MUM (Multitask Unified Model)

How MUM differs from BERT

	BERT	MUM
Launched	October 2019	May 2021 announcement
Capability	Understanding query context	Understanding + generation across modalities
Languages	English-first	75 languages, transferable knowledge
Modalities	Text only	Text + images + video
Use case	General query understanding	Complex multi-step queries

MUM is bigger and more sophisticated, but Google uses it selectively for query types where the additional capability pays off.

Example MUM-style query

“I’ve hiked Mt. Adams and now I want to hike Mt. Fuji next fall. What should I do differently to prepare?”

This is a multi-faceted query requiring:

• Understanding both mountains
• Understanding “next fall” as a time reference
• Comparing differences (terrain, weather, altitude)
• Generating actionable advice

Before MUM, this would require 8-10 separate searches. MUM-class models can synthesize an answer in one pass.

MUM and AI Overviews

When Google’s AI Overviews answer complex queries, MUM (or its descendants) handle the:

• Cross-language information retrieval — pulling from non-English sources to answer English queries
• Multi-source synthesis — combining information from many web pages into one answer
• Multimodal understanding — interpreting query intent that includes images

The AI Overview infrastructure isn’t pure MUM — it combines MUM-style models with retrieval systems and Gemini-class generation. But MUM was the breakthrough that made the AI Overview UX feasible.

SEO implications

MUM doesn’t directly change ranking signals (it operates at the query-understanding and answer-generation layer, not the ranking layer). But it changes what optimization needs to address:

• Cross-language content can be retrieved for any-language queries — a high-quality English article on a topic can be cited in answers given to Japanese users
• Comprehensive coverage of a topic wins — MUM-class synthesis prefers sources that cover multiple facets of a question
• Multimodal content adds surface area — videos, images, charts all become potentially-retrievable sources

Optimization in a MUM-aware world

Don’t optimize FOR MUM specifically. Optimize for:

• Comprehensive topical depth (cover the full question, not just keywords)
• Cross-language accessibility (proper hreflang, locale-specific content)
• Multimodal content (relevant images, embedded video, charts)
• Clear answer patterns (Quick Answer Blocks, FAQs) that synthesis systems can extract

Resocial perspective

MUM is part of why we built our agent workforce around topical-depth-first content strategy. The brands that win in an MUM-mediated future are the ones with comprehensive, multi-faceted coverage of their domain — not the ones with keyword-targeted but shallow content. Our content strategy framework and AI Search optimization both reflect this.