Early Google was somewhat intentionally optimized for long-tail searches, the thought being that common searches are served fine by most search engines and it's

Early Google was somewhat intentionally optimized for long-tail searches, the thought being that common searches are served fine by most search engines and it's the difficult long-tail searches where people actually try other search engines and cause people to switch their defaults/habits. I suspect it's very difficult to properly define a machine learning objective to specifically optimize for the long-tail. Without very careful crafting, I suspect learn-to-rank optimizes more evenly, resulting in poorer ranking performance in the difficult long-tail. Back 20 years ago when I worked on web search infra at Google, a bunch of the "do what I mean, not what I say" features were disabled if you used any of the search operators in your query. If I recall correctly, it even used its "Kansas" database of per-user data to check if you'd recently used any search operators. Most of that would be very difficult and/or costly to implement with learn-to-rank, though, I guess you could pass a "power_user" boolean to learn-to-rank to try and regain some of the lost functionality.