wavelength geoid components (>4000 km) related to lower mantle sources. We have derived the transfer function (admittance) of geoid over topography as a function of wavelength λ in the range 1000-4000 km. The observed admittance is positive at all λ and presents a maximum of 2.5-3.5 m/km at λ≂2750 km. Comparing this result with the theoretical admittance computed for the lithospheric cooling plate model, we ascribe most of the admittance peak at λ=2750 km to errors in the depth-age and geoid-age correction. This suggests that a standard age correction applied to all seafloor is not valid. At shorter wavelengths, the admittance remains roughly constant, on the order of 1.5 m/km, and can be explained in terms of sediment loading and crustal thickening effects. Comparing spatially the residual geoid and depth anomaly maps, we note a positive correlation over a few areas coinciding with hot spot swells, volcanic plateaus or seafloor deeps; the geoid versus depth ratio amounts then to ~ 6-8 m/km in agreement with the values predicted above upwellings and downwellings by models of constant viscosity convection. However, the global admittance analysis leads to much lower values at all wavelengths, suggesting that these areas are too limited in number and in extent to produce any significant signal.