We conduct a global analysis of GRACE-reconstructed gravity gradients from July 2004 to February 2011, to test whether the deep signals preceding the March 2011 Tohoku earthquake can be detected before the event as a specific feature originating from solid Earth. First, we improve the angular resolution of the gravity gradients using two overlapping ranges of azimuthal sensitivity to investigate short-term signals of large amplitude aligned with the orientation of the Northwestern Pacific subduction. Then, we set-up a method to identify consistent solid Earth signals shared by different GRACE gravity models. Robust signals in a model are selected based on their spatial overlap and relative intensity with the signals of another model, so that their sensitivity to the GRACE data processing and ocean dealiasing product can be tested. We show that the dipolar gravity gradient anomaly before the Tohoku earthquake is nearly unique in space and time in the GRACE GRGS03 solutions. A well-resolved dipolar spatial pattern, typical of dislocations within the solid Earth and poorly sensitive to the ocean dealiasing model, is detected. In addition, the preseismic gravity gradient increase is highly consistent between the GRGS03 and CSR06 solutions, independently from their respective oceanic corrections, and can be clearly distinguished from rare anomalies of similar amplitudes all associated with the water cycle over continental areas. Our approach offers solutions for the continuous monitoring of the Pacific subduction belt to document transient slabs motions in real time from global satellite gravity fields, and their relation with shallower deformations and seismic events.