In late October 2019 Google told the world it had made one of the biggest changes to how it reads search queries in years. The change was called BERT, and unlike most "updates" SEOs talk about, it was not a penalty, a quality filter or a crackdown on any tactic. It was a new way for Google to understand language, especially the small words such as "to", "for" and "no" that change what a person actually means.
BERT still matters in 2026 because it marked the moment Google's search systems moved decisively from matching keywords to modelling meaning. Almost everything that followed, from passage ranking to MUM to today's AI Overviews, builds on that shift.
In this guide you will learn what Google actually announced, how the rollout expanded from US English to almost every English query, what Google's own documentation says BERT does today, and why the honest answer to "how do I optimise for BERT?" is that you don't, at least not directly.
Key Takeaways
- Google announced BERT in Search on 25 October 2019, saying it would help with about one in 10 searches in the US in English.
- By December 2019 BERT was reported to be used for search in more than 70 languages, and in October 2020 Google said it was used in almost every English query.
- BERT is a language-understanding system, not a quality or spam update, so there is nothing to "fix" if rankings moved around that time.
- BERT did not replace RankBrain; it was an additional way to understand queries and content.
- The practical lesson is to write naturally for real questions and match search intent, rather than forcing exact-match keywords.
BERT at a glance
Here are the confirmed facts, with the source for each in the References section.
| Item | Detail |
|---|---|
| Name | BERT (Bidirectional Encoder Representations from Transformers) |
| Announced in Search | 25 October 2019 (Google blog post by Pandu Nayak) |
| Type | Language-understanding system used in ranking and featured snippets, not a quality or spam update |
| Confirmed by Google? | Yes |
| Initial scope | About one in 10 searches in the US in English; featured snippets in the two dozen countries where they were available |
| December 2019 | Reported as adopted for Google Search in more than 70 languages |
| October 2020 | Google said BERT was used in almost every query in English |
| Status in 2026 | Listed as an active system in Google's ranking systems guide |
What Google announced in October 2019
Google's announcement explained that its systems had struggled with longer, conversational queries where prepositions and word order carry the meaning. BERT, a transformer-based model, reads words in relation to all the other words around them, rather than one by one in order. That "bidirectional" context is what lets it understand that "to" in a travel query decides who is travelling where.
Google said BERT would help Search better understand one in 10 searches in the US in English and that it was also using a BERT model to improve featured snippets in the two dozen countries where that feature was available. Search Engine Land's launch coverage added that BERT did not replace RankBrain; it was another method for understanding queries and content. If you want background on that earlier system, see our explainer on Google RankBrain.
“Bidirectional Encoder Representations from Transformers (BERT) is an AI system Google uses that allows us to understand how combinations of words express different meanings and intent.”
— Google Search Central, A guide to Google Search ranking systems
The examples Google used
Google illustrated the change with several before-and-after queries. They are worth studying because they show exactly what kind of misunderstanding BERT was built to fix.
| Example query (from Google's post) | What went wrong before | What BERT understood |
|---|---|---|
| 2019 brazil traveler to usa need a visa | Results were about US citizens travelling to Brazil | The word "to" means a Brazilian travelling to the US |
| do estheticians stand a lot at work | The system matched "stand" with "stand-alone" | The query is about the physical demands of the job |
| parking on a hill with no curb | Too much weight on "curb", ignoring "no" | The negation changes the answer the searcher needs |
How the rollout expanded
BERT did not stay limited to US English for long. Wikipedia's BERT article notes that on 9 December 2019 it was reported that Google Search had adopted BERT for over 70 languages. Then, at Google's Search On event on 15 October 2020, Google said that BERT was now used in almost every English query. Search Engine Land's 2020 review summed this up as BERT going from about 10% of queries to almost all English-language queries within a year.
This is a good example of how Google's language systems tend to roll out: first in one market, then expanded quietly. That pattern repeated with passage ranking and later with MUM.
Who was affected, and how
Because BERT changes how Google interprets queries, its effects were spread thinly across many searches rather than concentrated on particular sites. Google did not publish a list of winners or losers, and you should be sceptical of anyone who claims to know which sites "BERT hit". What can be said with confidence:
- Longer, conversational and question-style queries were the main target, according to Google's own examples.
- Pages that answered the exact intent behind a nuanced question stood a better chance of being matched to it.
- Pages that ranked mainly because they repeated a phrase, without actually answering the query, had less of an advantage when the phrase's meaning was misread.
- Featured snippets were also affected, since Google applied a BERT model to snippet selection.
BERT also arrived only weeks after the September 2019 core update, so if you are analysing historical data from autumn 2019, be careful not to attribute every movement to one change.
Can you optimise for BERT?
Google's announcement did not include any optimisation advice, and BERT is not something a site can "pass" or "fail". Search Engine Land's launch coverage compared it to RankBrain, noting Google had previously told SEOs they can't really optimise for RankBrain. The same logic applies here: BERT helps Google understand language; your job is to give it clear, useful language to understand.
In practice that means the fundamentals described in our guides to search intent and semantic SEO. Here is a simple process for reviewing a page through a "language understanding" lens:
- Write down the three to five real questions a searcher would type to reach the page, in full natural sentences.
- Check that the page answers each question directly, in plain words, near the relevant heading.
- Look for places where you used an awkward exact-match keyword phrase instead of how people actually speak, and rewrite them.
- Make sure qualifiers that change meaning ("for beginners", "without a car", "in winter") are reflected in the content when they matter to the reader.
- Compare your page with what currently ranks for those questions to confirm you match the dominant intent.
- Re-check Search Console queries after changes to see whether the page now appears for the longer, more specific searches you intended.
Common mistakes
- Treating BERT as a penalty. It was not a quality judgement, so there is nothing to "recover" from in the way you might after a core or spam update.
- Buying "BERT optimisation" services. Any tool or agency promising a special BERT score is selling something Google never described.
- Removing stop words from titles and copy. Small words carry meaning; Google's examples show they matter.
- Writing for keywords instead of questions. Repeating a phrase without answering what people mean is exactly what BERT made less effective.
- Blaming BERT for every autumn 2019 drop. Check other changes and seasonality first, using a structured ranking drop diagnosis.
Where BERT fits in the bigger timeline
BERT sits between Hummingbird (2013) and RankBrain on one side, and passage ranking, MUM and generative AI search on the other. Google's ranking systems guide still lists BERT as an active system, describing it as helping Google understand how combinations of words express meaning and intent. For a wider view of how these systems relate, see our overview of Google algorithm updates and the ranking systems guide explained.
Related Guides
- Featured Snippet Deduplication (2020): Why Snippet Pages Lost Their Second Listing
- Google's January 2020 Core Update Explained
- Google's December 2020 Core Update: A Big Update at Year-End
Frequently Asked Questions
When did the Google BERT update launch?
Google announced BERT in Search on 25 October 2019, initially for about one in 10 searches in the US in English.
Is BERT still used by Google?
Yes. Google's ranking systems guide lists BERT as an active system used to understand how combinations of words express meaning and intent.
Did BERT replace RankBrain?
No. Search Engine Land's launch coverage notes BERT was an additional method for understanding queries and content, not a replacement for RankBrain.
How many languages does BERT cover?
It was reported in December 2019 that Google Search used BERT for more than 70 languages, and in October 2020 Google said it was used in almost every English query.
Can I optimise my website for BERT?
Not directly. BERT helps Google understand language, so the best approach is clear content that answers real questions and matches search intent.
Conclusion
BERT was a turning point in how Google understands what people mean, but it never asked site owners to do anything new. Clear writing, honest answers and good intent matching were always the target; BERT simply made Google better at recognising them. If you want help aligning your content with how modern search reads language, our content writing and SEO services teams can review your key pages, or you can request a free quote.
References
- Google: Understanding searches better than ever before (Oct 2019)
- Google Search Central: A guide to Google Search ranking systems
- Search Engine Land: Welcome BERT, Google's AI for understanding search queries
- Wikipedia: BERT (language model)
- Google: Search On 2020 announcements (15 October 2020)
- Search Engine Land: Google algorithm updates 2020 in review
