BERT update (2019): what it changed and what it means now

What Google changed#

On 25 October 2019, Google’s Pandu Nayak announced that Search had started using BERT (Bidirectional Encoder Representations from Transformers), a neural network technique for natural language processing. Google called it “the biggest leap forward in the past five years, and one of the biggest leaps forward in the history of Search.”

BERT reads the words in a query in relation to each other rather than one by one. Google said this helps most with longer, conversational queries, and with searches where small words such as “for” and “to” change the meaning. Google applied BERT to ranking and to featured snippets, and needed new hardware to do it: the models ran on its Cloud TPUs.

When it rolled out#

Google’s post is dated 25 October 2019, and Google SearchLiaison posted that day that BERT was already in use for US English searches. Search Engine Roundtable reported that Google said the rollout had begun earlier that week; Google gave no exact start day, so the date here is the announcement date.

On 9 December 2019 Google SearchLiaison posted that BERT was “now rolling out to over 70 languages worldwide.” In October 2020 Google said BERT was “now used in almost every query in English.”

What is known about its impact#

The figures come from Google. At launch it said: “BERT will help Search better understand one in 10 searches in the U.S. in English.” For featured snippets, Google used a BERT model in the two dozen countries where featured snippets were available, and reported significant improvements in languages including Korean, Hindi and Portuguese. Google did not publish site-level traffic effects.

What site owners should do#

BERT is still in use. Google’s ranking systems guide lists it among current systems as an AI system that helps Google “understand how combinations of words express different meanings and intent.” There is no BERT setting to change and no BERT score. What helps is what Google’s helpful content guidance asks for:

• Write for people, in the natural language your readers use when they search.
• Answer the question the page targets clearly and completely.
• Avoid writing for keyword variations at the expense of meaning.

How to check whether it affected you#

Search Console keeps only 16 months of data, so its reports no longer cover late 2019. Archived analytics exports from October to December 2019 are the only record. Look at long, conversational queries, which Google said BERT helped most. The Google Algorithm Updates timeline shows the core updates either side of it.

Related updates#

• RankBrain was Google’s earlier machine learning ranking system.
• Passage ranking was the next language understanding change Google announced.
• Featured snippet deduplication changed featured snippets three months later.