However, another commenter pointed out above that some implementations allow peeking into the GraphQL query. That way if you know you're going to need the user anyway, you can already perform the join proactively when you fetch the ads.
In a few cases we've done this because we always want the joined results so we just do the join regardless of the query we get.
I still don't understand why this abstraction is useful, if not harmful.
I think the abstraction is somewhat useful if your GraphQL queries vary tremendously (one device asks for the user along with the ads, another device doesn't). Expand this out to a bunch of different data types and you'd find yourself writing a lot of variations of the queries. GraphQL with Data loader allows you to write these things exactly once albeit with potentially suboptimal queries.
All of that being said, we're finding that our queries vary very little and are somewhat questioning the value of data loaders and of GraphQL altogether in our case.