I think people often underestimate the power of much simpler methods. You'd probably get substantial improvements over the baseline just by penalizing sites for having the same words too many times. I forget the exact number, but I think it's something like 50% of the words in your average article appear only once. So if two articles have even 5% similarity among the words they only use once, they are probably quite similar.

AI of course can get you much farther still, but you definitely don't need something that powerful for a convincing upgrade from baseline.