We define transparent watermarking algorithms as those whose expected distortions of input images are at most c log l, as measured under the average absolute difference metric (l is the range of possible pixel values, and c is a positive constant.) Our definition is based on asymptotic analyses of the expected distortions caused by two prototypical watermarking methods generally considered as transparent: the Patchwork and NEC methods. We also propose some shift-resistant variants of these distortion metrics that incorporate alignment techniques used in DNA string comparisons. Experiments show that these new distortion metrics yield much smaller values when a small number of columns are deleted.