Most research results in video watermarking refer to the embedding of messages into uncompressed video, which needs full decoding and then full encoding of the content. In some applications, such a method is too complex, and watermarking of compressed bitstreams is required. Watermarking of compressed bitstreams is more difficult as the number of bits is very limited, and the compressed video is very sensitive to all modifications that may yield visible artifacts in the decoded watermarked video. In the paper, an original watermarking method is described for HEVC-compressed video. The main idea of the technique is to embed a message into nonzero residual transform coefficients in such a way that potential drift is limited. The watermark detection is based on reference, unwatermarked video. The experimental verification demonstrates the strong robustness of the embedded watermarks.