For PDF you can do this in a much more subtle way. In a typical block of text every individual letter comes with its own kerning adjustment. You can adjust those in a way that's invisible to the reader but still allows fingerprinting. There's probably 1000 different options too - don't think of moving words as in swapping positions in a sentence. (I know parent suggested it, but that's silly)

These probably wouldn't survive extraction of the pure text, would they?

Replacing characters with identical-looking unicode chars, adding extra spaces here and there, adding newlines (and more spaces :)), adding random typos, use dictionary with "safe" word/phrase replacements etc. And don't forget about formulas, charts etc - pure text version is not too useful on its own

If you deal with fiction and the like where you basically have just text then I think that's correct: it would be trivial to detect the watermarks in various copies by simply comparing them. I was dealing with PDFs containing tables, formulas, illustrations, etc., so a plain-text version would be unusable.

Randomly choose 3 big paragraphs in the entire ebook to add an extra newline in the middle of at the end of a random sentence. This would be my choice if I had to do some kind of invisible watermarking, at least.

This is one of the many things that could be trivially detected and fixed when you have multiple watermarked copies of the same file.

Closer to home and a bit more extreme, a few transposed numbers in a scholarly article would be enough to rekindle another autism/vaccine conspiracy theory!