those also require real or simulated data to get good recommendations, which is sometimes hard. if you're going to go that far, you might as well create real indexes.
besides, in some environments access to the real data (if there's PII) is much more difficult to get, so simulating is as far as you can go.
I messed around with Postgres hypothetical indexes[1], which was a cheap way to create a lot of them and see what happens.
[1] - https://hypopg.readthedocs.io/en/rel1_stable/