Handcrafted European papers from the 13th to 19th centuries often feature watermarks. The papermaker added these to make their paper more recognizable. Today these watermarks play a vital role in dating and verifying historical documents, offering valuable insights into paper origins and production. We present a new automated method for finding similar watermarks based on the digitized collections of the German Museum of Books and Writing and the German National Library. Unlike previous approaches, which relied on image processing or neural network classification, our approach utilizes a CycleGAN neural network to generate artificial watermark sketches from scanned historical papers. The sketches are compared using a pre-trained ResNet18 neural network and the Spotify Annoy algorithm for efficient nearest neighbor search in the database. Our pipeline achieves a 63% accuracy in finding similar watermarks within the 50 nearest neighbors from a database of over 6200 photographed watermarks and traced watermarks.