Google uses several AI systems within Search, including neural matching, RankBrain, BERT and MUM. Google describes neural matching as matching representations of concepts in queries and pages, RankBrain as relating words to concepts, BERT as understanding how combinations of words express meaning and intent, and MUM as supporting specific Search applications rather than general ranking. Google has not published the full architecture of how these systems process pages, how they interact, or what internal pipeline they use. No separate BERT, RankBrain, or MUM optimisation checklist exists. Foundational SEO — satisfying search intent with accurate, well-structured content — remains the correct approach.
- Google describes RankBrain as a system that helps relate words to concepts. Google has not documented RankBrain as a dwell-time, pogo-sticking, or engagement-learning algorithm.
- BERT helps Google understand how combinations of words express meaning and intent. Google confirmed its use for understanding searches, particularly queries and featured snippets, in 2019.
- MUM is used for specific Search applications. Google explicitly states it is not used for general ranking.
- Neural matching helps Google match representations of concepts in queries and pages — it is a distinct system, not interchangeable with BERT or RankBrain.
- Google has not published a five-stage page-processing pipeline, a universal document embedding architecture, or a semantic-completeness scoring formula.
- Technical terms from machine learning — self-attention, embeddings, transformers — describe how these models work in principle. They do not prove how Google Search deploys them, in what order, or on which documents.
- The correct SEO response is to satisfy reader intent with accurate, clearly structured content. Do not optimise for imagined model scores.
A common pattern in SEO coverage of BERT, MUM and RankBrain is to take real machine-learning concepts, attach them to an undocumented Google architecture and derive optimisation advice from that assumed architecture. The result is articles that sound technical but teach a fictional version of how Search works.
This article takes a different approach. It separates what Google has confirmed from what Google has not disclosed, and derives SEO implications from the first category only. Where the architecture is unknown, this article says so.
For the foundational case that these systems matter for semantic SEO, see Article 1 in this cluster.
Evidence Framework
Before each system, claims in this article are labelled by evidence class:
| Label | Meaning |
|---|---|
| [Google-confirmed] | Explicitly documented by Google |
| [Google research background] | Described in Google research, not necessarily deployed in Search as stated |
| [General NLP concept] | Technically accurate machine-learning background; does not prove Google Search deployment |
| [Practitioner implication] | Safe SEO interpretation; not a disclosed ranking mechanism |
| [Unknown] | Google has not published the architecture or mechanism |
Search Uses Multiple Systems — Not One NLP Algorithm
A common misconception is that Google’s language understanding can be attributed to a single model. Google’s ranking systems guide identifies multiple distinct systems that work together:
- Neural matching — matches representations of concepts in queries and pages
- RankBrain — helps Google understand how words relate to concepts
- BERT — helps Google understand how combinations of words express meaning and intent
- MUM — used for specific Search applications
Google documents these as distinct systems or technologies. It does not publish their complete interaction, weighting or lifecycle within Search. (Google Ranking Systems Guide)
Neural Matching
[Google-confirmed] Google describes neural matching as a system that helps it understand representations of concepts in queries and pages. It helps Google connect queries to relevant content even when the exact words in the query do not appear in the page.
Google lists neural matching, BERT and RankBrain as distinct systems or technologies in its ranking-systems documentation. Their complete architectural relationship is not public. SEOs sometimes conflate neural matching with BERT, but they serve different described functions.
[Practitioner implication] Use accurate, recognisable terminology because it improves reader clarity and reduces ambiguity about the subject. Google confirms that neural matching can connect representations of concepts beyond exact wording, but it does not publish a terminology-density or neural-matching optimisation score.
RankBrain: Relating Words to Concepts
What Google Has Confirmed
[Google-confirmed] Google describes RankBrain as an AI system that helps it understand how words relate to concepts. This allows Search to return relevant content even when a page does not contain every exact word used in a query. Google introduced RankBrain in 2015. (Google Ranking Systems Guide)
What Google Has Not Confirmed
[Unknown] Google has not documented RankBrain as a system that learns from:
- click-through rates
- dwell time or time on page
- pogo-sticking (users returning to the SERP after clicking a result)
- page depth
- return visits
- bounce rates
Google’s current ranking-systems guide does not describe RankBrain as a behavioural-learning feedback loop of this kind. Describing RankBrain as an engagement-optimisation system is a practitioner theory, not a documented Google mechanism.
The Pogo-Sticking Myth
The claim that RankBrain learns which results satisfy queries by monitoring whether users return to the SERP after clicking has circulated widely in SEO since 2016. It is plausible as a machine-learning design — but Google has not confirmed it as how RankBrain works. Google’s documented description of RankBrain is that it helps understand how words relate to concepts. That is a narrower and more specific statement than “it learns from engagement behaviour.”
[Practitioner implication] Write pages that satisfy the query’s meaning and use clear, recognisable terminology. Do not attempt to optimise a supposed RankBrain engagement score, dwell-time target, or pogo-sticking rate.
BERT: Understanding How Words Express Meaning Together
Transformer Background
[General NLP concept] BERT — Bidirectional Encoder Representations from Transformers — is a transformer model published by Google researchers in 2018. Transformers process sequences of text using self-attention, which allows the model to represent the relationship between tokens across the full sequence simultaneously — not just adjacent words. This is the “bidirectional” in BERT: the model is pretrained by jointly conditioning on both left and right context, producing representations that account for surrounding language in both directions. (arXiv — BERT: Pre-training of Deep Bidirectional Transformers for Language Understanding)
Self-attention and bidirectionality are properties of how the transformer architecture works in principle. They do not prove that Google Search runs BERT on every indexed page, applies it at a specific stage, or uses it in a particular way. The capability and the deployment are separate questions.
What Google Has Confirmed About BERT in Search
[Google-confirmed] Google integrated BERT into Search in October 2019, describing it as “one of the biggest leaps forward in the history of Search.” Google’s announcement described BERT’s role as helping to understand searches — particularly how combinations of words express context and intent — and noted its application to ranking and featured snippets. (blog.google — Understanding searches better than ever before)
Google’s current ranking-systems guide describes BERT as a system that helps understand how combinations of words express meaning and intent. Google has not published a complete architecture diagram of BERT’s integration into Search, or confirmation that BERT comprehensively processes indexed documents as a document-understanding pipeline.
What BERT Means for Queries
[Google-confirmed] BERT’s most clearly documented function is query understanding. Google’s 2019 announcement cited the query “can you get medicine for someone pharmacy” — before BERT, Google returned results about pharmacies; with BERT, it understood the query was about picking up a prescription for someone else. (blog.google — Understanding searches better than ever before)
This demonstrates BERT improving query interpretation for conversational queries where word order and context affect meaning.
Grammar Is Not a BERT Ranking Signal
[Unknown] Google does not publish a grammar score or confirm that grammatically polished prose ranks more effectively because transformers parse it better. Write clearly because readers need to understand the content. Clear sentence structure reduces ambiguity. That is the justification — not a special BERT writing standard.
No BERT Optimisation Checklist
[Practitioner implication] BERT does not create a list of preferred keywords, a recommended sentence structure, or a targetable coherence score. Write accurate content in natural language. Do not attempt to insert context words or synonyms to “trigger BERT.” The model processes whatever language you write — the goal is to write language that is clear, accurate, and directly relevant to the query’s intent.
MUM: Specific Applications, Not General Ranking
What Google Announced
[Google-confirmed] Google announced MUM — Multitask Unified Model — in 2021. Google described MUM as a model designed for complex, multi-aspect information tasks, trained across 75 languages. Google described its ability to understand information across text and images, with broader multimodal possibilities discussed as future potential. Google described MUM as “1,000 times more powerful than BERT” without defining that statement as a parameter-count comparison or a ranking-quality multiplier. (blog.google — Introducing MUM)
MUM’s Role in Search: Not General Ranking
[Google-confirmed] Google’s ranking-systems guide states that MUM is used for specific applications in Search and is not used for general ranking. This directly contradicts the assumption underlying most “MUM optimisation” guidance — that writing comprehensively across multiple formats or languages improves rankings because MUM evaluates it. (Google Ranking Systems Guide)
Multimodal Scope
[Google-confirmed] Google’s 2021 announcement described MUM as capable of understanding information across text and images. Broader modalities were discussed as future potential, not confirmed deployment. Do not assume that every video or audio file on a page is processed by MUM for ranking purposes. (blog.google — Introducing MUM)
No MUM Optimisation Checklist
[Unknown] There is no documented MUM content requirement. Google does not publish a framework for “making content MUM-ready.” Create media types — images, video, documents — when they genuinely serve the user’s task.
[Practitioner implication] MUM’s confirmed role does not change the foundational SEO approach: satisfy the query’s task, use accurate information, structure content so readers can find what they need.
What Google Publicly Explains About Processing Pages
The prior version of this article described a five-stage page-processing pipeline: entity extraction, topic classification, intent inference, quality assessment, and semantic embedding — “each step using transformer-based NLP models.” Google has not published this pipeline.
[Google-confirmed] Google publicly describes Search at a high level as: crawling, indexing, and serving search results. (Google — How Search Works)
Google confirms that language-understanding systems — including neural matching, RankBrain, and BERT — help interpret concepts and intent. It has not published a universal sequence showing exactly how every page is processed, which models run on which documents, or how a page is converted to any specific representation.
[Unknown] Whether every indexed page receives a single vector embedding, how many representations are stored, and how those representations interact with other ranking signals are not publicly disclosed.
Safe SEO Implications
These recommendations are justified through reader clarity, intent satisfaction, and evidence — not through undocumented model behaviour.
Use recognisable terminology for your subject. Neural matching and BERT help Google interpret language and concepts. Using accurate, standard terminology for the subject you cover helps readers understand the content and reduces ambiguity. This is not a claim that correct terminology triggers a transformer relevance score.
Answer the query’s actual task. Write pages that address what the searcher is trying to accomplish. For informational queries, that means providing accurate information. For transactional queries, it means making the relevant action clear.
Structure direct answers near descriptive headings. A descriptive heading followed by a direct answer can improve reader comprehension and make the passage easier to quote or summarise. Use this structure where it genuinely fits the information; do not treat it as a guaranteed featured-snippet or AI-citation formula.
Support consequential claims with reliable evidence. This is especially important for health, financial, legal and other high-stakes topics. Reliable sourcing improves reader trust and editorial accountability. E-E-A-T comes from Google’s quality-evaluation framework, while GEO citation readiness is an editorial framework used here to assess whether claims can be extracted and attributed safely. Neither is a disclosed numerical ranking score.
Separate distinct intents into separate pages. A page trying to serve multiple incompatible intents does none of them well. This is an information-architecture decision, not an NLP-targeting decision.
How to Evaluate an NLP SEO Claim
Use this when reviewing vendor tools, agency proposals or SEO content recommendations that reference BERT, MUM, RankBrain or Google’s NLP systems.
| Claim | Decision |
|---|---|
| “We improve your Google BERT score” | Reject unless clearly identified as the vendor’s own proxy metric |
| “Increase dwell time to improve RankBrain” | Unsupported — Google has not confirmed this mechanism |
| “Add semantic terms to target MUM” | Unsupported — MUM is not used for general ranking |
| “Use accurate concepts needed for the search task” | Reasonable editorial practice |
| “Third-party embeddings replicate Google’s index” | Treat as an analytical model only; not Google’s architecture |
| “Question-and-answer headings guarantee AI citations” | Unsupported — no guaranteed citation formula exists |
| “Use clear answers and evidence for readers” | Sound information design |
| “Create multimedia because MUM rewards it” | Create it only when it serves the user’s actual task |
| “Optimise your semantic salience score” | Third-party metric; not a disclosed Google ranking input |
Myths and Unsupported Metrics
| Claim | Evidence status | Correct position |
|---|---|---|
| RankBrain learns from dwell time and pogo-sticking | [Unknown] — not in Google’s documentation | Google describes RankBrain as relating words to concepts |
| BERT evaluates semantic coherence instead of keywords | [Unknown] — false either/or | Keywords remain descriptive signals; BERT adds context understanding |
| MUM rewards comprehensive, multi-format content | [Unknown] — Google says MUM is not for general ranking | Create formats that serve user need |
| Google converts every page into one semantic vector | [Unknown] — not published | Google uses representations of concepts; the exact architecture is not disclosed |
| Grammatically polished prose ranks better because BERT parses it better | [Unknown] — no Google documentation supports this | Write clearly for readers |
| Precise terminology signals domain expertise to NLP models | [Unknown] — no confirmed NLP expertise signal | Use precise terminology for reader clarity |
| “Semantic completeness” is evaluated by Google’s NLP models | [Unknown] — Google has not published a completeness score | Cover sub-topics that serve the reader’s actual task |
| BERT and MUM generate featured snippets and AI Overview answers | [Partially confirmed / overstated] — Google confirmed BERT for featured snippets in 2019; AI Overview architecture is not the same | Follow foundational SEO for all Search features |
| Optimise for your “BERT score” using third-party NLP tools | [Unknown] — third-party NLP scores are not Google’s scores | Third-party tools model language; they do not replicate Google’s ranking system |
NLP Systems and AI Overviews
[Google-confirmed] Google said BERT was applied to featured snippets in 2019 to better understand the language in queries and documents. That does not establish that BERT or MUM currently generates every featured snippet or AI Overview. (blog.google — Understanding searches better than ever before)
AI Overviews and AI Mode are generative AI features grounded in Google Search systems and indexed web content. Google does not publish a complete model-by-model architecture or a universal citation-selection formula. The shift toward AI search has changed how results are presented, but the underlying SEO principles remain consistent with what Google confirms about its ranking systems.
Understanding how LLMs retrieve and cite information is separately useful for GEO — but that process operates on top of Search’s foundational quality and relevance signals, not instead of them.
Google’s current guidance says foundational SEO practices remain relevant for AI features — creating helpful, accurate content that demonstrates expertise. Google does not instruct SEOs to target BERT or MUM specifically for AI Overview inclusion. (Google — AI Optimisation Guide)
[Practitioner implication] For AI Overviews, apply the same practices that apply to Search generally: accurate information, clear structure, supported claims, and content that directly satisfies the query’s task.
What Not to Measure
The following are not Google-disclosed ranking metrics:
- keyword density as a BERT target
- “BERT score” from any third-party tool presented as Google’s evaluation
- NLP salience score as a confirmed ranking input
- semantic-term count as a completeness measure
- dwell time as a RankBrain performance metric
- pogo-sticking rate as a ranking signal
- MUM readiness score
- embedding similarity from a third-party tool presented as Google’s similarity score
Third-party tools that model these concepts are building their own proxies. Some may correlate with ranking outcomes. None reproduce Google’s internal evaluation.
Evidence Boundary
The following is what Google has confirmed about the systems covered in this article:
Neural matching: Helps Google match representations of concepts in queries and pages. (Google Ranking Systems Guide)
RankBrain: An AI system that helps Google understand how words relate to concepts. (Google Ranking Systems Guide)
BERT: Helps Google understand how combinations of words express meaning and intent. Applied to ranking and featured snippets, announced October 2019. (blog.google — Understanding searches better than ever before)
MUM: Trained across 75 languages. Capable of understanding text and images. Used for specific Search applications. Not used for general ranking. (blog.google — Introducing MUM; Google Ranking Systems Guide)
Page processing: Google publicly describes Search as crawling, indexing, and serving results. The specific NLP pipeline, document representation, and model deployment architecture are not publicly disclosed. (Google — How Search Works)
The architecture between these confirmed facts — how the systems interact, which models run on which documents, in what sequence, and at what cost — is not published by Google. Where SEO articles fill that gap with plausible machine-learning design, they are presenting inference as fact.
Primary Sources
- Google Ranking Systems Guide: developers.google.com/search/docs/appearance/ranking-systems-guide
- Google BERT announcement (2019): blog.google — Understanding searches better than ever before
- BERT research paper: arXiv — BERT: Pre-training of Deep Bidirectional Transformers for Language Understanding
- Google MUM announcement (2021): blog.google — Introducing MUM
- How Search Works: developers.google.com/search/docs/fundamentals/how-search-works
- AI Optimisation Guide: developers.google.com/search/docs/fundamentals/ai-optimization-guide
- Search Essentials: developers.google.com/search/docs/essentials
Summary
Google uses several AI systems within Search to interpret language, concepts and intent. Neural matching is explicitly described as matching representations of concepts in queries and pages, while RankBrain, BERT and MUM have distinct documented roles and undisclosed implementation details. Google describes RankBrain as relating words to concepts, BERT as understanding how combinations of words express meaning and intent, and MUM as a model used for specific Search applications rather than general ranking.
These confirmed descriptions are narrower than what circulates in most SEO coverage. RankBrain is not a documented dwell-time algorithm. BERT is not a confirmed document-processing pipeline that runs on every indexed page. MUM is not a general ranking system that rewards comprehensive content.
No separate BERT, RankBrain, or MUM optimisation checklist follows from what Google has confirmed. The correct approach is to satisfy the reader’s intent with accurate, clearly structured content — the same foundational principle that applies to semantic SEO generally. Where the architecture is unknown, treat it as unknown rather than converting a plausible machine-learning design into a ranking rule.
The next article in this cluster examines Google’s Knowledge Graph, entity relationships and how websites can provide clear, consistent and verifiable entity information without treating structured data as an authority shortcut.
Next: The Knowledge Graph and Search: How Google Stores Meaning, Not Just Words
TL;DR Google uses several AI systems within Search, including neural matching, RankBrain, BERT and MUM. Google describes neural matching as matching representations of concepts in…
