On October 25, 2019, Google announced in a blog post by Pandu Nayak, a Google Fellow and Vice President of Search, that it had begun applying BERT, the transformer-based language model released by Google AI the year before, to understanding search queries. Nayak called it “the biggest leap forward in the past five years, and one of the biggest leaps forward in the history of Search.”
The motivation was that short search queries often hinge on small words, such as prepositions like “to” or “for,” that change the meaning of the request. Earlier keyword-style matching could miss this, returning results based on the important-looking words while ignoring how they fit together. BERT’s strength is reading a query in context, both directions at once, so it can grasp the intent behind a natural phrasing. Google said BERT would help it better understand roughly one in ten searches in the United States in English, with the largest gains on longer, conversational queries, and that it was also using BERT to improve featured snippets across two dozen countries.
The announcement marked a visible turning point where modern deep language models moved from research into the most-used software product in the world, and it presaged the broader shift toward language understanding and dense retrieval in search.
For a business reader, this is when search stopped being mainly about matching words and started being about understanding sentences, changing both how people phrase queries and how content must be written to be found.