HATEOAS was and is implemented widely and effortlessly in hypermedia, since Fielding was describing the existing web architecture. It has failed fairly dramatically in JSON APIs, because JSON is not a native hypertext. Early API designers tried to shoehorn HATEOAS concepts into their APIs, which was somewhat plausible when APIs were XML, which, when squinted at, looked sort of like a hypermedia. Once we kicked over to JSON (and back to RPC as a network architecture) it became silly.
I’m not arguing against that, just that huge chunks of the web break the HATEOAS abstraction, as it were, and yet, as you pointed out, “the web is pretty successful overall”. It’s not super clear to me that this is because of hypermedia or not in the sense that correlation is not causality.
The Web went all JavaScript instead of HTML sometime between 2000(2005?) and 2010(2015?). But by 2000 it was already wildly successful, so apparently it must have been HTML -- "hypermedia" -- that was the key to success, not JavaScript / JSON.
