BERT and SEO: what Google's BERT system does to a search, what Google has said about it, and what it changes in how you write

BERT is a language understanding system that Google applied to Search in October 2019. It reads a query as a whole sentence, so that small words like "for", "to" and "no" change the result the way they change the meaning. For SEO, BERT is not a setting to switch on or a score to chase. It rewards pages that answer the exact situation a person describes, in the words they would use, and it removes most of the benefit of writing in keyword fragments. This page sets out what Google itself has said about BERT, in its own posts, and what follows for the way you write.

What Google announced in October 2019

Google's announcement, "Understanding searches better than ever before", was written by Pandu Nayak and dated 25 October 2019. The main points, in Google's words where they matter:

• BERT stands for Bidirectional Encoder Representations from Transformers. Google had "introduced and open-sourced" it the year before as a technique for natural language processing.
• Transformers process "words in relation to all the other words in a sentence, rather than one-by-one in order". So BERT considers the words before and after each word.
• Google applied BERT to both ranking and featured snippets. For ranking, it said BERT would help Search "better understand one in 10 searches in the U.S. in English", with more languages to follow.
• For featured snippets, Google used a BERT model in the two dozen countries where they were available, and named improvements in Korean, Hindi and Portuguese.
• Some BERT models were complex enough that Google used Cloud TPUs to serve results for the first time.

The post also said that 15 percent of the searches Google sees every day are ones it has not seen before. That is the problem BERT was built for: queries nobody could predict, written the way people talk.

The examples Google gave, and what each one teaches

Google's examples are the most useful part of the post for a writer, because each one shows a small word deciding the answer.

Query	What went wrong before BERT	The lesson for a page
2019 brazil traveler to usa need a visa	Results about US citizens travelling to Brazil	Direction matters. Say who is going where.
do estheticians stand a lot at work	Matched "stand-alone" to "stand"	Context decides the sense of a word.
can you get medicine for someone pharmacy	General results about filling prescriptions	"For someone" is the whole question.
a hill-parking query containing "no curb"	Ignored "no" and returned advice for a hill with a curb	A negative changes the answer completely.

Google was also open about the limits. Even with BERT, a search for "what state is south of Nebraska" returned a guess about a community called South Nebraska. Language understanding, the post said, "remains an ongoing challenge".

What Google said about BERT in 2022

In "How AI powers great search results", dated 3 February 2022 and also written by Pandu Nayak, Google explained how BERT sits beside its other AI systems. The key line: "Today, BERT plays a critical role in almost every English query." Google said BERT helps with two tasks, ranking and retrieving, and that it had improved older systems with BERT training. It also stressed that BERT is "never working alone" but is "part of an ensemble of systems".

The same post placed BERT in a sequence:

• RankBrain, launched in 2015, Google's first deep learning system in Search, which relates words to concepts.
• Neural matching, introduced in 2018, which matches the concepts in a query to the concepts in a page.
• BERT, launched in 2019, which understands how combinations of words express meaning.
• MUM, which Google described as "a thousand times more powerful than BERT" and able to both understand and generate language. Google said MUM was not then used to help rank results the way RankBrain, neural matching and BERT are.

Where BERT sits in Google's documentation now

Google's "guide to Google Search ranking systems" on Search Central, last updated 10 December 2025, still lists BERT as a system it uses, describing it as an AI system "that allows us to understand how combinations of words express different meanings and intent." The same guide says MUM is "not currently used for general ranking in Search" and is used for specific applications. So BERT is not a 2019 event that came and went. It is part of how Google reads queries today.

What BERT changes in how you write

None of Google's posts gives site owners a BERT checklist. They describe how Google reads a query. The writing guidance follows from that:

1. Write for the situation, not the keyword. "Visa for a Brazilian visiting the US" and "visa for an American visiting Brazil" share every keyword and need different pages. Name who, where and which direction in the first lines.
2. Keep the small words. Titles and headings written as fragments ("pharmacy medicine someone") drop the very words BERT reads. Write headings as the question a person asks.
3. Answer one question per section. Long, conversational searches often match one passage on a page. A heading that states the question, with the answer in the first sentence below it, is easy to match and easy to quote in a featured snippet.
4. Cover the negative and the exception. If people search "without", "no" or "instead of", give that case its own answer rather than a passing mention.
5. Read the results page before you write. If the top results answer a different sense of the words than you planned, the query means something else. Our guide to search intent analysis shows how to classify a query from its results page before writing.

The old habit BERT punishes is "keyword-ese", Google's own word for typing strings of words people think a search engine will understand. Pages written in keyword-ese read badly to people, and since 2019 they have also stopped helping with the search engine.

The same full-sentence question, asked of an assistant

BERT taught Google to read a question as a sentence. Assistants like ChatGPT, Gemini and Claude go further: people type the whole situation, often in two or three sentences, and get a written answer that names a few businesses or sources. The pages that do well there are the same ones BERT favours: specific, in plain language, with the answer stated early. If you sell a service, it is worth asking those assistants the conversational version of your best search, the way a customer would phrase it, and seeing who they name. AI Knows Us runs those questions across the main assistants and Google AI Overviews and records which businesses and pages each answer names.

Questions about BERT and SEO

What does BERT stand for?

Bidirectional Encoder Representations from Transformers. "Bidirectional" means the model reads the words on both sides of each word to understand it.

Is Google still using BERT?

Yes. Google said in 2022 that BERT "plays a critical role in almost every English query", and its ranking systems guide, updated in December 2025, still lists BERT.

Can I optimise a page for BERT?

Not with a BERT-specific trick. Google describes BERT as a way to understand queries. Clear writing that answers the exact question, in natural language, is what it rewards.

Is BERT the same as MUM or RankBrain?

No. RankBrain (2015) relates words to concepts, BERT (2019) reads how words combine, and MUM can understand and generate language. Google says MUM is not used for general ranking.

Did BERT cause my traffic to change in late 2019?

It may have, mostly on longer queries. Google said BERT affected about one in 10 US English searches at launch. Compare organic landing-page traffic in your analytics for the weeks either side of 25 October 2019, starting with pages that answer conversational questions.