Generates a personalized content feed by treating each saved user interest as an implicit search query, scoring fresh results against the interest plus user-engagement signals, and surfacing the highest-value items as a continuously-refreshed stream.
Patent Overview
- Inventor
- Srinivasan Venkatachary
- Assignee
- Google LLC
- Filed
- 2017-02-28
- Granted
- 2024-07-09
- Application Number
- US 15/445,127
The Challenge
The Challenge
Users have persistent interests but typically express them only through one-shot queries. The system needs a way to satisfy those interests continuously through a feed of fresh, relevant content rather than waiting for the user to re-query each time.
- Search Is Reactive, Interests Are Persistent — A user interested in machine learning does not want to re-query 'machine learning news' every day. The system can serve them continuously if it treats the interest as a standing query.
- Interest Profiles Must Be Curated — User interests come from saved searches, followed topics, engagement patterns. The profile must be explicit enough that users can curate it (add, remove, refine) and inferred enough to handle implicit signals.
- Fresh Content Must Be Continuously Scored — Each new piece of content the crawler ingests must be scored against active user interests. The pipeline is streaming, not batch.
- Feed Quality Beats Quantity — A feed of weak matches loses user attention. Per-user, only top-scoring items should surface. The threshold must be high enough to keep the feed valuable.
- Engagement Feedback Refines Selection — Which feed items users click and dwell on refines future selection. The feed becomes more accurate with use as feedback accumulates.
Innovation
How The System Works
The system maintains per-user interest profiles, scores incoming content against each user's interests using the same retrieval and ranking machinery search uses, applies engagement-feedback weighting, surfaces top items as a personalized feed, and updates the interest profile from feed-interaction signals.
- Build Per-User Interest Profile — Profiles aggregate explicit interests (followed topics, saved searches), inferred interests (engagement history), and locale signals. Each entry is a structured interest record.
- Treat Each Interest As Implicit Query — Each saved interest becomes a standing implicit query the system continuously runs against new content. The mechanism leverages the existing search retrieval and ranking infrastructure.
- Score Fresh Content Per Interest — Each new crawled content item is scored against each user's interests. Standard retrieval relevance plus interest-specific weighting produces per-item scores.
- Apply Engagement Feedback — Past feed-interaction data refines scoring. Sources the user engages with rise; sources they skip fall. Per-user fine-tuning happens continuously.
- Compose The Feed — Top-scoring items across all interests compose the user's feed. Diversity rules prevent any single interest from dominating.
- Render And Capture Interaction — The feed renders in Discover or a comparable surface. User clicks, dwell, skips, and explicit feedback are captured.
- Update Interest Profile — Interactions update the interest profile. New engagement patterns spawn new inferred interests; abandoned interests decay.
Interests As Standing Queries
The patent's load-bearing idea is to treat each user interest as a standing implicit query that the retrieval system runs continuously. The feed emerges from the union of standing-query results, scored per-user and refined by engagement.
From Reactive To Continuous Search
Traditional search responds to explicit queries. Interest-driven feeds respond to persistent intents. The shift makes search continuous rather than transactional, serving users between query events.
- Interest Profile Curation — Per-user interest profile combines explicit and inferred signals. Users curate it explicitly; engagement refines it implicitly.
- Standing-Query Execution — Each interest is a standing query. The retrieval system runs it continuously against new content, surfacing matches as they appear.
- Engagement-Refined Selection — Per-user feedback refines selection. The feed becomes more accurate with use as engagement patterns inform future selection.
Technical Foundation
Technical Foundation
The patent specifies the interest-profile schema, the standing-query execution model, the per-user scoring pipeline, the engagement-feedback integration, and the feed-rendering layer.
- Interest Profile Schema — Per-user, the profile records explicit and inferred interests with metadata (source, confidence, decay rate). Structured for fast scoring.
- Standing-Query Index — Standing queries are indexed for fast match against streaming content. The structure mirrors inverted indexes but optimized for continuous re-evaluation.
- Per-User Scorer — For each user, the scorer evaluates new content items against their interests. Personalization plus engagement-feedback weighting produces per-user per-item scores.
- Diversity Compositor — Selects top items across interests subject to diversity constraints, so no single interest dominates the feed.
- Engagement Feedback Pipeline — Feed-interaction signals (clicks, dwell, skips, explicit feedback) feed back into the profile and the scorer. The system learns continuously.
- Rendering Surface — The feed renders in Discover, search-app home, or comparable surfaces. Format adapts to surface constraints and device.
The Process
The Process
The pipeline runs as a continuous stream from crawled content to user feeds, with per-user scoring happening as fresh content arrives and as users open their feeds.
- Content Crawled — Crawler ingests new content. Indexer extracts content, metadata, and content embeddings.
- Score Against Standing Queries — For each item, the system scores it against active standing queries (user interests). Per-item, per-user scores are computed and cached.
- Apply User Engagement Weights — Per-user engagement feedback adjusts scores. Sources the user engages with get lift; abandoned sources sink.
- User Opens Feed — When the user opens the feed, the compositor selects top items across their interests subject to diversity constraints.
- Render Feed — Items render in feed surfaces (Discover, etc.). Layout adapts to device and surface.
- Capture Interactions — Clicks, dwell, skips, and explicit feedback log per item per user. Logs feed engagement and profile updates.
- Update Profile — Interaction patterns update the interest profile. New interests emerge; abandoned interests decay.
Quality Control
Quality Control
Feed quality determines whether users return. The patent specifies safeguards against feed degradation, filter bubbles, and privacy leaks.
- Quality Threshold — Items below the score threshold do not appear in the feed. Better to show fewer items than weak ones.
- Diversity Enforcement — Diversity rules across interests, sources, and content types prevent monoculture feeds. Users see breadth, not just depth.
- Filter Bubble Mitigation — The system injects diversity items that broaden the user's exposure beyond established interests, preventing narrow feedback loops.
- Privacy Boundary Enforcement — Sensitive interest categories (health, finance) are handled with stricter rules or excluded from inferred profiling.
- User Control Surfaces — Users can add or remove interests, hide sources, mute topics, and clear feedback history. Controls are first-class.
Real-World Application
Interest-driven feeds power Google Discover, the search-app home stream, and the topical-content surfaces across Android and Chrome. The patent's primitives shape how Google serves persistent information needs continuously.
- Standing-query Execution Model — Each interest is a standing query the retrieval system runs continuously. The pipeline is streaming, not batch.
- Per-user Scoring Scope — Content scores per user against their interests. Two users see different feeds reflecting different interest profiles.
- Engagement-refined Continuous Learning — Per-user feedback refines selection continuously. The feed gets more accurate with use.
Why Discover Is A New Traffic Channel
Pages aligned with user-interest patterns surface in Discover without users explicitly searching for them. For sites covering specific topics consistently, Discover becomes a substantial traffic source that bypasses query-based ranking entirely.
Why Topical Consistency Builds Feed Visibility
Sites that publish consistently on a topic match user-interest standing queries reliably. The pattern rewards topical focus over breadth, since narrow specialty content fits standing queries cleanly while broad-spectrum content fits weakly.
<\/section>What This Means for SEO
What This Means for SEO
The patent treats each saved user interest as a standing implicit query run continuously, scoring fresh content against the interest plus engagement to build a personalized feed. SEO implication: consistent topical publishing matches standing interest queries and surfaces in Discover-style feeds without users searching.
- Discover Is A Query-Free Traffic Channel — Pages aligned with user-interest patterns surface in feeds without explicit searches. For sites covering specific topics consistently, this becomes a substantial traffic source that bypasses query-based ranking entirely. Plan for it as its own channel.
- Topical Consistency Builds Feed Visibility — Sites publishing consistently on a topic match standing interest queries reliably. Narrow specialty content fits standing queries cleanly while broad-spectrum content fits weakly. Topical focus is the structural advantage for feed placement.
- Fresh Content Is The Trigger — The system scores incoming content against interests continuously. Regular publishing of fresh, relevant material is what keeps surfacing you into feeds, whereas static content rarely re-enters the continuously-scored stream.
- Same Ranking Machinery Applies — The feed uses the same retrieval and ranking machinery as search. Search fundamentals (relevance, quality, authority) still gate feed eligibility, so feed visibility builds on, not instead of, sound on-page work.
- Engagement Feedback Weights Selection — Engagement-feedback weighting and feed-interaction signals refine the profile. Content that earns engagement in the feed strengthens future selection for similar users, so compelling headlines and lead imagery compound feed reach.
- Interests Are Persistent, Not One-Shot — The shift is from reactive query response to continuous interest service. Covering a topic deeply enough to be a standing answer to a persistent interest earns recurring exposure between the user's explicit search events.
- Match The Interest, Not A Keyword — Standing queries represent interests, not exact keywords. Serving the underlying interest comprehensively aligns you with the standing query better than optimizing for a specific phrase the user never explicitly types into the feed.