How exactly does Google's AI ranking work? RankBrain, BERT, MUM - I'm confused

James, I’ll break this down. These systems are complementary, not replacements.

The ensemble approach:

Google’s ranking uses multiple AI systems working together. They trigger at different times and in different combinations depending on query type.

How they work together:

1. RankBrain handles the 15% of queries Google has never seen
2. BERT understands the meaning of your specific query
3. Neural Matching finds pages that match the concepts (not just keywords)
4. MUM handles complex, multimodal tasks

Key insight:

Google asks: “Which page best answers this user’s intent?” Not: “Which page has the most keyword matches?”

Your observation about ranking lower for broader queries is likely RankBrain + BERT working together - they understand users want different content for broad queries than what you’re providing.

Technical explanation of how RankBrain measures quality:

RankBrain monitors two key signals:

1. Click-through rate (CTR) - Do users click your result?
2. Dwell time - How long do they stay?

The feedback loop:

User searches → Sees results → Clicks result → Either:
  - Stays (positive signal) → Ranking boost
  - Returns quickly (pogo-sticking) → Ranking drop

Research findings:

Google tested RankBrain against human engineers to identify the best page for searches. RankBrain outperformed humans by 10%.

What this means for you:

Your title tag is now more important than ever. It needs to earn the click AND your content needs to satisfy the search intent.

Let me address the “is keyword optimization dead” question.

Short answer: Traditional keyword optimization is dead. Semantic optimization is essential.

What RankBrain killed:

The practice of creating separate pages for minor keyword variations:

• “best keyword research tool”
• “best tool for keyword research”
• “keyword research tool best”

RankBrain understands these are identical queries. Google shows nearly identical results for all of them.

What works now:

1. One comprehensive page per topic
2. Semantic coverage - related terms and concepts
3. Topic clusters - supporting pages that link to pillar content
4. Entity optimization - cover all aspects of the topic

Example:

Old approach (5 pages):

• best-crm-software.html
• top-crm-tools.html
• crm-software-comparison.html
• best-crm-for-business.html
• crm-tool-reviews.html

New approach (1 comprehensive page):

• best-crm-software.html (covers all angles, 3000+ words)
• Supporting pages link to it for specific use cases

The single comprehensive page ranks for thousands of keyword variations automatically.

Neural Matching deserves more attention here.

What Neural Matching does:

It understands broader representations of concepts, not just keywords.

Example query: “insights how to manage a green”

Traditional search: Struggles because words don’t match any pages

Neural Matching: Understands this is about the “green” personality type from color-based personality guides, returns management tips for that personality type

Why this matters:

Your content can rank for queries that don’t contain your exact keywords if:

1. The concepts match
2. Your content addresses the underlying intent
3. You cover the topic comprehensively

Optimization strategy:

Think about all the ways people might ask about your topic:

• Direct questions
• Indirect references
• Related concepts
• Adjacent topics

Cover them all, and Neural Matching will connect the dots.

Let’s talk about MUM - the future of Google search.

MUM capabilities:

• 1000x more powerful than BERT
• Can understand and generate language
• Trained on 75 languages simultaneously
• Multimodal (text, images, potentially video)

Current MUM applications:

• COVID-19 vaccine information
• Google Lens visual + text searches
• Not yet used for general ranking

What to expect:

MUM will eventually power:

• Complex multi-turn queries
• Cross-language search (search in English, find results in Japanese)
• Image + text combined queries
• Deeper reasoning chains

Strategic implication:

Future-proof your content by:

1. Including visual elements (images, diagrams)
2. Covering topics comprehensively
3. Building topical authority (not just single-page optimization)
4. Thinking globally (consistent messaging across languages)

How AI ranking affects local search specifically:

Location + intent understanding:

Google’s AI systems understand that “football” means different things in different places:

• Chicago → American football, Bears
• London → Soccer, Premier League

Local relevance signals AI evaluates:

For local businesses:

Don’t just optimize for keywords. Optimize for:

• Your specific location context
• The problems local users are trying to solve
• The language patterns your local audience uses

RankBrain and BERT understand local context. Use it to your advantage.

Enterprise perspective on AI ranking:

The challenge:

Large sites with thousands of pages can’t optimize each page individually. We need scalable strategies.

Our approach:

1. Topic architecture - Organize content into clear hierarchies
2. Template optimization - Ensure templates include proper semantic elements
3. Automated quality signals - Author attribution, publish dates, structured data
4. Internal linking - Let Google understand relationships

What AI ranking means for enterprise:

Results:

After restructuring around topics instead of keywords:

• 47% increase in long-tail traffic
• 23% better engagement metrics
• Featured snippet capture up 180%

AI ranking rewards sites that are organized around topics, not keywords.

The CRO angle on AI ranking:

RankBrain’s engagement signals create a feedback loop:

Good content → Users stay → Rankings improve → More traffic → More data → Better rankings

The opposite is also true:

Poor match → Users bounce → Rankings drop → Less traffic → Worse rankings

Practical improvements:

1. Above-the-fold answer - Give users what they need immediately
2. Scannable format - Headers, bullets, short paragraphs
3. Visual hierarchy - Guide eyes to key information
4. Clear next steps - What should users do after reading?

Our test results:

Page with answer buried in paragraph 3:

• Avg time on page: 23 seconds
• Bounce rate: 78%

Same content with answer in first paragraph:

• Avg time on page: 3:47
• Bounce rate: 34%

RankBrain noticed. Rankings improved by 12 positions over 6 weeks.

Don’t forget: Google AI ranking ≠ AI search platforms.

Google’s AI ranking:

• Determines which pages rank in traditional search
• Uses RankBrain, BERT, Neural Matching, MUM
• Still shows list of links (mostly)

AI search platforms (ChatGPT, Perplexity, Claude):

• Generate answers, not rankings
• May cite sources inline
• Different optimization strategies

The overlap:

Content that ranks well in Google AI ranking is often cited by AI platforms too. But not always.

Monitor both:

Tools like Am I Cited let you track visibility across:

• Traditional Google rankings
• Google AI Overviews
• ChatGPT citations
• Perplexity citations

Your Google optimization and AI optimization strategies should complement each other, not compete.

This thread clarified a lot. Here’s my updated understanding:

How Google’s AI systems work together:

1. RankBrain - Handles new queries, measures engagement signals (CTR, dwell time)
2. BERT - Understands the meaning of queries, especially small contextual words
3. Neural Matching - Connects concepts across queries and content
4. MUM - Future multimodal understanding (limited current use)

Key shifts in SEO strategy:

From → To:

• Keywords → Intent
• Multiple thin pages → One comprehensive page
• Keyword density → Semantic coverage
• Exact match → Concept matching
• Page optimization → Topic clusters

Practical changes I’m making:

1. Consolidate similar pages into comprehensive resources
2. Optimize titles for CTR (RankBrain cares about clicks)
3. Answer questions directly in first paragraph (engagement signals)
4. Cover topics comprehensively (Neural Matching connects concepts)
5. Match exact user language patterns (BERT understands context)

The big insight:

Google’s AI is trying to understand what users actually want and find pages that satisfy that intent. Optimize for user satisfaction, and the AI will reward you.

Thanks everyone for breaking down the complexity into actionable insights.