When to open this: If you do SEO / run content sites / want to understand 10 years of Google ranking evolution / want to know exactly what the official original sources behind these 5 abbreviations say. This wiki compares Google's official announcements for the 5 milestones,and is the "constitution" that anyone doing SEO must read first.
One-sentence core:RankBrain (2015, words → concepts) → Neural Matching (2018, fuzzy matching) → BERT (2019, word-to-word relationships) → MUM (2021, multimodal + multilingual) → HCU (2022, site-level "written for people") → AI Overviews (2024, generative search). Google now runs hundreds of models working together, not replacing one with another — any SEO advice that says "optimize specifically for algorithm X" is wrong.
0. Quick reference table for the 5 major milestones
| Milestone | Year | Author | Core | Impact scope |
|---|---|---|---|---|
| RankBrain | 2015-10 | (Pandu Nayak 2022 overview) | First deep learning ranking system / words → concepts | "one of the major AI systems", still a core system today |
| Neural Matching | 2018 | (Pandu Nayak 2022 overview) | Neural network fuzzy matching between queries and documents | "critical part of retrieval" |
| BERT | 2019-10 | Pandu Nayak | Transformer / word-to-word relationships within sentences | 10% of US English queries / now "almost every English query" |
| MUM | 2021-05 | Pandu Nayak | T5 / 1000x BERT / 75 languages / multimodal / capable of generation | "join the chorus", works in coordination with other models |
| HCU | 2022-08 | Danny Sullivan | Site-level ML classifier / written for people vs written for search engines | Launched first across the entire English-speaking world, later expanded |
| AI Overviews | 2024-05 | Liz Reid | custom Gemini / generative answers + multi-step + planning + video | Launched first in the US, expanded in 2024 |
1. RankBrain (2015) — Google's first deep learning
The 5 Major AI Algorithm Milestones in Google Search (RankBrain/BERT/MUM/HCU/AIO) source:2022-02-how-ai-powers-search.md §RankBrain
1.1 Official definition
RankBrain launched in 2015 as "the first deep learning system deployed in Search." It represented a breakthrough because it helped understand "how words relate to concepts."
1.2 Classic example (official)
Query:"what's the title of the consumer at the highest level of a food chain"
RankBrain's understanding:Relates to animals rather than human consumers → returns "apex predator"
1.3 Still a core system today
"RankBrain remains one of the major AI systems powering Search today."
1.4 No standalone 2015 announcement
⚠️ Important: Google did not publish a standalone 2015 RankBrain blog post. It was first disclosed in a Bloomberg interview with Jack Clark on 2015-10-26. The authoritative official retrospective is in Pandu Nayak's February 2022 overview.
2. Neural Matching (2018) — Fuzzy matching
The 5 Major AI Algorithm Milestones in Google Search (RankBrain/BERT/MUM/HCU/AIO) source:2022-02-how-ai-powers-search.md §Neural Matching
2.1 Official definition
Neural matching enabled Search to "understand how queries relate to pages" using neural networks. This system analyzes "fuzzier representations of concepts" rather than just keywords.
2.2 Example
Query: "insights how to manage a green"
Neural Matching's understanding: can match to a management guide that describes personality using colors(green-personality type content), even if that document does not contain the word "insights".
2.3 Status
"Neural matching is such a critical part of how we retrieve relevant documents."
3. BERT (2019) — Transformer enters search
The 5 Major AI Algorithm Milestones in Google Search (RankBrain/BERT/MUM/HCU/AIO) source:2019-10-bert-understanding-searches.md
3.1 Key proposition
"The biggest leap forward in the past five years, and one of the biggest leaps forward in the history of Search." —— Pandu Nayak
3.2 Core technology
- BERT = Bidirectional Encoder Representations from Transformers
- Transformer: processes the relationship between a word and every other word in the sentence, rather than processing one by one in a single direction
- BERT looks at the surrounding context of a word to understand query intent
- First time using Cloud TPUs to serve search results(the model was too complex for traditional hardware)
3.3 Key data
| Dimension | Number |
|---|---|
| Share of ranking queries affected(at launch) | 10% of US English queries |
| Impact scope(as of the 2022 overview) | Almost every English query |
| Improved language in featured snippets | Korean / Hindi / Portuguese, etc. |
| 15% of queries have never been seen before(background data) | Continues to be used today |
3.4 Classic examples (all given in Pandu Nayak's original text)
| Query | Difficulty | BERT's solution |
|---|---|---|
| "2019 brazil traveler to usa need a visa" | "to" determines direction | Understands it means a Brazilian person traveling to the US (not the other way around) |
| "do estheticians stand a lot at work" | "stand" can easily match stand-alone | Understands "stand" = the physical intensity of the job |
| "can you get medicine for someone pharmacy" | "for someone" is key | Understands it means picking up medicine for another person |
| "parking on a hill with no curb" | Old systems ignored "no" | Understands "no" is critical |
| "book the young adult" | "adult" tends to surface 18+ content | Understands "young adult" refers to the YA literature category |
3.5 Google itself admits BERT is not perfect
"If you search for 'what state is south of Nebraska,' BERT's best guess is a community called 'South Nebraska.' (Spoiler: it's Kansas.)" — original text by Pandu Nayak
4. MUM (2021) — Multimodal + multilingual
The 5 Major AI Algorithm Milestones in Google Search (RankBrain/BERT/MUM/HCU/AIO) source:2021-05-mum-ai-milestone.md
4.1 Key data (must memorize)
| Dimension | Number |
|---|---|
| MUM = Multitask Unified Model | — |
| Base model | T5 text-to-text framework |
| Capability multiplier | 1,000 times that of BERT |
| Languages | 75 languages |
| Number of tasks | Trained on multiple tasks simultaneously |
| Modalities | Text + images(to expand to video / audio in the future) |
| Generative capability | Not only understands, but also generates(Generative) |
4.2 Core scenario: complex tasks
User insight:
"We find that people issue eight queries on average for complex tasks" — an average complex task requires 8 queries
Example(used repeatedly throughout the full text):
"You've climbed Mt. Adams, and want to climb Mt. Fuji next autumn. If you asked a mountaineering expert, you would only need to ask 'what should I prepare differently?' to get the answer — including a height comparison between the two mountains, that Mt. Fuji in autumn is in rainy season so you need to bring a waterproof jacket, training advice, and so on."
4.3 Cross-lingual knowledge transfer
"Say there's really helpful information about Mt. Fuji written in Japanese; today, you probably won't find it if you don't search in Japanese. But MUM could transfer knowledge from sources across languages."
SEO implication:English content no longer holds exclusive weight — good information in content in Japanese / Chinese / French and other languages can feed back into English search via MUM.
4.4 Future multimodal scenarios
"Eventually, you might be able to take a photo of your hiking boots and ask, 'can I use these to hike Mt. Fuji?' MUM would understand the image and connect it with your question."
4.5 MUM works in coordination with other systems
"MUM will... join the chorus of systems that run together." — does not replace BERT/RankBrain, works in coordination with them
5. HCU (2022-08) — Site-level "written for people vs written for search engines"
The 5 Major AI Algorithm Milestones in Google Search (RankBrain/BERT/MUM/HCU/AIO) source:2022-08-helpful-content-update.md + technical detail version published simultaneously on the GSC blog
5.1 Key proposition
"Content that seems to have been primarily created for ranking well in search engines rather than to help or inform people."(definition of content targeted precisely)
5.2 Key mechanisms (must know for SEO)
- HCU = Helpful Content Update
- Site-level ML classifier(site-level, not per-page)
- Automatic determination, no manual review appeal channel
- Runs continuously, not a one-time update
5.3 Key categories
"The update particularly impacts online education, entertainment, shopping, and technology categories."
5.4 Relationship with AI content
Key understanding(see wiki-seo-in-llm-era.md §3 for details):
- HCU Does not target "whether it is AI-generated"
- HCU targets"written for people or written for ranking"(intent)
- March 2024 HCU merged into core ranking: from this point there is no independent "recovery date" for helpful content
5.5 Subsequent evolution (see GSC Blog: Core Updates and Algorithm Iteration Evolution (2010-2026) §HCU)
| Time | Action |
|---|---|
| 2022-08 | First launch of HCU, launched in English first |
| 2022-12 | Expanded to non-English languages |
| 2024-03 | HCU merged into core (no longer independent) |
| After September 2024 | "scaled content abuse" replaces the single HCU signal |
6. AI Overviews (2024-05) — Gemini enters search results
The 5 Major AI Algorithm Milestones in Google Search (RankBrain/BERT/MUM/HCU/AIO) source:2024-05-ai-overviews-launch.md
6.1 Key proposition
- Former SGE (Search Generative Experience) officially GA → AI Overviews
- Powered by:custom Gemini model(not an open API, but a search-specific fine-tuned version)
- Initial launch:Full rollout in the US(2024-05-14 Google I/O)
6.2 4 major new capabilities
| Capability | Example |
|---|---|
| Multi-step reasoning | Searching for "find yoga studios + reviews + new member discounts" in one go (originally required 3 queries) |
| Planning | Generating adjustable meal plans (one-click switch to vegetarian) |
| AI-organized results | Inspiration-type search: automatic categorized display for restaurants/movies/books/shopping |
| Video search | Upload a video and ask "what is this / how to fix it", no need to type a description |
6.3 SEO impact (synthesized from Shensi Circle + Wang Kai + newtype)
See wiki-seo-in-llm-era.md §6 for details:
- Not zero-click: Google emphasizes there will still be clicks
- Cited pages receive citations
- Optimization advice: make your page "the one that gets cited"
- 2025/05 Succeeding in AI Search 5 points: original / page experience / structure / multi-modal / trust signals
7. Common patterns across the 5 milestones (practical SEO implications)
7.1 Not "replacing one with another", but a "coordinated chorus"
"Google's Search infrastructure runs hundreds of algorithms and machine learning models that work in combination rather than isolation." —— Pandu Nayak 2022-02
SEO implication:Any advice to "optimize specifically for algorithm X" is wrong(such as "optimize titles for BERT").
7.2 All are a progression of "semantic understanding > keyword matching"
| Era | Primary method |
|---|---|
| pre-2015 | Keyword matching + link analysis (PageRank) |
| After RankBrain | Words → concepts |
| After Neural Matching | Query → document fuzzy matching |
| After BERT | Word-to-word relationships within sentences |
| After MUM | Cross-task + cross-language + cross-modal |
| After AIO | Generative answers + reasoning |
SEO implication:"Keyword density / keyword stuffing" has been dead for 10 years.
7.3 All are ML, not rules
All 5 are ML models, which means:
- Cannot be reverse-engineered
- Cannot be appealed
- You can only improve content quality, you cannot "hack the algorithm"
7.4 Site-level vs page-level (unique to HCU)
HCU is the only site-level determination — one piece of low-quality content can drag down the entire site.
SEO implication:Cleaning up low-quality pages on your site is more important than adding new good pages(when HCU is triggered).
7.5 Cross-language + Multimodal (starting with MUM)
After MUM:
- English content no longer holds exclusive weight
- Images / videos become signal sources
- The quality of "non-English" content such as Chinese / Japanese / Spanish directly affects global search
8. How to Use — Milestone Path by SEO Task
| If you want to | Primary milestone | Starting point |
|---|---|---|
| Understand how Google ranks now | §7 5 Commonalities + Collaborative chorus | Do not optimize specifically for a single algorithm |
| How content sites can avoid being hit by HCU | §5 Site-level ML + Write for people | Clean up low-quality content on site / Real human authors / First-hand content |
| Write long-tail content to capture traffic | §3 BERT + §4 MUM | Natural language + Cross-language transfer |
| Bet on AI Overviews traffic | §6 AIO 4 Core Capabilities + Succeeding | original + structure + multi-modal + trust |
| International content site strategy | §4 MUM 75 languages | High-quality non-English content can feed back |
| Multimodal content (image/video) optimization | §4 MUM + §6 AIO video | Add descriptions / Add structured data / Main body of video content |
General disciplines:
- It is not "specifically targeting algorithm X" —— Hundreds of models collaborate, optimize for quality
- HCU is site-level —— Cleaning up low-quality content > Adding new good pages
- Keyword stuffing has been dead for 10 years —— Write in natural language + From a real person's perspective
- Cross-language feedback after MUM —— Non-English content also matters
- AIO is not zero-click —— Compete to be "the one that gets cited"
9. Reference List
9.1 5 Official Original Articles (all sources for this wiki)
- 2019-10 BERT — Understanding Searches Better Than Ever Before —— Pandu Nayak
- 2021-05 MUM — A New AI Milestone for Understanding Information —— Pandu Nayak
- 2022-02 How AI Powers Great Search Results —— Pandu Nayak(The only place that covers all 4 of RankBrain + Neural Matching + BERT + MUM together)
- 2022-08 More content by people, for people in Search —— Danny Sullivan
- 2024-05 Generative AI in Search —— Liz Reid
9.2 GSC Synchronized Technical Articles
sources/reference/google-search-central/blog/2022/08/helpful-content-update.en.md—— HCU technical details version
9.3 Related wikis
- SEO × LLM Era (cross-author topic synthesis) —— SEO × LLM Era synthesis (this wiki has been merged into it)
- GSC Blog: Evolution of AI Search and Google's Stance on AI Content (2015-2026) —— Google's stance on AI content (this wiki serves as its "unread" supplement)
- GSC Blog: Core Updates and Algorithm Iteration Evolution (2010-2026) —— Core Updates evolution (subsequent trajectory of HCU)
- GSC Blog: Quality Policies and E-E-A-T Evolution (Webmaster Guidelines → Search Essentials) —— E-E-A-T quality policies
9.4 Notes without independent blog posts
- RankBrain 2015:Google did not publish an independent 2015 blog post. The earliest disclosure was the 2015-10-26 Bloomberg interview with Jack Clark. The official retrospective is in the February 2022 overview by Pandu Nayak cited in §1 of this wiki.
- Neural Matching 2018:Similar situation, mainly disclosed in John Mueller's September 2018 Twitter post + the 2022 overview.
Sources and related materials
- https://blog.google/products/search/search-language-understanding-bert/
- https://blog.google/products/search/introducing-mum/
- https://blog.google/products-and-platforms/products/search/how-ai-powers-great-search-results/
- https://blog.google/products/search/more-content-by-people-for-people-in-search/
- https://blog.google/products/search/generative-ai-google-search-may-2024/
- https://developers.google.com/search/blog/2022/08/helpful-content-update?hl=en
- SEO × LLM Era (cross-author topic synthesis)
- GSC Blog: Evolution of AI Search and Google's Stance on AI Content (2015-2026)
- GSC Blog: Core Updates and Algorithm Iteration Evolution (2010-2026)
- GSC Blog: Quality Policies and E-E-A-T Evolution (Webmaster Guidelines → Search Essentials)