Fresh

Google Discover Architecture

Google Discover is not search. Users do not type queries. The system predicts what content a user wants to see based on behavioral signals, entity interests, and real-time events. This documentation is derived from analysis of 87,498 decompiled classes from the Google app's Android APK.

---

Methodology

How This Research Was Conducted

Metehan decompiled the Google app (Android APK) and analyzed 87,498 Java/Kotlin classes related to content surfacing, feed ranking, and personalization. The findings below are based on code structure, class names, method signatures, protobuf definitions, and gRPC endpoint analysis -- not speculation.

---

The 8-Stage Content Pipeline

Content moves through 8 stages before appearing in a user's Discover feed:

---

Stage 2: Metadata Parsing Hierarchy

Discover parses content metadata in a strict priority chain. If the preferred source is missing, it falls through to the next.

Title Resolution

Priority	Source	Notes
1	og:title	Open Graph title tag
2	<title>	HTML title element
3	headline (schema)	Article structured data
4	First H1 on page	Fallback extraction

Author Resolution

Priority	Source	Notes
1	author (schema)	Article or Person schema
2	article:author (OG)	Open Graph author meta
3	Byline extraction	Pattern matching in page content
4	Domain name	Final fallback

Image Resolution

Priority	Source	Notes
1	og:image	Must be >= 1200px wide
2	image (schema)	Article schema image
3	First large image in content	Minimum dimension requirements
4	No image	Card displays without visual (severely penalized in feed ranking)

Image Requirements

Discover strongly favors content with high-quality images >= 1200px wide. Content without images receives significantly lower placement in the feed. Use og:image with proper dimensions on every article.

---

Stage 3: Paywall Classification

Discover classifies every article into one of three access tiers:

Classification	Definition	Feed Treatment
Free	No paywall detected	Normal ranking
Metered	Limited free articles, then paywall	Shown with "X free articles remaining" indicator
Locked	Full paywall, no free access	Deprioritized unless user has subscription signals

The system detects paywalls through:

• isAccessibleForFree schema property
• JavaScript-based paywall detection patterns
• Response header analysis
• Historical click-through and bounce rate patterns (locked content = high bounce)

---

Stages 4-5: Two-Level Filtering

Collection Filtering (Stage 4)

Collections are broad topic categories (Sports, Technology, Entertainment, etc.). A piece of content is first assigned to one or more collections based on entity and topic analysis. Collection-level filters remove:

• Content from blocked domains (spam, policy violations)
• Duplicate content (same story from multiple sources)
• Content that does not match any collection the user has shown interest in

Entity Filtering (Stage 5)

Entity-level filtering is more granular. Each piece of content is associated with specific entities (people, organizations, products, events). Entity filters remove:

• Content about entities the user has explicitly muted
• Content about entities with no engagement history
• Content about entities that have been "tombstoned"

---

Tombstoning System

Tombstoning is Discover's mechanism for removing stale or problematic content from the feed without deleting it from the index.

A content card is tombstoned when:

• The source URL returns a 404 or redirect
• The content has been significantly modified after initial indexing
• A manual action or policy violation is applied
• Engagement metrics fall below minimum thresholds after initial serving

Tombstoned content is not deleted -- it is flagged as ineligible for feed inclusion. It can be "un-tombstoned" if the underlying issue is resolved.

---

Stage 6: NAIADES Personalization

NAIADES is Google's personalization engine for Discover. It evaluates user interest across 8 subtypes:

Subtype	What It Measures	Signal Source
Explicit Interest	Topics the user explicitly followed	Google app settings, Search history
Implicit Interest	Topics inferred from behavior	Browsing history, app usage, location
Trending Interest	Rapidly growing topics in user's graph	Social signals, search volume spikes
Seasonal Interest	Recurring time-based interests	Historical engagement by month/season
Local Interest	Location-based relevance	GPS, IP, location history
Social Interest	Topics popular in user's social graph	Contacts, shared content, group memberships
Professional Interest	Work-related topics	LinkedIn signals, work app usage, business hours browsing
Entertainment Interest	Leisure content preferences	YouTube history, streaming app data, weekend browsing

Each subtype generates a score per entity/topic, and these scores are combined to determine overall user-content relevance.

---

13 Cluster Types

Discover groups content into clusters -- sets of related articles about the same topic or event. There are 13 cluster types:

#	Cluster Type	Example	Card Format
1	Breaking News	Major world event	Full-width card, multiple sources
2	Developing Story	Ongoing news story with updates	Timeline card
3	Evergreen Topic	"How to invest in index funds"	Standard card
4	Product Launch	New iPhone announced	Product card with specs
5	Sports Event	Live game or match	Score card with live updates
6	Entertainment	Movie release, TV show	Media card with trailer
7	Local News	City-level event	Location-tagged card
8	Weather Event	Storm, extreme weather	Alert card
9	Financial Event	Market movement, earnings	Data card with charts
10	Science/Research	New study published	Citation card
11	How-To/Tutorial	Step-by-step content	List card
12	Opinion/Analysis	Commentary on trending topic	Author-focused card
13	Visual Story	Photo essay, infographic-heavy	Gallery card

---

Freshness Bucketing Logic

Discover assigns each piece of content to a freshness bucket that determines its eligibility window in the feed.

Bucket	Age	Treatment	Best For
Real-time	0-4 hours	Highest priority, pushed to interested users	Breaking news, live events
Fresh	4-24 hours	High priority, standard feed inclusion	News, timely content
Recent	1-3 days	Moderate priority, competes on engagement	Analysis, reactions
Current	3-7 days	Lower priority, needs strong engagement signal	Evergreen-ish content
Evergreen	7-30 days	Lowest priority, only resurfaces on strong user match	Tutorials, guides
Archived	30+ days	Not shown unless directly relevant to trending topic	Historical reference

Freshness is Not Just Date

Discover uses datePublished, dateModified, HTTP Last-Modified headers, AND content-level freshness signals (new data points, updated statistics, current year references) to determine freshness. Changing the date without changing the content does not fool the system.

---

Beacon Push System

For time-sensitive content types, Discover uses a "beacon" push system that bypasses normal feed assembly.

Sports Beacons

• Triggered by live game events (goals, touchdowns, set wins)
• Pushed to users with demonstrated interest in the specific team/player
• Include real-time score data
• Override normal freshness bucketing

Finance Beacons

• Triggered by significant market movements (>2% index change, earnings surprises)
• Pushed to users with portfolio or market tracking behavior
• Include price data and change percentages
• Time-limited (expire after market close or next trading session)

---

Real-Time Feed Architecture (6 gRPC Endpoints)

Discover's feed is served through 6 gRPC endpoints, each handling a different aspect of feed assembly:

Endpoint	Function	Latency Target
/feed.GetFeedItems	Primary feed content retrieval	<200ms
/feed.GetClusterUpdate	Real-time updates to existing clusters	<100ms
/feed.GetBeaconPush	Time-sensitive push notifications	<50ms
/feed.ReportEngagement	User engagement telemetry (clicks, dwell, dismissals)	Fire-and-forget
/feed.UpdatePreferences	User preference changes (follow, mute, hide)	<150ms
/feed.GetFeedConfig	Client-side configuration (card layouts, feature flags)	Cached, <500ms

---

Publisher Recommendations

Content That Performs in Discover

1. High-quality images are mandatory. Use og:image with images >= 1200px wide on every article. Content without images is severely penalized.

2. Write compelling titles in og:title. Discover pulls from og:title first. This is your headline in the feed. Make it specific and engaging without being clickbait.

3. Publish consistently. Discover's personalization engine builds stronger entity associations with publishers who produce regular content on consistent topics.

4. Match content to entity clusters. Write about specific entities (people, products, organizations) rather than vague topics. Discover's entity filtering is more powerful than its topic filtering.

5. Use Article schema with all fields populated. datePublished, dateModified, author, image, headline -- every field gives Discover more metadata to work with.

6. Do not fake freshness. Updating dateModified without changing content is detectable and will hurt trust signals.

7. Earn Tier 1 engagement. Dwell time > 60 seconds is the threshold for highest engagement scoring. Deliver value fast but make content worth reading fully.

8. Monitor tombstoning. If your Discover traffic drops suddenly for a specific URL, check for 404s, redirects, or content modifications that may have triggered tombstoning.

The Discover Mindset

Discover is not search. You are not answering a query. You are predicting what someone wants to know before they ask. Write for entity-level interest matching, not keyword queries.