The first version of the feature heavily prioritized two things: helping the user understand that the playlist was being generated in the moment based on their search, and having Spotify's AI DJ "take credit" for the playlist. These features tested well in usability testing. Users were enthusiastic, and were able to identify that "Spotify is making the playlist for me in real time".
However... our first A/B test revealed something different. Users weren't tapping on the playlist, and consumption from the second item skyrocketed. But at the same time, the users who were making it into the playlist were having longer-than-average listening sessions. The playlist was working well, but the entry point was not. A combination of survey responses and data analysis revealed that banner blindness was the issue. Users were so used to their current search workflows that they were missing the top result entirely.