BERT—Bidirectional Encoder Representations from Transformers—is a language-representation technique Google open-sourced in 2018 and applied to Search in October 2019. It helps Google understand how combinations of words express meaning and intent by considering surrounding context. BERT improved language understanding; it did not create a special SEO score, plugin, schema type or penalty that site owners can directly optimize.
October 2019
What changed
The original BERT research trained deep bidirectional representations from unlabelled text and then fine-tuned them for tasks such as question answering and language inference. In plain language, the representation of a word can change according to the words before and after it. “Bank” in a financial account and “bank” beside a river should not mean the same thing.
When Google launched BERT in Search, its public examples focused on queries where small words and relationships changed the task: who was travelling to which country, whether medicine was for another person, whether a curb was absent, or whether “stand” described physical work rather than “stand-alone.” Google said it applied BERT models to ranking and featured snippets.
The current Google ranking systems guide still lists BERT as an AI system that helps understand how word combinations express different meanings and intent. That does not mean the original 2019 model, coverage or implementation is frozen. Search uses many systems together, and Google does not publish a BERT report for individual sites or queries.
BERT timeline: research technique to Search system
Coverage figures below are dated historical statements, not permanent percentages for today.
| Date | Documented milestone | What the evidence supports |
|---|---|---|
| October–November 2018 | Google researchers introduced and open-sourced BERT as an NLP pre-training technique. | BERT began as reusable language research, not an SEO update. |
| 2018–2019 | The research reported strong results across 11 natural-language tasks and was published at NAACL 2019. | Benchmark performance explains technical importance but does not reveal Search ranking weights. |
| 25 October 2019 | Google announced BERT in U.S. English Search ranking and in featured snippets across multiple countries. | At launch, Google said it helped understand about one in ten U.S. English searches. |
| 9 December 2019 | Google announced Search rollout across more than 70 languages. | BERT became multilingual quickly; identical quality across every language was not promised. |
| 15 October 2020 | Google said BERT was used in almost every English query. | The original ten-percent figure had already become outdated within a year. |
| Current documentation | Google continues to list BERT alongside other notable ranking systems. | BERT remains relevant context, but modern Search is not explained by BERT alone. |
Translate BERT terminology without turning it into SEO mythology
The research concept, Search product and content recommendation are related—but they are not the same claim.
| Term | Accurate working meaning | Do not infer |
|---|---|---|
| Bidirectional context | A word representation can use context on both sides during BERT pre-training. | Google reads a page exactly like a person or understands every statement. |
| Transformer | A neural architecture that relates input elements through attention mechanisms. | A website needs Transformer markup or a special JavaScript library. |
| Pre-training | A general language representation is learned from large unlabelled text before task-specific fine-tuning. | Google trains directly on a website whenever its wording changes. |
| Fine-tuning | A pre-trained model is adapted for a downstream task. | Site owners can fine-tune Google Search by repeating query variants. |
| Search ranking use | Google said BERT helps understand query meaning when selecting and ordering useful results. | BERT is the only ranking system or a public page-quality score. |
| Featured-snippet use | Google also applied BERT models to improve snippet understanding. | Writing a short paragraph guarantees a featured snippet. |
What it means for SEO
Write naturally, but do not confuse natural language with casual or vague language. Make the actor, action, object, direction, condition, time, location, quantity and exception explicit whenever they change the answer. A grammatically smooth sentence can still be ambiguous or factually empty.
Keep keyword research. Google recommends using words people use in prominent places such as the title and main heading, while noting that language-matching systems can relate a page to queries without every exact variation. The practical shift is away from awkward keyword strings and toward one clear task with complete context.
Separate relevance from quality. BERT can help interpret what a query or passage means; it does not independently prove that the information is accurate, original, current or trustworthy. Evidence, authorship, links, page experience and other systems still matter.
Misconception vs responsible interpretation
| Misconception | Responsible interpretation |
|---|---|
| BERT made keywords obsolete. | User vocabulary still helps discovery and clarity; exact-match repetition is not a substitute for intent. |
| There is a BERT score to improve. | Google exposes no BERT score, Search Console filter or page-level diagnostic. |
| A BERT plugin or schema is required. | No BERT-specific markup or plugin exists for Google Search eligibility. |
| Long conversational copy is automatically better. | Useful context can be concise; filler and unnecessary complexity make the task harder. |
| Use every synonym and related entity. | Choose precise natural terms that support the page purpose, not a generated vocabulary list. |
| BERT rewards correct grammar. | BERT is a language-understanding system, not a public grammar quality factor. Clear editing still benefits users. |
| BERT is a fact checker. | Understanding a claim’s language does not establish whether the claim is true or trustworthy. |
| BERT, RankBrain and Neural Matching are the same. | Google documents them as distinct systems with related language-understanding roles. |
| BERT created AI Overviews. | BERT predates generative Search experiences and is not the same as Gemini-based answer generation. |
| A traffic change in October 2019 proves BERT impact. | Google provides no site-level BERT report; technical, demand, intent, competitors and other systems must be investigated. |
BERT, RankBrain, Neural Matching and other language systems
Use Google’s public descriptions to keep the systems distinct. Their real production interaction is more complex than a one-system SEO diagram.
| System | Google’s documented emphasis | What not to claim |
|---|---|---|
| Hummingbird | A major 2013 improvement to Google’s overall ranking systems; now listed as historical. | Every later language system is merely another name for Hummingbird. |
| RankBrain | Helps understand how words relate to concepts so relevant content can be returned. | It is a measurable user-signal score or synonym counter. |
| Neural Matching | Understands representations of concepts in queries and pages and matches them. | It is an LSI-keyword checklist. |
| BERT | Understands how combinations of words express different meanings and intent. | It is a page-quality score or one-time penalty. |
| Passage Ranking | Identifies individual sections of a page to better understand page relevance. | Google indexes a passage as a separate URL. |
| MUM | Understands and generates language for certain specific Search applications. | Google lists MUM as a general ranking system for every query. |
| Gemini in Search | Supports newer generative Search experiences and capabilities. | Every AI Overview citation can be explained by BERT. |
Practical response
- Define one primary user task, intended audience and successful outcome for the page before drafting.
- Collect real query language from Search Console, customer conversations, internal search, communities and keyword research.
- Group variants that share the same intent and expected answer; do not create one page for every phrasing.
- State the direct answer early using complete, precise sentences rather than a string of keywords.
- Name the actor and object when pronouns such as it, they or this could refer to several things.
- Preserve direction, negation, comparison, units, eligibility, location, date and other small details that can reverse meaning.
- Use descriptive headings based on decisions and subquestions, then keep each section within its promised scope.
- Define specialist terminology and connect it to the language a beginner or buyer actually uses.
- Add examples, calculations, screenshots, evidence or exceptions when they resolve real uncertainty.
- Edit AI-assisted drafts for missing referents, invented certainty, contradictory conditions and unsupported claims.
- Link to genuinely related supporting pages with concise anchors; avoid building thin pages for every long-tail query.
- Measure query clusters, landing-page outcomes and conversions, then improve misunderstood or incomplete sections from evidence.
Context-clarity audit for one page
This audit improves communication for people and machines without pretending to calculate a BERT score.
| Context element | Pass question | Warning sign |
|---|---|---|
| Actor | Is it clear who performs or receives the action? | A pronoun could refer to multiple people, companies or products. |
| Direction | Do from, to, for, by and between express the correct relationship? | Reversing origin and destination changes the answer. |
| Negation | Are no, not, without and exceptions preserved? | A summary removes a negative and reverses the instruction. |
| Scope | Does the statement specify audience, product, location or situation? | A conditional answer is presented as universal advice. |
| Time | Are dates, versions and effective periods explicit? | The page mixes historical and current rules. |
| Quantity | Do numbers include units, denominator and comparison basis? | A percentage appears without sample or timeframe. |
| Reference | Can the reader identify what this, that, former and latter refer to? | The referent is in another paragraph or missing. |
| Terminology | Is the official term used and explained in audience language? | Synonym variation creates several names for one concept. |
| Answer completeness | Are the decision, reason, next step and relevant limitation present? | The page matches keywords but sends the reader back to Search. |
What this does not mean
- Google does not provide a BERT score, impact report, manual action, recovery request or optimization switch.
- The one-in-ten figure described the October 2019 U.S. English launch, not current coverage.
- The more-than-70-languages statement described the December 2019 expansion and did not promise identical performance in every language.
- “Almost every English query” was Google’s October 2020 statement; it should not be converted into an exact present-day percentage.
- Google’s public examples illustrate language challenges but do not expose all production features or ranking weights.
- BERT helps language understanding; it does not guarantee crawling, indexing, ranking, featured snippets, AI citations or clicks.
- BERT is not a substitute for factual accuracy, first-hand evidence, source quality or technical eligibility.
- Writing longer, adding questions or using conversational filler is not BERT optimization.
- Third-party NLP and content scores can support editing but cannot verify Google’s internal interpretation.
- A Search Console change cannot isolate BERT from demand, intent, SERP layout, competitors or other Search systems.
Frequently asked questions
What does BERT stand for?
Bidirectional Encoder Representations from Transformers, a technique for pre-training contextual language representations.
When did Google start using BERT in Search?
Google announced its use in U.S. English Search ranking on 25 October 2019 and soon expanded it across more than 70 languages.
Is BERT still used by Google?
Yes. Google’s current ranking systems guide continues to list BERT as a notable AI system for understanding word combinations, meaning and intent.
Can I optimize specifically for BERT?
There is no special setting. Improve the page’s task, contextual clarity, evidence and usefulness instead.
Did BERT replace keywords?
No. Use the words audiences use, especially in titles and headings, but avoid awkward repetition and pages for every exact variant.
Is BERT the same as RankBrain?
No. Google documents BERT for combinations of words and intent, while RankBrain helps relate words to concepts.
Is BERT the same as Neural Matching?
No. Google describes Neural Matching as matching concept representations in queries and pages; the systems can have related roles without being identical.
Does BERT judge content quality?
Not by itself. Language understanding helps relevance, while Google uses many other systems and signals to assess usefulness and reliability.
Does structured data help BERT?
Structured data can support eligible Search features, but Google provides no BERT-specific schema or markup.
How should multilingual sites respond?
Write and review each language for native meaning, preserve direction and conditions, and avoid literal translation that changes the user’s task.
Official sources and primary references
- Google Search: Understanding searches better than ever before
- Google Research: Open sourcing BERT
- Google Research publication: BERT
- Google Search ranking systems guide: BERT
- Google Search: BERT used in almost every English query
- Search Engine Land: BERT rollout to more than 70 languages
- Google Search: How Google delivers reliable information
- Google Search: Creating helpful, reliable, people-first content
- Google Search Central: SEO Starter Guide
- Ahrefs: Search intent and the three Cs
- Semrush: How the Google Search algorithm works
Continue the Google Search systems series
Review visibility, content, links and technical health without chasing an invented algorithm score.
How Google Ranking Evolved: From PageRank to Modern Search SystemsSeptember 2, 2026
Google AI Content and SEO: What Is Allowed, What Is Spam, and How to Publish SafelySeptember 2, 2026
Google Florida Update (2003): What We Know, What Remains TheorySeptember 2, 2026