This paper presents a completely automatic artifact rejection method for electroencephalographic data based on Riemannian geometry. The proposed method is tested on two P300-based brain-computer interface databases and is shown to allow superior accuracy and information transfer rate as compared to two state-of-the-art methods.