BERT (Bidirectional Encoder Representations from Transformers) is a Google natural-language processing model that reads search queries bidirectionally — considering the words on both sides of each term — to understand context, prepositions, and nuance. Google rolled BERT into Search in October 2019 affecting 10% of English queries; it now runs on almost every English search and 70+ languages.
Before BERT, Google treated queries as bags of keywords. If you searched "2019 brazil traveler to usa need a visa", the "to" barely mattered — Google might have returned pages about US citizens travelling to Brazil. BERT reads the "to" and knows the direction matters. That is the whole story.
What is BERT?
BERT is a transformer-based deep learning model for natural language processing (NLP). Google Research published the original paper in October 2018, open-sourced the code, and integrated a version into Google Search one year later. The name expands to Bidirectional Encoder Representations from Transformers — bidirectional because it reads the words to the left and right of each term simultaneously, unlike earlier models that read left-to-right only.
In plain terms: BERT is how Google finally started understanding grammar. Prepositions like "to", "for", and "from" now carry meaning. Query length benefits — the longer and more conversational the search, the more BERT helps. Google specifically called out that BERT improves results for the kind of queries you type into a chatbot.
At launch Google said BERT affected 1 in 10 English queries — its "biggest leap forward in the past five years." By late 2020 Google confirmed BERT is used on almost every English query, and the underlying transformer architecture now feeds later systems like MUM and Google's AI Overviews.
Why BERT matters for SEO
- Long-tail queries improved. Conversational, multi-word queries — the ones voice search creates — are exactly what BERT was built for. If those account for a big share of your traffic, BERT already lifted you or hurt you.
- Keyword stuffing stopped working. BERT understands topics and relationships, not just term repetition. Content written for search engines reads worse to BERT than content written for people.
- Intent-matching beats keyword-matching. A page that satisfies the searcher's actual goal now outranks a page that only stuffs the exact-match phrase.
- Featured snippets got better. Google confirmed BERT improves passage selection for snippets in 25+ countries, meaning the best-written passage wins even if it is not the best-formatted page.
- Foundation for later systems. BERT's transformer architecture is the same family used in MUM, RankBrain updates, and modern AI overview systems. Understanding BERT prepares you for what comes after.
How BERT works
BERT is a transformer neural network trained on billions of words from Wikipedia and books. Two mechanics matter for SEO practitioners:
Bidirectional context
Older NLP models read text one direction. BERT reads both directions at once, so each word's meaning is informed by the words before and after it. In the query "can you get medicine for someone pharmacy", BERT understands "for someone" is critical — the search is about picking up a prescription for another person.
Masked language modelling
During training, BERT was fed sentences with random words hidden and had to predict them. This trains it to weight every position in a sentence against the whole. In practice, it means BERT handles prepositions, negations, and modifiers accurately.
Applied to search
Google applies BERT at query time to interpret what the user meant, and at index time to understand passages within pages. Both sides of the equation get better simultaneously.
BERT vs other Google NLP systems
| System | Launched | What it does |
|---|---|---|
| Hummingbird | 2013 | Entity + intent-based ranking rewrite |
| RankBrain | 2015 | Machine-learned query interpretation |
| BERT | Oct 2019 | Bidirectional NLP for context + prepositions |
| Passage Ranking | Feb 2021 | Rank specific passages inside long pages |
| MUM | Jun 2021 | Multimodal, multi-task successor to BERT |
| AI Overviews | 2024 | Generative summaries built on transformer stack |
Real BERT examples from Google
1. "2019 brazil traveler to usa need a visa"
Before BERT, results showed pages about US citizens travelling to Brazil. The preposition "to" was ignored. After BERT, results correctly show pages about Brazilian citizens travelling to the US. A single preposition flipped the ranking.
2. "Can you get medicine for someone pharmacy"
Before BERT, Google returned generic prescription-filling pages. After BERT, results answer the actual question — picking up a prescription for another person — because BERT understood "for someone" as the critical clause.
3. "Do estheticians stand a lot at work"
Google's own example. Before BERT, results were about esthetician training. After BERT, results were about physical job demands of estheticians. The phrase "stand a lot" now carries meaning.
BERT vs RankBrain — how they relate
RankBrain (2015)
- Learns from past query patterns
- Handles ambiguous / never-seen queries
- Refines results based on click behaviour
- Machine learning applied to ranking
- Example: interpreting acronyms it has not seen
BERT (2019)
- Reads the query as language, not signal
- Handles context, prepositions, negations
- Applied at query and document level
- Transformer-based NLP
- Example: understanding "to" changes direction
They are complementary, not competing. Google uses both simultaneously — BERT for linguistic understanding, RankBrain for interpreting rare queries. Later systems like MUM build on both.
6 best practices for the BERT era
- Write naturally for humans. Google has said there is nothing specific to optimise for BERT. Content that reads well to a person reads well to BERT.
- Cover topics, not keywords. BERT understands topical depth. A page that fully answers the intent outranks one that stuffs the exact-match phrase.
- Use prepositions and modifiers precisely. If the query is "book a hotel near the airport", the answer page should be about proximity, not any airport hotel.
- Answer conversational questions directly. Long-tail, voice-shaped queries are BERT's sweet spot. Structure content with clear question-answer chunks.
- Optimise for passage ranking. Because BERT can pick a paragraph out of a long page, headings and self-contained passages help Google find the right chunk.
- Match search intent completely. A page that half-answers the query, no matter how well-optimised, will lose to a page that fully satisfies BERT-interpreted intent.
You cannot. Google engineers have said this on the record. Any SEO service claiming BERT-specific optimisation is either misleading you or repackaging general content advice. Focus on writing clearly and covering intent — those are the BERT strategies.
Common BERT-era mistakes to avoid
- Keyword-stuffing landing pages. BERT reads context. Stuffed pages read as low-quality and rank lower.
- Thin content covering broad topics. BERT surfaces the page that fully answers a specific intent, not the vaguely-related shorter one.
- Ignoring long-tail queries. Long-tail is exactly what BERT improved for. Skipping conversational content leaves traffic on the table.
- Neglecting non-English markets. BERT now covers 70+ languages. Global sites benefit as much as English-first ones.
- Confusing BERT with a ranking algorithm. BERT is not a "ranking factor" — it is how Google interprets language. It affects everything but is not something you tune for.
How theStacc helps
Content that works in the BERT era is the kind theStacc plans and writes — deeply-covered topics, clear question-answer structure, and passages built for both readers and Google's language models. Combined with entity SEO and technical hygiene, it is what earns durable rankings in a world where generative engines extend BERT's approach further.
Frequently asked questions
BERT stands for Bidirectional Encoder Representations from Transformers. It is a transformer-based NLP model published by Google Research in 2018 and rolled out to Google Search in October 2019.
At launch, BERT affected 10% of English searches. Google confirmed in 2020 that BERT is now used on almost every English query, and it has since been extended to over 70 languages.
Not directly. Google has said there is nothing to optimise for BERT specifically — write naturally, cover intent completely, and match the language real users use. BERT rewards content written for people.
RankBrain (2015) learns from past queries to interpret ambiguous searches. BERT (2019) reads the query bidirectionally to understand grammar and context. Both run together — BERT does not replace RankBrain.
Yes. Google announced BERT is used to improve featured snippet selection in 25+ countries. It helps Google pick the passage that most precisely matches the query intent.
Related glossary terms
Sources
