My guess (complete speculation, but I worked with the Search Quality folks about a decade ago):
1 - By interpreting your search, it leads to better "search quality" by having one model say "i think this is what they want" and another execute the search. P90 accuracy is increased at the cost of P99 accuracy.
2 - If you search for a literal string you know exists, you expect to find it. By interpreting, fewer search inputs with literal strings make it to the search function.
3 - Since Google is interpreting more searches, this gives ad-placement a route in to favorably interpret "they want to buy something" even when this isn't the case. This makes Google money.
4 - People that used to use literal searches either stop, learn how, or switch search engines.
5 - After a couple years, business metrics show that literal searches represent 0.1% of queries and make less money, is it really worth investing in? When it was a P99 issue it fell off the radar and now the P999 is lost.
So a series or rational decisions by rational actors leads to a decline of a used feature because of business incentives and chasing P90s at the cost of P99s.