Obviously the future is to train the model with the ads, so that they're indistinguishable from the core of the answer.

I kid, but also hope I'm wrong.

Hm, m$ also runs a few giant adtech platforms, maybe they can just inject tracking code at the source.