Generates per-query refinements based on session context, click history, and topical model. Powers the related-searches and refinement chips on the SERP — the system that helps users iterate from broad queries to specific intent.
Patent Overview
- Inventor
- Paul Haahr, others
- Assignee
- Google LLC
- Filed
- 2008
- Granted
- 2019-03-05
The Challenge
The Challenge
Users frequently submit broad queries that map to multiple intents. The system needs to suggest refinements that help users converge on their actual intent — without surfacing irrelevant or misleading suggestions.
- Broad Queries Have Multiple Intents — A broad query like 'java' can mean coffee, the language, or the island. The system can't fully resolve intent from the query alone.
- Refinements Must Be Relevant — Random refinements waste user attention. Suggestions must align with likely user intents revealed by aggregate behavior.
- Session Context Reveals Intent — Prior queries and clicks in the session reveal the user's actual topic. Refinements should use this context.
- Topical Coherence Matters — Refinements should form a coherent topical exploration path. Random topical jumps degrade UX.
- Adaptation Must Be Fast — Refinements appear in milliseconds. Generation and ranking must fit a tight latency budget.
Innovation
How The System Works
The system mines aggregate query patterns and session traces, builds per-query refinement candidates, scores candidates by relevance and topical coherence, applies session context, and surfaces top refinements as SERP chips.
- Mine Query Co-Occurrence Patterns — Offline, aggregate query logs reveal which queries follow which queries in sessions. Co-occurrence patterns become refinement candidates.
- Build Per-Query Candidate Set — Per query, retrieve candidate refinements from co-occurrence data plus topical-model expansions.
- Score Candidates — Per candidate, score relevance based on co-occurrence strength, topical alignment, and click-success rate.
- Apply Session Context — Per query within session, session-context vector modulates candidate scores. Refinements aligned with session topic earn boost.
- Topical Coherence Filter — Refinements form a coherent topical exploration path. Random topical jumps filtered.
- Surface Top-N — Top-N refinements by combined score surface as SERP chips or related-search suggestions.
- Learn From Clicks — Per refinement, click-through tracking feeds back into scoring. Refinements that work get reinforced.
Refinements Bridge Broad To Specific
The patent's load-bearing idea is that refinements are not random suggestions but learned bridges from broad queries to specific intent. Co-occurrence plus session context plus click feedback produces refinements that match real user trajectories.
Sessions Reveal Intent Pathways
User sessions show how broad queries decompose into specific intents. Aggregate session traces reveal the natural refinement pathways. The system mirrors them.
- Query Co-Occurrence Mining — Aggregate query logs reveal which queries follow which. Co-occurrence is the primary candidate source.
- Session-Context Modulation — Per-session topical context modulates candidate scores. Refinements aligned with session topic earn boost.
- Click-Feedback Reinforcement — Refinements that drive successful clicks earn reinforcement. The loop self-improves.
Technical Foundation
Technical Foundation
The patent specifies the co-occurrence miner, candidate builder, candidate scorer, session-context modulator, coherence filter, surface layer, and click-feedback loop.
- Co-Occurrence Miner — Offline, mines aggregate query logs for query-pair co-occurrence in sessions.
- Candidate Builder — Per query, builds candidate refinement set from co-occurrence data plus topical-model expansions.
- Candidate Scorer — Per candidate, scores relevance based on co-occurrence strength, topical alignment, click-success rate.
- Session-Context Modulator — Per session, topical context vector modulates candidate scores. Session-aligned refinements earn boost.
- Coherence Filter — Filters refinements that don't form coherent topical exploration paths.
- Click-Feedback Loop — Per refinement, click-through tracking feeds back into scoring. Self-improving loop.
The Process
The Process
Mining runs offline; refinement generation runs per query at query time.
- Offline Co-Occurrence Mining — Aggregate query logs mined for query-pair co-occurrence patterns.
- Receive Query — Query arrives. Refinement pipeline activates.
- Build Candidate Set — Per query, retrieve candidate refinements.
- Score Candidates — Per candidate, relevance score computed.
- Apply Session Context — Session-context vector modulates scores.
- Filter And Surface — Coherence filter applied; top-N surfaced as SERP chips.
- Track Clicks — Per refinement, click-through tracked and fed back into scoring.
Quality Control
Quality Control
Wrong refinements waste user attention. The patent specifies safeguards.
- Relevance-Threshold Filtering — Candidates below relevance threshold filtered. Avoids surfacing weak refinements.
- Topical-Coherence Check — Refinements that break topical coherence filtered. Path coherence preserved.
- Click-Feedback Monitoring — Per refinement, click-through rate monitored. Refinements failing to drive clicks demoted.
- Adversarial Defense — Refinement candidates from manipulated query log signals filtered. Prevents query-log spam from polluting refinements.
- Continuous Recalibration — Scoring weights and filters recalibrate against fresh data.
Real-World Application
Refinement systems power related-search chips, autocomplete suggestions, and the 'people also ask' question discovery on modern SERPs. The pattern of co-occurrence-mined, session-modulated, click-reinforced suggestions is the foundational architecture.
- Co-occurrence Primary Signal — Aggregate query logs mined for query-pair co-occurrence patterns.
- Session-modulated Personalization — Per-session topical context modulates candidate scores. Refinements align with session intent.
- Click-reinforced Feedback Loop — Refinements that drive successful clicks earn reinforcement. Self-improving system.
Why Topical Cluster Coverage Wins
Refinement systems surface topical neighbors. Sites covering a topical cluster well show up in multiple refinement paths from broad seeds — capturing discovery traffic across the user's exploration journey.
Why Click-Worthy Titles Matter In Related-Search
Refinement clicks feed back into scoring. Pages that earn clicks from refined queries get reinforced as targets for those refinements — a compounding loop favoring pages with click-attracting titles and SERP appearance.
<\/section>What This Means for SEO
What This Means for SEO
This patent powers related-search chips and refinement suggestions by mining query co-occurrence in sessions, modulating by session context, and reinforcing refinements that earn clicks. SEO implication: covering a topical cluster places you on multiple refinement paths from broad seeds, capturing discovery traffic across the user's exploration journey.
- Topical Cluster Coverage Captures Refinement Paths — Refinement systems surface topical neighbors of a broad query. Sites that cover a cluster well appear across multiple refinement paths from broad seeds, capturing discovery traffic throughout the user's journey.
- Click-Worthy Titles Win Related Searches — Refinement clicks feed back into scoring, so pages that earn clicks from refined queries get reinforced as targets. Strong titles and SERP appearance create a compounding loop that favors you on those refinements.
- Refinements Mirror Real Session Paths — Suggestions come from how users actually decompose broad queries into specific ones in sessions. Structuring content to answer the natural next questions in a journey aligns you with these learned paths.
- Session Context Boosts Aligned Content — Per-session topical context modulates which refinements surface. Content coherent with a user's evolving topic earns boost along their refinement path, so topical consistency across a journey helps.
- Be The Answer To The Refined Query — Refinements bridge broad to specific intent. Targeting the specific refined queries, not just the broad seed, positions you to win the traffic at the converging end of the funnel.
- Weak Refinement Targets Get Demoted — Refinements that fail to drive clicks are demoted. Earning genuine engagement from refined queries is what keeps you on the suggestion surface; weak pages fall off it.
- Manipulated Query-Log Signals Are Filtered — Refinement candidates from manipulated query-log activity are filtered out. You cannot inject yourself into refinements with fake query patterns; you earn the slot through real co-occurrence and clicks.