Video copy detection is mainly required for protecting owners against unauthorized use of their content. Content-based copy detection methods rely on the evaluation of the similarity between potential copies and the original videos. Scalability is the key issue in making these methods practical for very large video databases. To address this challenge, we put forward here an optimized similarity-based search method that takes into account the local characteristics of the space of content signatures. First, refined models of the distortions undergone by the signatures during the copy creation process allow to search in a more appropriately defined area of the description space, increasing query selectivity and improving detection quality. Second, by identifying in the description space those regions where the local density of content signatures is high, a significant additional reduction of the computation cost is obtained. An evaluation on ground truth databases shows that the proposed solution is reliable. Scalability is then demonstrated on larger databases of up to 280,000 hours of video.