I think you can maintain the feedback loop without degrading the user experience if you actually ask them what they want to be shown. The problem lies in that

I think you can maintain the feedback loop without degrading the user experience if you actually ask them what they want to be shown. The problem lies in that "engagement" definition : if the user clicked on a thumb, then they must have been attracted to it. In reality, there is considerably more explanation for that click : They could be looking for a specific title, mistakenly thought _this_ actor was playing in the movie, clicked on it to reset the search, wondered if this was a show or a film, missclicked, just picked the least annoying in an ocean of shitty propositions... An algorithm alone can never find its way for better suggestions if it lacks _intent_. The outcome of this lack of data and the lack of data scientists' imagination, will always be stuck in a local minima. So, what solution do you have to gather this precious intent and act on it ? Well, you start small on predictable things : "what would you recommend to Agatha if they liked _Expendables_ ?" "Is _RED_ a good addition to the list if I'm looking for Action/Comedy films ?" And later "Are you looking for Action/Comedy films ? here is some proposals" It works the same for different approaches, for picking thumbnails, or promoting movies. Just ask user for what's motivate them. If Netflix had asked me last week, I would have told them their catalog is getting worse and worse so we prefer using disney+ instead.