I'm not convinced this is true. It's more likely they're moving away from a keyword model of search toward a more natural language model because the AI hypetrain goes choo-choo.

It's genuinely difficult to do both of these at the same time. You can do keyword search well, you can do natural language search well, you can't do both at the same time. What you get in a mixed paradigm is confusing and unpredictable results.

This is mostly a limitation in query understanding. The traditional boolean index is still there.

Natural language search engines typically first do a series of traditional keyword searches to identify candidate results, and then re-score them using a natural language understanding of the query.