Metehan: Google's AI Search architecture
- th3s3rp4nt
- 8. Dez. 2025
- 2 Min. Lesezeit
Key Takeaways:
Metehan discovered in Google's Discovery Engine 7 ranking signals that likely power AI Mode (Vertex AI search)
Base Ranking → core algorithm output
Gecko Score → embedding similarity
Jetstream → cross-attention model (handles negation better than embeddings)
BM25 → yes, keyword matching still matters
PCTR → 3-tier engagement system
Freshness → time-sensitive query adjustment
Boost/Bury → manual business rules
His key takeaways:
Ancestor headings travel WITH each chunk
Prediction is here. PCTR unlocks in tiers: Popularity → PCTR → Personalized PCTR
Personalized ranking only activates after 100K+ queries
Query Suggestions Model Options:
Automatic: System decides the best source
Document: Suggestions derived from indexed document content
Search History: Suggestions based on past searches
User Events: Suggestions driven by user interaction data
Completable Fields: Suggestions from schema-defined completable attributes
Actionable insights:
Optimize for grounded synthesis. The LLM is constrained to ground answers in retrieved content. Your content needs clear, extractable statements.
Avoid adversarial patterns. Content that triggers adversarial detection gets filtered. Write straightforwardly without manipulation patterns.
Maintain high relevance density. Low relevance content gets filtered. Every section should contribute to query relevance.
Structure for multi-turn. AI Mode supports follow-ups. Content that answers related questions in the same topic cluster may have advantages.
1. Seven explicit signals, not a black box
The Signal Viewer exposes seven distinct ranking signals. This isn’t a single neural network making opaque decisions. It’s a multi-signal fusion system with interpretable components.
2. Gecko + Jetstream = Semantic layer
Google uses both embedding similarity (Gecko) and cross-attention (Jetstream) for semantic understanding. Jetstream’s negation handling suggests nuanced query understanding beyond simple similarity.
3. PCTR is a three-tier system, not a single signal
Engagement signals operate in tiers: Popularity → PCTR → Personalized PCTR. Each tier has quality thresholds and unlocks with more user interaction data. Personalized ranking only activates after 100,000+ queries. (And of course, this is about the Discovery Engine)
4. 500-token chunk limit with optional heading context
The retrieval unit has a ~500 token maximum, with an option to include ancestor headings. Structure your content accordingly.
5. Three product tiers from one pipeline
Traditional search, AI Overviews, and AI Mode are all served from the same pipeline with different configurations. The architecture is unified.
6. Adversarial and relevance gates
Final safety filters can block content even after it ranks well. Write naturally and maintain high relevance throughout.
Sources:
