Brain-Computer Interfaces (BCI) face a great challenge: how to harness the wide variability of brain signals from a user to another. The most visible problem is the lack of a sound framework to capture the specificity of a user brain waves. A first attempt to leverage this issue is to design user-specific spatial filters, carefully adjusted with a lengthy calibration phase. A second, more recent, opening is the systematic study of brain signals through their covariance, in an appropriate space from a geometric point of view. Riemannian geometry allows to efficiently characterize the variability of inter-subject EEG, even with noisy or scarce data. This contribution is the first attempt for SSVEP-based BCI to make the most of the available data from a user, relying on Riemannian geometry to estimate the similarity with a multiuser dataset. The proposed method is built in the framework of transfer learning and borrows the notion of composite mean to partition the space. This method is evaluated on 12 subjects performing an SSVEP task for the control of an exoskeleton arm and the results show the contribution of Riemannian geometry and of the user-specific composite mean, whereas there is only a few data available for a subject.