When people say "the watermark", they usually mean two unrelated technologies that happen to share a name. Understanding which one you are looking at explains most of what a removal tool can and cannot do for you.
Visible watermarks
A visible watermark is pixels. A logo, a wordmark or a diagonal band is composited over the finished image, usually at partial opacity so the picture still reads underneath it. Because it is applied last and applied uniformly, it has properties a program can exploit: it sits in the same place across every image from the same source, it has the same shape every time, and the original pixels underneath were dimmed toward the watermark's colour by a consistent amount.
That regularity is what makes visible watermarks removable at all. A tool can locate the mark by matching its known shape, estimate how much of each pixel is watermark and how much is photograph, and reconstruct what was underneath from the surrounding image.
It is reconstruction, not recovery. The information genuinely destroyed by the overlay is gone. What a good repair does is produce pixels that are consistent with their surroundings — plausible rather than true. On smooth backgrounds like sky, wall or blurred bokeh, plausible and true are almost the same thing. Across a face, a line of text or a detailed pattern, they are not, and this is exactly where results get disappointing.
Invisible watermarks
The second kind is a statistical signature woven through the whole image rather than stamped on one corner. Google's SynthID is the best known. Instead of darkening a rectangle, it makes many tiny adjustments spread across the picture, in a pattern that a matching detector can recognise but a person cannot see. There is no logo to find and no region to repair.
This is worth being blunt about: a visible-watermark tool does not touch these. Removing a corner logo has no effect on an embedded signature, because they are not the same mark and are not in the same place. Anything claiming to strip both is describing two very different jobs as though they were one.
The same goes for C2PA "content credentials", which travel as metadata attached to the file. Stripping metadata is easy — but platforms increasingly detect generated images from the pixels themselves, so a clean metadata block does not make an image read as camera-original.
What this means in practice
If your goal is a clean version of an image you made and own, visible watermark removal does exactly what you expect. If your goal is to make an AI-generated image undetectable as AI-generated, no watermark tool achieves that, and platforms label content on their own analysis regardless of what the file says.