That's a bit out there, but Google has mentioned in several different ways that pages and sites have thousands of derived features and attributes they feed into their various ML pipelines.
I assume Google is turning all the site's pages, js, inbound/outbound links, traffic patterns, etc...into large numbers of sometimes obscure datapoints like "does it have a favicon", "is it a unique favicon?", "do people scroll past the initial viewport?", "does it have this known uncommon attribute?".
Maybe those aren't the right guesses, but if a page has thousands of derived features and attributes, maybe they are on the list.
So, some SEO's take the idea that they can identify sites that Google clearly showers with traffic, and try to recreate as close a list of those features/attributes as they can for the site they are being paid to boost.
I agree it's an odd approach, but I also can't prove it's wrong.