There's no mystery, they just have a lot of actual user data, so they've been able to refine the question answering behavior. They've also baked some common problem solving strategies like chain of thought into the model via training.