Google Introduces BERT Update – Improved Search for Conversational Queries

What is BERT?

BERT is a neural network for natural language processing that Google published and open-sourced in 2018 (arXiv). It is based on transformers: models that process each word in relation to all the other words in a sentence, rather than one by one in order. That lets it read context in both directions, before and after a word. Google says BERT was so demanding that it used its latest Cloud TPUs to serve search results for the first time (Google).

In Search, BERT helps with longer, more conversational queries where small words like “for” and “to” change the meaning. Google’s own example: for “2019 brazil traveler to usa need a visa”, earlier systems returned pages about US citizens travelling to Brazil; with BERT, Search understood that the traveller was going from Brazil to the US. BERT also improved featured snippets.

How did BERT develop after 2019?

Date	What happened
2015	RankBrain: Google’s first machine-learning system for understanding queries
Nov 2018	Google publishes and open-sources BERT
25 Oct 2019	BERT launches in Search for about 1 in 10 US English queries and featured snippets (Google)
9 Dec 2019	BERT rolls out to Search in over 70 languages, including Polish (SEL)
Oct 2020	BERT is used in almost every English query (SEL)
May 2021	Google announces MUM, a multitask, multilingual model it describes as 1,000 times more powerful than BERT (Google)
2024–2025	AI Overviews, built on Gemini, launch in the US and reach Poland in March 2025 (SER)
2026	BERT is still listed among Google’s ranking systems (Google)

What does BERT mean for your content?

BERT changed how Google understands queries, not how it judges sites, so there is no “BERT optimisation”. What it did was make keyword matching less important than meaning. The same is true, even more so, for AI answers. In practice (Google):

• Write for the question, not the keyword. Find what people really want to know (search intent) and answer it directly.
• Use natural language. Stuffed phrases like “best cheap shoes Warsaw buy” don’t help a model that reads full sentences.
• Answer first, then explain. Put a clear answer in the first sentences of each section; that suits featured snippets and AI Overviews.
• Cover the topic, not one phrase. Related questions, definitions and examples help models see that your page answers the whole need.
• Check pages that lost rankings for long queries and rewrite them around what those queries ask. Keep doing it; one-off optimization doesn’t last.

FAQ

Can you optimise for BERT?

Not directly. BERT helps Google understand queries; the best response is clear, helpful content written in natural language.

Does BERT work in Polish?

Yes. Since December 2019 BERT has been used in Google Search in over 70 languages, including Polish.

What is the difference between BERT and RankBrain?

Both help Google understand queries. RankBrain (2015) relates words to concepts; BERT reads each word in the context of the whole sentence. Google uses both.

Is BERT the same as the AI in AI Overviews?

No. AI Overviews use Gemini models, which generate answers. BERT is an earlier, smaller model used to understand queries and rank results. See AI search myths.

Sources

• Google: Understanding searches better than ever before (2019)
• Devlin et al.: BERT – Pre-training of Deep Bidirectional Transformers (arXiv)
• Search Engine Land: BERT rolling out in over 70 languages
• Search Engine Land: BERT now used on almost every English query
• Google: MUM – a new AI milestone
• Google Search Central: A guide to Google Search ranking systems