BERT & MUM · Natural Language Understanding in Google Search

BERT (Bidirectional Encoder Representations from Transformers) is a neural network architecture for natural language processing developed by Google Research and published in a 2018 paper. Google incorporated BERT into its search systems in October 2019, initially for ~10% of US English queries.

The breakthrough BERT introduced over previous NLP models was bidirectionality. Earlier models read text in one direction (left to right, or right to left) when building understanding of each word's meaning. BERT reads the entire sequence of words simultaneously, building a representation of each word that incorporates the context of all words around it — in both directions.

This means BERT understands that "bank" means something different in "I went to the bank to deposit money" vs "I sat on the river bank" — and it makes this determination by looking at all the surrounding words simultaneously rather than sequentially.

Google gave a specific example at BERT's launch that illustrates its capabilities: the query "2019 brazil traveler to usa need a visa". Before BERT, the word "to" was not well-understood in context — Google would interpret this as a query about Brazilian travellers in general, returning results about US citizens travelling to Brazil. BERT understands "to" as directional context — a Brazilian traveller going to the USA — and returns the correct information about visas required by Brazilians entering the United States.

This example highlights that prepositions, conjunctions, and small contextual words that were previously noise in keyword-based systems became meaningful signals in BERT. Words like "to", "for", "without", "not" fundamentally change query meaning and are now correctly interpreted.

Context changes meaning

BERT also handles negation and modification correctly. "What medicines can I take if I have diabetes but not high blood pressure" requires understanding that the user has diabetes, does not have high blood pressure, and needs medicine that is appropriate for the first condition but not contraindicated by the second. This multi-condition intent is beyond keyword matching and requires BERT-level contextual understanding.

Google's official position on BERT for SEO is consistent: "There's nothing to optimise for with BERT, nor anything for you to be doing differently for pages that may be impacted by this change." This is technically accurate — BERT improved Google's understanding of content and queries, rewarding content that was already written naturally and clearly. The SEO implications are therefore about avoiding practices that conflict with natural language, not about adding new ones.

What BERT rewards

• Clear, natural writing. Content written in natural sentences that a human would use to explain a topic performs better under BERT than content written with unnatural keyword insertions.
• Specific, precise answers. BERT enables Google to extract precise answers to specific questions. Content that directly and specifically answers questions (featured snippet-style) benefits from BERT's ability to match precise answers to precise queries.
• Context and nuance. Content that acknowledges nuance, edge cases, and context-dependent answers aligns with BERT's ability to process complex, conditional queries.

What BERT penalises (through better matching)

• Content with keyword insertions that break natural sentence flow becomes less likely to match the actual intent of BERT-processed queries
• Pages targeting high-volume keywords that are vague in their actual coverage — "best marketing strategies" as a title with generic content — are outranked by specific, precise content that BERT matches to specific queries

MUM (Multitask Unified Model) was announced by Google in May 2021. Where BERT handles text-based natural language understanding, MUM is 1,000x more powerful than BERT according to Google, and extends NLP capabilities in two key dimensions: multimodality and multilinguality.

Multimodal understanding: MUM can process and understand text, images, and potentially video simultaneously. A user could show Google a photo of hiking boots and ask "what socks work well with these for mountain hiking in rainy conditions" — MUM understands the visual characteristics of the boots from the image and the specific functional requirements from the text query.

Multilingual understanding: MUM is trained across 75 languages simultaneously and understands information across languages without translation. It can find the best answer to a query in one language from a source document written in a different language.

Multi-step reasoning: MUM is designed to handle complex, multi-part queries that previously required multiple searches. A query like "I'm preparing for my first marathon in 3 months — I've been running for 2 years, have mild knee pain, and live in a cold climate. What training plan, gear, and nutrition changes do I need?" requires synthesising information from multiple domains simultaneously.

Google has deployed MUM capabilities selectively — primarily for complex topic and question understanding rather than as a wholesale replacement for all query processing. Current MUM applications in Google Search include:

• Topic refinements and subtopics. Google uses MUM to understand the full topic space around a query and surface relevant subtopics users might not have thought to search for.
• Images in search. Google Lens uses MUM-powered visual understanding to identify objects and answer questions about images.
• Cross-language information. Finding relevant information published in other languages and surfacing it (translated) for users whose language has limited coverage of a topic.

MUM's long-term implications for SEO include increasing importance of comprehensive, multi-format content (text + images + video), the continued importance of topical authority (MUM requires comprehensive topic coverage to answer complex queries), and the growing relevance of multilingual content for international sites.

The practical guidance for writing content that performs well under BERT, MUM, and the NLP-driven Google Search of 2026 is essentially the same advice good writers have always followed:

• Write for your reader, not for a search engine. NLP models evaluate content the way a sophisticated human reader would. Natural writing that explains things clearly performs better than keyword-dense writing that reads awkwardly.
• Be specific and precise. NLP models match specific answers to specific questions. Content that answers precisely — with specific numbers, named examples, concrete details — matches more queries more accurately than vague generalisations.
• Cover related concepts naturally. NLP models understand semantic relationships. Writing that naturally covers related entities, synonyms, and concept relationships (because the topic genuinely involves them) signals comprehensive coverage.
• Structure for question answering. Use headings that match natural question phrasings. Provide concise, direct answers at the start of sections before elaborating. This structure aligns with how NLP models extract answers for featured snippets, People Also Ask, and AI Overviews.

Authentic Sources