Google has already pivoted in that direction, they already try to answer questions directly in the response window instead of just providing links. All they would need to do is make a language model provide those answers instead when it is sure enough.

And that would also solve other issues such as trying to validate those answers. The language model could get the first response page and base its answer on those, basically summarising the first 10 pages making it way more reliable than ChatGPT is now.