Inside Google’s Secret Search Systems
Investigating Google’s Inner Workings
Over recent months, we’ve delved deep into the workings of Google, uncovering fascinating insights we’re excited to share. Although we can’t reveal everything, what follows sheds light on how Google generates and ranks its search results.
What We Learned from 1,200 Experiments
We accessed nearly 1,200 Google experiments, with over 800 still active as of June 2025. This extensive dataset confirms that many components from the 2024 leaks, like Mustang and Twiddlers, remain key players in Google’s search algorithm. What’s more, we came across several intriguing new codenames, including Harmony, Thor, Whisper, Moonstone, and Solar.
Among the highlights:
- DeepNow: A successor to Google Now, paired with NowBoost.
- SuperGlue: Potentially set to replace Glue, similar to NavBoost for universal searches.
Continuous Evolution Over Major Overhaul
Unlike typical websites that update every few years, Google thrives on continuous evolution. Instead of launching a major new version, they make ongoing micro-changes that steadily transition from experimentation to full integration. This approach helps mitigate risks, as any failures only impact a small subset of users, and keeps innovation jumping at a pace that traditional redesigns can’t match.
The experiments cover diverse domains, including:
- AI (with multiple Magi and AI Mode variants)
- Shopping (featuring over 50 dedicated experiments)
- Various verticals like sports, finance, weather, and travel.
Every vertical is tucked into its own unique “domain,” such as ShoppingOverlappingDomain and SportsOverlappingDomain. This sophisticated structure allows for parallel testing among different product teams without conflict.
The Role of Entities in Google’s Ecosystem
Entities play a crucial role in all aspects of Google’s services, discussed in detail during a recent presentation, “Entities Everywhere,” by experts Damien Andell and Sylvain Deauré from 1492.vision. Their findings reveal that the Knowledge Graph underpins everything from Search to YouTube, Maps, and more.
The Knowledge Graph: Google’s Central Nervous System
The Knowledge Graph serves as Google’s essential framework, functioning far beyond a mere sidebar assistant. It drives Search, Discover, YouTube, Maps, Assistant, and more. Central to this is Livegraph, which assesses the reliability of information before approving it.
Data types are classified in a way that directly influences trust:
- kc: Highly validated information (e.g., government records)
- ss: Extracted web facts, less reliable
- hw: Information curated by humans
Ghost Entities and Real-Time Adaptation
One compelling discovery is the existence of ghost entities—temporary elements in the Knowledge Graph that allow Google to react almost immediately to new events. Unlike permanent entities, these can be dynamically generated and progressively validated before being surfaced in search results.
SEO Implications: Build a Validated Entity
For SEO professionals, the implications are clear: your brand must exist as a validated entity within Google’s expansive Knowledge Graph. Google assesses entire sites, measuring factors like thematic coherence, which boosts focused content and penalizes scattered material. Chrome data supports this by updating trust signals and identifying emerging trends.
Inside Google’s AI Mode: A Multitude of Projects
Our investigation also led to the discovery of an internal Google debug menu showing a new build dated May 28, 2025, which lists nearly 90 projects in development—an increase of over 40 from earlier.
A Strategy of Ultra-Specialized Agents
What’s most striking is Google’s multi-agent strategy. Rather than creating one all-purpose assistant, they’re launching a variety of specialized agents:
- MedExplainer for health
- Travel Agent and Flight Deals for trips
- Neural Chef, Food Analyzer, and Smart Recipe for cooking
Project Magi: The Backbone of AI Mode
Many of these projects fall under Project Magi, Google’s framework for AI Mode, boasting over 50 active experiments. Their rollout follows a structured approach with components such as:
- MagiModelLayerDomain: Core infrastructure
- MagitV2p5Launch: Aligns with Gemini 2.5
- SuperglueMagiAlignment: Similar to the Glue system that tracks user interaction
SEO Takeaways from AI Mode
- Hyper-specialization is vital—content must align with expert-level agents.
- Multi-modality is a must; text, images, videos, and structured data all contribute.
- Personalization is taken to a new level, influenced by context from an entire session rather than a single query.
The Profiling Engine: Transforming Interactions into Embeddings
Hidden within Google’s infrastructure is a system that converts every interaction into a mathematical embedding—an encoded vector encapsulating your online identity.
At the heart of this is Nephesh, creating vector representations of your behaviors and preferences across Google platforms.
Dual Embeddings for Google Discover
For the Discover service, Google utilizes a dual embedding system called Picasso and VanGogh:
- Picasso: Constructs a long-term profile from months of user interactions.
- VanGogh: Operates in real-time to capture immediate signals.
A Range of Specialized Embeddings
Google has various specialized embeddings including vertical, temporal, and contextual embeddings, allowing for refined personalization of user interactions. The HULK system detects behaviors in real-time, gauging user context to deliver pertinent results.
Query Understanding: Expansion and Real-Time Scoring
We also uncovered Google’s query expansion engine and a real-time scoring mechanism. For instance, searching “cycling tour france” triggers a transformation where “cycling tour” converts into “cyclingtour” and branches out further to related terms.
Geographic Intelligence
For location-driven queries, Google’s system maps geographic categories and translates terms as needed based on local intent.
These findings reaffirm that the architecture revealed in the 2024 leaks is still active and evolving.
