Google has long ranked through an interplay of several AI systems rather than a single algorithm – and thanks to the DOJ trial (2023) and the API leak (2024), it is now documented which signals truly count: from semantic retrieval to real click data.
- AI is the core: RankBrain, Neural Matching, DeepRank, and NavBoost work in a cascading pipeline – each system has a specific role between query interpretation and final re-ranking. Since the DOJ trial in 2023, we know for the first time how this architecture actually interacts.
- User signals are officially confirmed: NavBoost uses 13 months of click data with goodClicks, badClicks, and unicornClicks as ranking signals. CTR and dwell time are no longer myths – proven by Pandu Nayak’s sworn testimony and the 2024 API leak.
- Gemini 3 shifts the playing field: With AI Mode, Query Fan-Out, and Generative UI, “citation in the AI answer” becomes the new SEO metric alongside classic rankings. Those who strengthen E-E-A-T, semantic depth, and structured data today position themselves for both worlds.
Google has integrated more artificial intelligence into its search algorithm in the last ten years than in the entire period before. What began in 2015 with RankBrain is now a complex interplay of different machine learning systems that jointly determine which content you see in search results.
Understanding these systems isn’t academic curiosity – it’s essential for modern SEO. Anyone who doesn’t understand how RankBrain interprets long-tail queries, how DeepRank evaluates documents at a semantic level, or how NavBoost translates user satisfaction into ranking signals, is optimizing blindfolded.
From my work as a Product Developer at iGaming.com, I see daily how radically the ranking landscape has shifted. Pages that were stable at position 1 three years ago are losing visibility – not because of an update, but because the AI systems have learned to read better signals. The DOJ antitrust trial in 2023 and the API leak of 2024 gave us our first genuine look behind the curtain.
This cornerstone article is part of my comprehensive guide to the Google Search Algorithm: From Crawling to Ranking. While the article on Semantic Search & Knowledge Graph explains how Google understands meaning and entities, here we dive deep into the concrete AI systems that turn that understanding into measurable rankings.
RankBrain: Google’s First Machine Learning System
RankBrain was launched in October 2015 and was Google’s first use of deep learning at the core of its search algorithm. According to Google, it was already the “third most important ranking signal” at the time.
What RankBrain Actually Does
RankBrain solves a specific problem: approximately 15% of all daily search queries are entirely new – combinations Google has never seen before. Before RankBrain, Google could only guess at such queries based on keyword matching. RankBrain understands the relationship between words and concepts instead.
An example from Google itself: when searching “what is the consumer at the top of a food chain called?”, RankBrain recognizes this is about the concept of a food chain (not human consumers) and that “top of the food chain” points to an “apex predator” – even though those words never appear in the query.
The Technology Behind RankBrain
RankBrain is based on a feed-forward neural network, a simpler architecture than later transformer models. It uses word embeddings (similar to Word2Vec) to translate words into a mathematical vector space. In this space, semantically similar words are close together. A detailed explanation is available in Google’s official blog post on AI in search.
RankBrain in 2026: Still Relevant?
In the DOJ trial 2023, Pandu Nayak confirmed that RankBrain is still active but is increasingly being superseded by DeepRank:
“DeepRank is taking on more and more of that capability now. At the time that this was done, maybe they were more complementary, but over time they are becoming less complementary.”
– Pandu Nayak, Google VP of Search, DOJ Trial October 2023
RankBrain remains especially valuable for its speed: it’s “cheaper to train” than transformer models and can therefore run for all queries, while DeepRank is only used for the final top results.
Neural Matching & RankEmbed: Semantic Retrieval
Neural Matching was introduced in 2018 and operates in a different phase than RankBrain: it’s part of document retrieval, the pre-selection of potentially relevant documents from the index.
The Problem Neural Matching Solves
Imagine someone searches for “tips for tying shoelaces.” Classic keyword matching would find pages containing “shoelaces” and “tying.” But what about an excellent page about “how to knot your laces properly”? Without Neural Matching, it might never make it into the candidate list for ranking.
RankEmbed: The Technical Name
In the DOJ trial, it was revealed that Neural Matching is internally called RankEmbed. Pandu Nayak explained:
“RankEmbed identifies a few more documents to add to those identified by the traditional retrieval. […] RankEmbed evaluates semantic similarity between query and document.”
– Pandu Nayak, DOJ Trial 2023
RankEmbed supplements classic keyword-based retrieval with semantic retrieval. It’s a dual-encoder model that encodes query and document separately, then compares them via dot product or cosine similarity.
Behind this semantic retrieval lies the same vector mechanic as any modern meaning-based search: text becomes embeddings, similarity is measured via cosine similarity. How that works in detail – including SBERT and a worked example – I broke down here: Embeddings, Vectors and SBERT.
BERT & DeepRank: Contextual Language Understanding for Ranking
BERT (Bidirectional Encoder Representations from Transformers) was a quantum leap in language understanding in 2019. Unlike previous systems that analyzed words in isolation or only left-to-right, BERT examines bidirectional context – looking simultaneously at all words before and after a term.
DeepRank: BERT for Ranking
In the DOJ trial, it became clear: when BERT is used for ranking, it’s internally called DeepRank. This isn’t just a name difference – DeepRank was specifically trained for ranking relevance. More on how Google combines language understanding and ranking in my article How Google Evaluates AI Texts.
“DeepRank is BERT when BERT is used for ranking. […] DeepRank not only gives significant relevance gains, but also ties ranking more tightly to the broader field of language understanding.”
– From an internal Google document, cited in the DOJ trial
Language Understanding + World Knowledge
“You get a lot of world knowledge from the web. In search, you can get the world knowledge because you have an index and you retrieve documents, and those documents that you retrieve gives you world knowledge.”
– Pandu Nayak, DOJ Trial
NavBoost: How User Signals Influence Rankings
Perhaps the biggest revelation of the DOJ trial was NavBoost – a system running since 2005 that uses click data to refine rankings. For years, Google had denied that click rates influence ranking. NavBoost proves the opposite. A detailed analysis is available in the How Google Search Ranking Works article on Search Engine Land.
NavBoost collects and analyzes user interaction data from the last 13 months. Further details emerged in the May 2024 leak of the “Google Content Warehouse API.” SEO experts Rand Fishkin (SparkToro) and Mike King (iPullRank) decoded the NavBoost metrics. A detailed analysis is in Fishkin’s API Leak analysis.
| Signal | Meaning | Interpretation |
|---|---|---|
| goodClicks | Clicks with positive signals | Longer dwell time, no return to SERP |
| badClicks | Clicks with negative signals | Quick return to SERP (pogo-sticking) |
| lastLongestClicks | Most recent longest interactions | Current user satisfaction |
| unicornClicks | Exceptionally positive interactions | Strong satisfaction signal |
The Interplay: How the AI Systems Cooperate
A common misconception: these systems don’t work in isolation but in an orchestrated pipeline. Each system has its strength and handles a specific task. The technical foundations of how documents enter this pipeline are explained in my article on Crawling & Indexing.
The Ranking Pipeline Overview
When you enter a search query, it passes through multiple phases. First comes Query Understanding: RankBrain interprets the query, recognizes search intent and context, and translates unfamiliar long-tail queries into understandable concept patterns. Simultaneously, Retrieval starts – the inverted index delivers candidates via keyword matching, while RankEmbed supplements semantically matching documents without exact word matches. How this query processing and retrieval works in detail – including tokenization, query expansion, and the inverted index architecture – is explained in my separate article Query Processing in Google Search.
From the tens of thousands of retrieval candidates, coarse ranking (RankBrain + traditional signals like PageRank) filters down to a few hundred. DeepRank then evaluates these in fine ranking with deep language understanding – this is where positions 1-3 are decided. Finally, NavBoost together with Twiddlers adjusts the final positions based on user signals, freshness, and diversity. For certain queries, Gemini additionally generates an AI Overview from the top sources.
The Three Ranking Pillars: Topicality, Quality, and Popularity
The DOJ trial files and subsequent Remedial Opinion revealed a surprisingly clear architecture. An excellent analysis is provided by Shaun Anderson’s DOJ revelations analysis.
Topicality (T*) forms the foundation – does this document match the query at all? It builds on the “ABC signals”: Anchors (link texts), Body (page content), and Clicks (historical click data). Without a sufficient T* score, a page never reaches the next evaluation stages.
Quality (Q*) evaluates source trustworthiness – independent of the specific query. This largely static score draws from PageRank, spam scores, and human quality rater evaluations.
Popularity (P*) is the dynamic corrective. Here, NavBoost signals flow in: Chrome visit data, click behavior, dwell time, and incoming links. While T* and Q* are “theoretical” evaluations, P* measures reality.
| Phase | Variable | What Happens |
|---|---|---|
| Retrieval | T* | From billions of documents, the topically matching ones are pre-selected |
| Scoring | Q* | Candidates are weighted by trustworthiness and authority |
| Re-Ranking | P* | NavBoost sorts finally by real user behavior of the last 13 months |
The formula for rankings 2026: Relevance (your content answers the question) × Quality (you’re a trustworthy source) × Popularity (users are satisfied) = Rankings. All three factors must be right.
Gemini 3 & AI Mode: The New Era of Search
While RankBrain, BERT, and NavBoost determine classic ranking, Gemini 3 is changing search itself. Since November 2025, Gemini 3 Pro powers AI Mode, since December 2025, Gemini 3 Flash is the standard model for AI Mode worldwide. With the Gemini 3.1 Pro Update (February 2026) and the new Deep Think mode, Google has deepened agentic search capabilities further.
AI Mode is a fully AI-powered search experience where Gemini analyzes the query, performs multiple searches, and generates a tailored answer. The exact differences are explained in my article Google AI Mode vs. AI Overviews.
At Google I/O 2025, CEO Sundar Pichai announced 1.5 billion monthly users of AI Overviews and 400 million active Gemini users.
With Gemini, the SEO metric shifts from “position in the SERPs” to “citation in the AI answer.” The good news: Gemini still uses Google’s classic ranking systems as its foundation. E-E-A-T signals are more important than ever.
SEO Optimization for AI Ranking Systems
1. Semantic Content Architecture
Structure your content so AI systems recognize concepts and relationships. The Hub-and-Spoke model provides a proven architecture: topic clusters with pillar pages and linked sub-pages signal topical authority.
2. Answer-First Content
AI systems prioritize content that directly answers questions. Front-load answers – the core answer in the first 100 words. Structured Q&A with schema markup and precise definitions make your content attractive for Gemini and AI Overviews.
3. Make E-E-A-T Visible
Since AI systems rely on trustworthy sources, E-E-A-T is the gatekeeper. In my client projects at SEO Kreativ, I consistently implement author bios with expertise proof, links to primary sources, visible publication and update dates, and consistent linking across Wikipedia, LinkedIn, and the website itself.
4. Implement Structured Data
Schema.org markup helps AI systems understand entities and relationships. In my practice, I observe that pages with clean Article, FAQ, and Person schema are cited significantly more often in AI Overviews.
Measuring and Improving User Signals
Click-Through Rate (CTR) from SERPs
Where to measure: Google Search Console → Performance → Clicks/Impressions per query.
How to improve: Craft title tags that are emotional, specific, and use numbers or brackets. Meta descriptions with CTAs and value propositions. Rich Snippets via FAQ, How-To, or Review schema.
Bounce Rate & Pogo-Sticking
Where to measure: Google Analytics 4 → Engagement Rate + Search Console for position fluctuation.
How to improve: Most important information visible above-the-fold. Optimize Core Web Vitals (LCP < 2.5s). And the most fundamental: does the page deliver what the title promises?
The Future: Agentic Search and Beyond
With Project Mariner, Google has initiated the transition from “search engine” to “action engine.” Since February 2026, Mariner is available through Google Labs as a working prototype. The Chrome extension autonomously navigates websites, fills out forms, and can make bookings.
Mariner follows the pattern Observe → Plan → Act → Reflect. From my work in technical SEO audits, I already see that pages with clean schema markup and clear information architecture are preferred by agentic systems.
Trust becomes currency – in a world where AI makes decisions, E-E-A-T becomes the gatekeeper. Diversification is essential: dependence on Google traffic is risky, direct traffic and communities become more important.
Frequently Asked Questions (FAQ)
What is the difference between RankBrain and BERT?
RankBrain (2015) is a feed-forward neural network that links queries with concepts – especially valuable for new, unknown queries. BERT (2019) is a transformer model that understands bidirectional word context. RankBrain asks “What does the user mean?”, BERT asks “How does each word change the meaning?” Both work together, with DeepRank (BERT for ranking) increasingly taking over RankBrain’s tasks.
Does click-through rate really influence ranking?
Yes, this was officially confirmed in the 2023 DOJ trial. NavBoost uses click data from the last 13 months as one of the most important ranking signals. It distinguishes between “goodClicks” (longer dwell time) and “badClicks” (quick return to SERP). However, NavBoost is robust against manipulation – fake clicks are filtered.
What is DeepRank and why does it matter?
DeepRank is the internal name for BERT when used for ranking. It combines deep language understanding with world knowledge from the index. Due to its computational intensity, DeepRank only runs for the top 20-30 results – deciding exactly the positions 1-10 that matter most for SEO.
How does Gemini 3 influence search?
Gemini 3 has powered AI Mode in Google Search since late 2025. It performs parallel searches (Query Fan-Out), synthesizes answers from multiple sources, and generates dynamic interfaces (Generative UI). For SEO, this means: alongside classic rankings, “citation in the AI answer” becomes an important metric.
What are “goodClicks” and “badClicks”?
These terms come from the 2024 Google API leak. “GoodClicks” are clicks after which users don’t return to the SERP – they found what they were looking for. “BadClicks” are the opposite: quick return to search (pogo-sticking). Google also evaluates “lastLongestClicks” and “unicornClicks” (exceptionally positive signals).
Does AI replace traditional SEO?
No, AI reinforces traditional SEO principles. The systems reward high-quality, relevant content and punish manipulation more effectively than ever. Technical SEO (crawling, indexing, Core Web Vitals) remains the foundation. What’s changing: the ability to be cited in AI-generated answers becomes the new core competency.
Conclusion: Understanding and Leveraging AI Ranking
The key insights at a glance: machine learning is not an add-on but the core of Google ranking. User signals are officially confirmed as a ranking factor through NavBoost. Neural Matching and DeepRank understand concepts rather than keywords – thematic depth is the new standard. And with Gemini, “SERP position” becomes less relevant than “citation in the AI answer.”
I still regularly see in my SEO audits that teams optimize for individual systems in isolation. That falls short. The systems work as a pipeline, and only those who serve all stages rank consistently.
Your next steps: evaluate user signals in GSC and GA4, check content for search intent alignment, implement structured data (Article, FAQ, Person, Organization), track AI visibility, and strengthen E-E-A-T signals.
