Computer vision can show greater-than-estimated-human-performance, while still failing hilariously unhuman once in a while. People remember the 1 in a 1000 wonky recommendation that made them do a double-take. Recommendation engines work best for the mean and stereotypical person. That way, you can use information of similar profitable people to effectively recommend.
Facebook, for instance, got mined for "suckers". If you are scummy, you want a list of gullible people who click the most stupidest, poorly designed, and shady ads. Ad tech knows where they are and delivers them on a silver platter. Going back 2 decades to serving ads without ML would kill a business. You don't think they thoroughly test a new recommendation engine and see relevant stats go up before they deploy it? You don't think they can serve you more relevant ads when they know you are a 17 year old male vs. a 42 year old woman? Both the data gathering and the algorithms have improved year over year. To say ad tech personalization is terrible, is akin to complaining we don't have AGI.