Here's a clear example of ML improving search: voice search. You might not use it, but it's extremely popular in India and other developing markets. "G search has gotten worse as they’ve focused on recency in the index, gotten more tolerant of synonyms, and gotten less strict about quoted phrases." None of these are "machine learning" - these are product decisions. If you wanted to say "Google is not an ML company," you'd point to the outsized human influence on search rankings (see, e.g. https://static.googleusercontent.com/media/guidelines.raterh...).
Google Maps is extremely valuable as a proprietary dataset, and we're all making it better whenever we do a captcha, doing object recognition from streetview. So are YouTube, News, Translate, and so many others.
There are so many papers detailing practical metric improvements from ML: https://arxiv.org/abs/1810.09591, is one ("Replacing the manual scoring functionwith a gradient boosted decision tree (GBDT) model gave one of the largest step improvements in homes bookings in Airbnb’s history, with many successful iterations to follow" and deeper neural nets offered significant improvements after that).