The present paper introduces a novel commercial insertion framework dedicated to video documents broadcasted over the Internet, which aims at ensuring the ads contextual relevance, with respect to the video content, while reducing the level of intrusiveness. Unlike most commercial systems that insert the ads at fixed locations, the proposed system is able to establish, completely automatic, the optimal insertion points determined at the level of the stories boundaries. The core of the proposed methodology concerns a video scene segmentation algorithm based on visual and semantic features extracted using various CNNs architectures.