My suspect is BERT and semantic search. Before it most word embedding models were monolingual and non contextual, so they were not useful for search. So Google

My suspect is BERT and semantic search. Before it most word embedding models were monolingual and non contextual, so they were not useful for search. So Google had to use inverse indexes and an algorithm fine tuned for years. With BERT they can have a model to embed all the search queries for 99.99% languages and works sort of fine for 99% of the queries, so they are fine with that as well. It would have been sort of easy detecting that a query works better with the previous index than the new one, but it seems that they decided against that. So we lost a lot of usefulness, like phone number search.