Most of my web searches are for looking up specific things, to find the specific link(s) that contains the information I need. These aren’t searches that are going to be made better or faster by an ML model-they’re not natural language queries, they’re just a bunch of terms.
Even before this, you've seen the search engines add features to cater to that kind of use, things like Siri handing off questions it didn't understand to a search engine (as well as the other assistants that can do that), indications of this behavior in how companies like Google show themselves being used in ads.
Of course, regardless of the true prevalence of that behavior, it's probably in Google et al.'s favor to encourage that. Regular search sort of inherently cedes some power and control to the pages the results are coming from, where you're sending users away to if their answer isn't right in the snippet. But the "answer box" features, or an LLM that just tells you "the answer to your question" directly on the page keeps you there, treating the search page as your source of information and not somewhere else.