sinapse-ai 9.3.0 → 9.5.0

package/squads/squad-content/knowledge-base/social-algorithms-master-reference.md

@@ -0,0 +1,1007 @@
+# Social Algorithms Master Reference
+
+> **Source:** MS-007 Research (Social Algorithms & Content Distribution)
+> **Date:** 2026-04-07 | **Sources:** 42+ consulted
+> **Scope:** Instagram, TikTok, YouTube, LinkedIn, Twitter/X, Facebook algorithms, recommendation systems theory, cross-platform strategy, engagement mechanics, creator economy, Brazilian context
+
+---
+
+## Table of Contents
+
+1. [Panorama & Evolution](#1-panorama--evolution)
+2. [Instagram Algorithm](#2-instagram-algorithm)
+3. [TikTok Algorithm](#3-tiktok-algorithm)
+4. [YouTube Algorithm](#4-youtube-algorithm)
+5. [LinkedIn Algorithm](#5-linkedin-algorithm)
+6. [Twitter/X Algorithm](#6-twitterx-algorithm)
+7. [Facebook Algorithm](#7-facebook-algorithm)
+8. [Emerging Platforms](#8-emerging-platforms)
+9. [Recommendation Systems Theory](#9-recommendation-systems-theory)
+10. [Cross-Platform Content Strategy](#10-cross-platform-content-strategy)
+11. [Engagement Mechanics](#11-engagement-mechanics)
+12. [Creator Economy & Monetization](#12-creator-economy--monetization)
+13. [Social Commerce](#13-social-commerce)
+14. [AI & Social Media](#14-ai--social-media)
+15. [Brazilian Context](#15-brazilian-context)
+16. [Key People & References](#16-key-people--references)
+17. [Actionable Checklists](#17-actionable-checklists)
+
+---
+
+## 1. Panorama & Evolution
+
+### 1.1 Three Eras of Social Feeds
+
+| Era | Period | Model | Key Characteristic |
+|-----|--------|-------|-------------------|
+| **1. Chronological** | 2004-2012 | Posts in time order | Simple but inefficient at scale |
+| **2. Algorithmic Social** | 2012-2020 | Content from followed accounts, ranked | Social graph as primary filter |
+| **3. Interest-Based Recommendation** | 2020-present | Content based on interests, not follows | TikTok proved interest graph > social graph |
+
+Facebook introduced EdgeRank in 2009 (affinity x weight x time decay). Instagram abandoned chronological in 2016. TikTok's FYP doesn't depend on who you follow -- it depends on what you watch.
+
+Adam Mosseri (2023): "Recommendations are the future of Instagram."
+
+### 1.2 The Attention Economy
+
+Herbert Simon (1971): "A wealth of information creates a poverty of attention."
+
+**Key numbers (2025):**
+- Average global social media time: 2h24min/day
+- Brazil average: **3h49min/day** -- highest in the world (DataReportal 2025)
+- Average Instagram feed: 500+ new posts/day from followed accounts
+- Only ~30% of available content is actually shown to users
+- Algorithms optimize for: session time, return frequency (DAU/MAU), and ultimately ad revenue
+
+### 1.3 Discovery Surfaces Taxonomy
+
+| Surface | Function | Examples |
+|---------|----------|----------|
+| **Main Feed** | Followed + recommended content | Instagram Feed, Twitter For You |
+| **Discover/Explore** | New content discovery | Instagram Explore, TikTok FYP |
+| **Short-form Video** | Algorithmic short video | Reels, Shorts, TikTok |
+| **Search** | Intentional search | YouTube Search, Instagram Search |
+| **Stories** | Ephemeral ranked content | Instagram/Facebook Stories |
+| **Messaging/DM** | Private sharing (strong signal) | WhatsApp forwards, Instagram DM shares |
+| **Notifications** | Re-engagement | Push notifications, email digests |
+
+All surfaces share fundamental signals: engagement signals, content signals, user signals, and context signals.
+
+### 1.4 Platform vs Creator vs User Tension
+
+- **Platforms** want to maximize: time on platform, ad impressions, DAU/MAU, revenue
+- **Creators** want to maximize: reach, followers, engagement, monetization
+- **Users** want: relevant content, social connection, entertainment, information
+
+What users say they want and what their behavior reveals are often divergent -- YouTube calls this "nutritious content vs junk food content."
+
+---
+
+## 2. Instagram Algorithm
+
+### 2.1 Multiple Algorithms
+
+Instagram does NOT have one algorithm. Feed, Explore, Reels, and Stories each use different algorithms. Infrastructure: PyTorch-based ML at billions of predictions/second for 2B+ MAU.
+
+### 2.2 Feed Ranking Signals (by weight)
+
+1. **Interest:** Probability of user interacting, based on past behavior. Model predicts likelihood of 5 actions: like, comment, save, share, view time
+2. **Relationship:** How close the user is to the author. Signals: mutual interactions, DMs, tags, profile searches
+3. **Timeliness:** Temporal decay -- 30-min-old post > 3-day-old post. Viral content with high engagement can resurface
+4. **Usage Frequency:** Users who open multiple times/day see more chronological feed; once/day users see "best of"
+5. **Diversity:** Prevents too many consecutive posts from same author or same content type
+
+**Feed composition (2024):** ~15-20% recommended content from non-followed accounts, growing.
+
+### 2.3 Reels Algorithm (Key Differences from Feed)
+
+Over 50% of Reels content comes from accounts the user does NOT follow. This is a discovery surface.
+
+**Signal hierarchy:**
+
+| Signal | Weight | Details |
+|--------|--------|---------|
+| **Watch Time / Completion Rate** | Highest | Reel watched to end or rewatched = massive boost. "Retention curve" -- early drop-off penalized |
+| **Shares (DM)** | Very High | Mosseri: "Sends are the most important signal for Reels ranking" |
+| **Saves** | High | Indicates lasting value (reference content) |
+| **Engagement Velocity** | High | First 30-60 minutes determine algorithmic fate |
+| **Trending Audio** | Medium | Reels using trending audio get boost |
+| **Originality** | Medium | Penalizes TikTok watermarks, reposted content, aggregators |
+
+**Known penalties:**
+- TikTok watermark: significant distribution reduction
+- Low resolution (<720p): penalized
+- Text covering >20% of screen: may reduce reach
+- Recycled content without added value: penalized
+- Community guideline violations: shadow restriction
+
+### 2.4 Stories Ranking
+
+Primarily by relationship proximity:
+- **View history:** Consistently watched accounts appear first
+- **Story interactions:** DM replies, emoji reactions, poll votes are strong signals
+- **Timeliness:** More recent Stories appear first
+- **Content type:** Algorithm learns format preference (video vs photo)
+
+**Strategic insight:** Stories with interactive elements (polls, quizzes, sliders, questions) generate 2-3x more engagement and train the algorithm to prioritize your profile.
+
+### 2.5 Explore -- Discovery
+
+Exclusively non-followed content. Process:
+
+1. **Candidate Generation:** Find accounts similar to recently engaged accounts (co-engagement)
+2. **Two-Tower Model:** User tower + content tower generate embeddings; proximity determines relevance
+3. **Ranking:** By engagement probability (like > save > share > comment for Explore)
+4. **Filtering:** Guideline violations, low quality filtered out
+5. **Diversity Injection:** Topical diversity to avoid excessive bubbles
+
+### 2.6 Shadowban Mechanics
+
+Instagram denies "shadowban" but acknowledges "distribution reductions":
+- **Community Guidelines violations:** Immediate reduction even without removal
+- **Recommendation Guidelines:** "Borderline" content removed from Explore/Reels but not from Feed
+- **Repeated violations:** Systemic reduction
+- **Engagement bait:** Excessive "like, comment, share" requests may be penalized
+
+**What does NOT cause reduction:** Account type (personal vs business/creator), posting frequency (within reason), links in bio, editing captions after posting.
+
+### 2.7 Algorithm Changes 2025-2026
+
+| Change | Impact |
+|--------|--------|
+| **"Your Algorithm" tool** (Dec 2025) | Users can see/control topics shaping Reels recommendations |
+| **Carousels dominant** | 10% avg engagement rate (vs 7% images, 6% Reels). 1.4x more reach, 3.1x more engagement vs single photos |
+| **Aggregator penalty** | Accounts with 10+ reposts in 30 days excluded from recommendations. Original creators saw 40-60% reach increase |
+| **Watch time as #1 signal** (Jan 2025) | First 3 seconds critical for initial distribution. Likes per reach and DM shares = 2nd and 3rd signals |
+| **DM/Comments/Stories weight** (Mar 2026) | Even greater weight to accounts with frequent DM, comment, and Story interactions |
+| **AI categorization** (2026) | Visual elements, text in images, video clips analyzed beyond hashtags -- hashtags less relevant |
+| **AI translation for Reels** (late 2025) | Auto-translation of text and audio (Hindi, Portuguese, English, Spanish) |
+
+### 2.8 Engagement Velocity -- The Critical Window
+
+1. Post shown to ~10% of followers
+2. Algorithm measures: engagement rate, view time, saves, shares
+3. If performance exceeds creator's baseline: shown to more followers
+4. Exceptional performance: enters Explore and Reels (for Reels)
+5. Each distribution round exponentially expands audience
+
+**Practical implication:** Post when followers are online (check Instagram Insights) to maximize engagement velocity.
+
+---
+
+## 3. TikTok Algorithm
+
+### 3.1 The For You Page Revolution
+
+**Regulatory context (2025-2026):** TikTok USDS Joint Venture LLC established Jan 22, 2026. ByteDance retained minority stake <20%. Oracle, Silver Lake, MGX each hold ~15%. US algorithm trained exclusively on Oracle-hosted servers.
+
+TikTok inverted the social media model: any video from any account can go viral regardless of follower count. FYP is based on **interest graph** (what you watch) not **social graph** (who you follow).
+
+### 3.2 Algorithm Signals
+
+**Tier 1 -- Content Signals (highest weight):**
+
+| Signal | Details |
+|--------|---------|
+| **Completion Rate** | Most powerful signal. 100% completion = max boost. Rewatched = additional boost |
+| **Watch Time Total** | Absolute time matters. 60s video watched 55s can beat 15s video at 100% |
+| **Shares** | Via DM or copied link = extremely strong. Valued above likes |
+| **Comments** | Long, detailed comments weigh more than emojis |
+| **Profile Visits** | Visiting creator profile after watching = strong interest signal |
+
+**Tier 2 -- User Signals:** Interaction history, created content types, followed accounts (less weight than Instagram), "Not Interested" marks (strong negative signal)
+
+**Tier 3 -- Device/Account Signals:** Language preference, country/location, device type (used as purchasing power proxy per leaked ByteDance docs), usage timing patterns
+
+### 3.3 Batch Testing (Pool System)
+
+| Pool | Views | What Happens |
+|------|-------|-------------|
+| **Pool 1** | ~200-500 | Shown to small diverse group. Measures completion rate, engagement, shares. Needs ~50% completion + above-baseline engagement to advance |
+| **Pool 2** | ~1,000-5,000 | Shown to users with similar interests. Metrics must maintain or improve |
+| **Pool 3** | ~10,000-100,000 | Broad distribution, geographic diversity. "Viral potential" |
+| **Pool 4+** | 100K-millions | Massive cross-geographic distribution. Algorithm saturates until engagement decays |
+
+**Critical insight:** A video can "die" in Pool 1 and "resurrect" days or weeks later. TikTok periodically retests old content.
+
+### 3.4 Interest Graph vs Social Graph
+
+| Aspect | Social Graph (Instagram/Facebook) | Interest Graph (TikTok) |
+|--------|-----------------------------------|------------------------|
+| **Base** | Who you follow | What you watch |
+| **New creator** | Must build followers | Can go viral on 1st video |
+| **Diversity** | Limited to social graph | High -- diverse FYP |
+| **Cold start** | Slow (depends on network) | Fast (5-10 watched videos create profile) |
+| **Lock-in** | High (followers are "property") | Low (algorithm can defavor anytime) |
+
+Eugene Wei, "TikTok and the Sorting Hat" (2020): TikTok functions like Harry Potter's Sorting Hat -- rapidly categorizes users by interest without requiring them to declare preferences.
+
+### 3.5 Content Diversity & Optimization
+
+**Diversity mechanisms:**
+- Category cap: limits consecutive videos of same topic
+- Creator diversity: avoids multiple videos from same creator in sequence
+- New content injection: videos from new creators periodically injected
+- Exploration allocation: ~10-20% of impressions for exploration outside user's interest profile
+
+**Optimization tips:**
+- **First 2-3 seconds:** Hook is everything. High skip rate = suppression
+- **Duration:** No universal "ideal." What matters is completion rate vs total time
+- **Trending audio:** One of most consistent boosts
+- **On-screen text:** TikTok indexes overlay text for categorization (works as "implicit hashtags")
+- **Hashtags:** Primarily for categorization, not discovery. #fyp/#foryou are irrelevant
+
+### 3.6 Monolith Architecture (ByteDance, 2022)
+
+Key innovations from the published paper:
+- **Real-time training:** Model trained continuously (not in daily batches)
+- **Collisionless embedding:** Improved feature representation accuracy
+- **Feature eviction:** Old irrelevant features automatically discarded
+- **Scale:** Billions of interactions processed daily in real-time
+
+---
+
+## 4. YouTube Algorithm
+
+### 4.1 The Core Formula: CTR x AVD
+
+**CTR (Click-Through Rate):** % of people who see thumbnail/title and click. Average: 2-10%. 8%+ is excellent.
+**AVD (Average View Duration):** Average time watching. YouTube values absolute retention (minutes) AND relative (%).
+
+- High CTR + low AVD = clickbait
+- High AVD + low CTR = good content, weak packaging
+- YouTube wants BOTH
+
+### 4.2 Discovery Surfaces
+
+| Surface | Traffic Share | Key Signals |
+|---------|-------------|-------------|
+| **Browse (Homepage)** | ~40-60% | Viewing history, subscriptions, trends. Highly personalized. Favors "novelty" |
+| **Suggested (Watch Next)** | ~30-40% | Topic correlation ("who watched X also watched Y"). Strongest signal: topical match |
+| **Search** | Varies | Title/description/tag relevance, watch time history, engagement, channel authority |
+| **Shorts** | Separate ecosystem | Completion/loop rate dominant. Cross-pollination with long-form limited |
+
+### 4.3 Satisfaction Surveys
+
+Unique YouTube innovation: in-app surveys asking "How would you rate this video?" (1-5 stars) or "Was this video worth your time?"
+
+This led to "responsible recommendation" -- balancing engagement with declared satisfaction. Junk food content (sensational, clickbait) may generate high watch time but low satisfaction.
+
+### 4.4 Thumbnail A/B Testing (2024+)
+
+- Creator uploads 2-3 thumbnails per video
+- YouTube shows each to equal audience portions
+- Winner determined by **watch time generated** (not just CTR, to avoid rewarding clickbait)
+- Automatically selected after 24-48h
+
+Mr. Beast: tests dozens of thumbnails before publishing. Thumbnails are the most controllable factor for CTR.
+
+### 4.5 Session Time & Ecosystem
+
+YouTube optimizes for the entire session, not just individual videos. A **session starter** (first video watched when opening YouTube) receives Browse boost. This favors regular, predictable publishing.
+
+**Subscriber bell:** Only ~10-20% of subscribers activate "All" notifications. If subscribers don't watch, it's a negative signal.
+
+### 4.6 YouTube Shorts Updates 2025-2026
+
+| Update | Details |
+|--------|---------|
+| **Extended duration** (2025) | Shorts can now be up to 3 minutes (was 60 seconds) |
+| **New view counting** (Mar 2025) | Any Short that starts playing or loops counts as view. Each loop = additional view |
+| **Algorithm separation** (late 2025) | Shorts recommendation engine fully separated from long-form |
+| **Search filter** (Jan 2026) | Dedicated "Shorts" filter in search type menu |
+| **Browse Feed change** (Dec 2025) | Homepage reduced long-form recommendations (~12 to ~2), prioritizing Shorts |
+| **Satisfaction > watch time** (2026) | Satisfaction surveys and post-viewing behavior now surpass raw watch time |
+
+### 4.7 YouTube Recommendations Paper (Covington et al., 2016)
+
+Two-stage architecture:
+1. **Candidate Generation:** Millions of videos reduced to hundreds using collaborative filtering with deep learning
+2. **Ranking:** Candidates ranked with complex model (watch time, engagement, freshness, upload frequency)
+
+Objective function: watch time (not clicks). "Example Age" feature: video age as feature to learn natural relevance decay.
+
+---
+
+## 5. LinkedIn Algorithm
+
+### 5.1 B2B Context
+
+~1B members (2025), but only ~3-5% publish content actively. Favorable supply/demand for creators.
+
+### 5.2 Signal Hierarchy
+
+| Signal | Weight | Details |
+|--------|--------|---------|
+| **Dwell Time** | Highest | Time spent reading, even without interaction. "Qualified dwell time" (actually read) vs "passive" (tab inactive). 30+ seconds = high quality |
+| **Meaningful Comments** | Very High | 50+ character comments: high weight. Short emoji comments: low weight. Thread replies: medium-high. 1st-degree connections' comments: extra weight |
+| **Shares/Reposts** | High | Shares with substantive comment >> simple reposts |
+| **Network Relevance** | Medium | 1st and 2nd degree connections prioritized. Strangers only if exceptional performance |
+
+LinkedIn explicitly reduces distribution of "engagement bait" ("comment YES if you agree", "repost if you've lived this").
+
+### 5.3 SSI (Social Selling Index)
+
+Score 0-100 across 4 dimensions:
+1. Establishing Your Professional Brand
+2. Finding the Right People
+3. Engaging with Insights
+4. Building Relationships
+
+Higher SSI correlates with greater organic reach (functions as implicit "creator score").
+
+### 5.4 Content Performance by Format
+
+| Format | Average Reach | Engagement | Best For |
+|--------|-------------|------------|----------|
+| **Document/Carousel** | High | High | Educational, frameworks, lists |
+| **Long text** | Medium-High | Medium | Storytelling, opinion, experience |
+| **Image + text** | Medium | Medium | News, celebrations |
+| **Video** | Medium | Medium-Low | Interviews, behind-the-scenes |
+| **External link** | Low | Low | LinkedIn penalizes external links |
+| **Newsletter** | Very High (via email) | High | Long recurring content |
+| **Poll** | High (declining) | High | Market research |
+
+**Critical insight:** LinkedIn penalizes posts with external links. Best practice: create native content. Putting link in first comment is also being penalized (2024+).
+
+### 5.5 LinkedIn Newsletters
+
+Best organic performance feature:
+- Subscribers get email AND push notification for each edition
+- Open rates 20-50% (vs ~2-5% organic reach for posts)
+- Each edition distributed as both post and email
+
+### 5.6 Employee Advocacy
+
+Employee posts have **561% more reach** than company page posts (MSLGroup). Personal profiles have 3-5x greater organic reach than company pages.
+
+### 5.7 Changes 2025-2026
+
+- **Saves and Sends** added to analytics (signals valued by algorithm)
+- **Anti-automation filter:** Excessive comments or automation tools can limit visibility
+- **Quality over quantity:** Limit to 1-3 highly relevant tags. Expert-led content prioritized
+- Algorithm functions increasingly as "trust filter"
+
+---
+
+## 6. Twitter/X Algorithm
+
+### 6.1 Open-Source Revealed Weights (March 2023)
+
+| Signal | Approximate Weight |
+|--------|-------------------|
+| **Reply** | 1x (baseline) |
+| **Like** | 0.5x |
+| **Retweet/Repost** | 1x |
+| **Quote Tweet** | 1x |
+| **Bookmark** | Confirmed signal (weight undisclosed) |
+| **Reading time** | Growing weight |
+| **Profile click** | Strong signal |
+| **Link click** | Medium signal |
+
+**Boost factors:**
+- Images: ~2x boost over plain text
+- Video: ~2x boost
+- External links: **penalized** (X wants users on platform)
+- Long threads: boosted by accumulated engagement
+- Premium subscribers: ~4x boost in For You ranking (pay-to-play)
+
+### 6.2 For You vs Following
+
+- **For You (default):** ~50% followed + ~50% recommended. "Heavy Ranker" ML model
+- **Following:** Strictly chronological, no recommendations
+
+### 6.3 Community Notes
+
+Decentralized fact-checking. Posts with Community Notes may have reduced distribution. Uses "bridging" algorithm -- notes only published if evaluators from different political perspectives agree.
+
+### 6.4 Grok AI Integration
+
+Grok (xAI) integrated directly into platform. Can summarize threads, answer questions about trending topics, analyze images. Evolving constantly.
+
+---
+
+## 7. Facebook Algorithm
+
+### 7.1 Evolution: EdgeRank to MSI
+
+- **EdgeRank (2009-2011):** Affinity x Weight x Decay
+- **ML Era (2011-2018):** Thousands of features, optimized for engagement (inadvertently favored sensational content)
+- **MSI - Meaningful Social Interactions (2018+):** Posts from friends/family prioritized over Pages. Content generating "meaningful conversations" (long comments, replies) prioritized. Publishers saw 40-60% organic reach drop
+
+### 7.2 Current State (2025)
+
+- ~30-40% of feed is recommended content from non-followed accounts (response to TikTok)
+- **Reels:** Maximum distribution priority on Facebook
+- **Groups:** Prioritized, especially active communities
+- **Link penalty:** Links receive lower distribution than native content (photos, videos, text)
+- Company Pages organic reach: frequently <2% of followers (lowest among all platforms)
+
+### 7.3 Facebook Reels vs Long-Form
+
+- **Reels:** Largest organic distribution of any Facebook format. Algorithm similar to Instagram Reels
+- **Long-form video:** Drastically deprioritized in timeline (2017 video pivot partially reversed)
+- **Facebook Live:** Still boosted during broadcast, but less impactful than 2018-2020
+
+---
+
+## 8. Emerging Platforms
+
+### 8.1 Threads (Meta)
+
+- 100M users in 5 days (fastest app launch ever, July 2023)
+- 400M MAU (Aug 2025), 450M MAU (early 2026)
+- Surpassed X in mobile DAU (141.5M vs 125M) in January 2026
+- Algorithm: Interest-based (similar to TikTok), not just follows
+- Instagram data influences recommendations
+- Replies and reposts are strongest signals
+- External links NOT penalized as strongly as on X
+
+### 8.2 Bluesky (AT Protocol)
+
+**Key innovation: Custom Feeds (Algorithmic Choice)**
+- Users choose between multiple algorithms or create their own
+- Chronological feed (default) + "Discover" + community-created feeds
+- **AT Protocol:** Data portability -- users can theoretically migrate account, followers, content to another provider
+- **Composable Moderation:** Users choose "labelers" and filter preferences
+
+### 8.3 WhatsApp Channels
+
+- One-way broadcast updates. Content disappears after 30 days
+- Ranking signals: recency, engagement (reactions), regional popularity
+- In Brazil, especially relevant given 99% WhatsApp penetration
+- Open rates significantly higher than email or traditional social media
+
+### 8.4 Pinterest
+
+- Visual search engine more than social network
+- Algorithm based on **intent** (discovery: decor, recipes, fashion)
+- Pins have extremely long lifespan (months to years of traffic)
+- SEO is fundamental (keywords in descriptions, board names)
+- Pinterest has highest purchase intent (83% use it to plan purchases)
+
+### 8.5 Reddit
+
+- Karma system (upvotes/downvotes determine visibility)
+- Subreddit governance with unique micro-cultures
+- "Hot" algorithm: karma + recency
+- Google partnership (2024): Reddit extensively indexed. Powerful indirect SEO platform
+
+---
+
+## 9. Recommendation Systems Theory
+
+### 9.1 Collaborative Filtering
+
+| Method | How It Works | Scale |
+|--------|-------------|-------|
+| **User-based** | "Users similar to you liked X" | Scales poorly (O(n^2)) |
+| **Item-based** | "Items frequently consumed together are similar" | More scalable (Amazon's "also bought") |
+| **Matrix Factorization** | Decomposes user-item matrix into latent factors. Dot product predicts interaction | Won Netflix Prize ($1M, 2009) |
+
+### 9.2 Content-Based Filtering
+
+Analyzes content properties instead of user behavior:
+- NLP for text (topics, sentiment, entities)
+- Computer Vision for images/video (objects, scenes, faces)
+- Audio analysis (genre, BPM, sentiment)
+- Metadata (tags, categories, hashtags, duration)
+
+Solves cold start for new items. Limitation: doesn't capture implicit/subtle preferences.
+
+### 9.3 Two-Tower Models (Dominant Architecture)
+
+```
+User Tower:              Content Tower:
+[history]                [content type]
+[demographics]  -->      [visual features]     --> Similarity Score
+[interactions]  Embedding [text features]       Embedding
+[device]                 [engagement stats]
+```
+
+Used by Instagram Explore, YouTube, TikTok. Pre-computed embeddings enable efficient approximate nearest neighbor (ANN) search.
+
+### 9.4 Exploration vs Exploitation (Multi-Armed Bandits)
+
+- **Exploitation:** Show content the algorithm knows user likes (high engagement probability)
+- **Exploration:** Show new/different content to discover new interests (lower immediate probability, higher discovery potential)
+
+| Method | Approach |
+|--------|---------|
+| **Epsilon-Greedy** | With probability epsilon (~5-20%), show random content |
+| **Thompson Sampling** | Bayesian -- maintains probability distribution per item |
+| **Upper Confidence Bound** | Favors items with high uncertainty + high potential |
+
+TikTok: ~15-20% of FYP is exploratory content.
+
+### 9.5 Cold Start Problem
+
+| Type | Solutions |
+|------|----------|
+| **New User** | Onboarding quiz (TikTok asks interests), popular/trending as default, demographics as proxy, cross-platform data transfer (Meta shares between Instagram/Facebook/Threads) |
+| **New Content** | Content-based features, creator history, batch testing (TikTok's approach), NLP/CV topic matching |
+
+### 9.6 Embedding Spaces
+
+- Each user: vector of ~100-500 dimensions
+- Each content piece: similar vector
+- Distance (cosine similarity, dot product) = relevance
+- Transformers (BERT, GPT) revolutionized embedding quality with semantic understanding
+- Platforms use CLIP/ViT for image embeddings, SentenceTransformers for text
+
+---
+
+## 10. Cross-Platform Content Strategy
+
+### 10.1 Format Optimization by Platform
+
+| Platform | Favored Format | Penalized Format |
+|----------|---------------|-----------------|
+| **Instagram** | Reels, Carousels | External links, plain text |
+| **TikTok** | Vertical video (9:16), trending audio | Content with watermarks |
+| **YouTube** | Long-form (8-20 min), Shorts | Short videos (<4 min) in long-form |
+| **LinkedIn** | Documents/Carousels, long text | External links, corporate posts |
+| **X/Twitter** | Threads, images, short video | External links (penalized) |
+| **Facebook** | Reels, Group posts | Links, Page posts |
+| **Pinterest** | Vertical pins (2:3), infographics | Content without keywords |
+
+### 10.2 Cross-Posting vs Native vs Repurposing
+
+- **Cross-posting (same piece everywhere):** Efficient but 30-50% lower performance than native
+- **Native per platform:** Ideal but impractical for solo creators
+- **Repurposing (recommended):** Create "master" piece, extract clips, transform insights per platform
+
+### 10.3 Hook Patterns (First 2-3 Seconds)
+
+| Pattern | Example |
+|---------|---------|
+| **Curiosity Gap** | "Most people don't know that..." |
+| **Contrarian** | "Stop doing X (everyone does it wrong)" |
+| **Result-First** | "I made R$50K in 30 days. Here's how." |
+| **Story Hook** | "3 months ago, I was broke..." |
+| **Direct Challenge** | "If you're a [persona], you NEED to know this" |
+| **Pattern Interrupt** | Unexpected visual, scene change, quick cut |
+| **Social Proof** | "10 million views. Here's the secret." |
+
+### 10.4 Storytelling Structures for Social
+
+| Framework | Structure |
+|-----------|----------|
+| **PAS** | Problem > Agitate consequences > Solution |
+| **AIDA** | Attention (hook) > Interest (facts) > Desire (benefits) > Action (CTA) |
+| **BAB** | Before (pain) > After (pleasure) > Bridge (your solution) |
+| **Hero's Journey** | Normal life > Challenge > Search for solution > Transformation > New reality |
+
+### 10.5 Optimal Posting Times (Brazil, BRT)
+
+| Platform | Best Times | Best Days |
+|----------|-----------|-----------|
+| **Instagram** | 11h-13h, 19h-21h | Tue, Wed, Thu |
+| **TikTok** | 12h-14h, 19h-22h | Tue-Fri |
+| **YouTube** | 14h-17h (upload) | Thu, Fri, Sat |
+| **LinkedIn** | 7h-9h, 12h | Tue, Wed, Thu |
+| **X/Twitter** | 9h-12h | Mon-Fri |
+| **Facebook** | 9h-11h, 13h-15h | Wed, Thu |
+
+**Caveat:** These are averages. Best time for any creator is determined by their own analytics.
+
+---
+
+## 11. Engagement Mechanics
+
+### 11.1 Engagement Rate Formulas
+
+**Instagram:**
+```
+ER (Post) = (Likes + Comments + Saves + Shares) / Followers x 100
+ER (Reach-based) = (Likes + Comments + Saves + Shares) / Reach x 100  [more accurate]
+```
+
+**Benchmarks Instagram (2025):**
+| Tier | Followers | ER |
+|------|----------|-----|
+| Nano | 1-10K | 3-5% |
+| Micro | 10-100K | 1.5-3% |
+| Mid | 100K-500K | 1-2% |
+| Macro | 500K-1M | 0.8-1.5% |
+| Mega | 1M+ | 0.5-1% |
+
+**TikTok:** `ER = (Likes + Comments + Shares) / Views x 100` -- 3-9% normal, >10% excellent
+**YouTube:** `ER = (Likes + Comments) / Views x 100` -- 3-7% good
+**LinkedIn:** `ER = (Reactions + Comments + Reposts) / Impressions x 100` -- 2-5% good, >5% excellent
+
+### 11.2 Vanity vs Actionable Metrics
+
+| Vanity (look good, don't indicate value) | Actionable (indicate real value) |
+|------------------------------------------|----------------------------------|
+| Total followers (without quality) | **Saves** (reference content) |
+| Total likes (easily inflated) | **Shares/DM sends** (genuine recommendation) |
+| Impressions (don't indicate interest) | **Quality comments** (emotional/intellectual connection) |
+| Reach without engagement context | **Profile visits** after content view |
+| | **Follower-to-lead ratio** (B2B) |
+| | **Revenue per follower** |
+
+### 11.3 Saves & Shares as Quality Signals
+
+Both Instagram and TikTok confirmed (2024-2025) that saves and shares outweigh likes.
+
+- **Like:** Low effort, quasi-reflexive, casual
+- **Comment:** Medium effort, can be superficial
+- **Save:** User wants to access again = lasting practical/emotional value
+- **Share (DM):** User recommending to someone specific = strongest genuine value signal
+
+**Implication:** Create "saveable" (tutorials, lists, frameworks, infographics) and "shareable" (relatable, surprising, useful to others) content rather than just "likeable" (beautiful but substanceless).
+
+### 11.4 Community Building Strategies
+
+1. Respond to comments systematically (especially first hours)
+2. Create participatory content (questions, challenges, polls)
+3. Name the community (give identity to follower group)
+4. Consistency creates expectation and habit
+5. "Insider" content (inside jokes, recurring references)
+6. DM engagement (respond genuinely -- strong algorithmic weight)
+
+### 11.5 DM Strategies
+
+DMs are the strongest relationship signal for most algorithms:
+- Instagram: frequent DM exchanges = content prioritized in feed
+- TikTok: DM shares = strongest signal for Reels
+- LinkedIn: InMail/DM interactions influence content ranking
+
+**Tactics:**
+- Respond to every genuine DM
+- Send exclusive content via DM (Instagram broadcast channels)
+- Use Stories to encourage DM replies
+- Create "close friends" lists with most engaged followers
+
+---
+
+## 12. Creator Economy & Monetization
+
+### 12.1 Market Overview
+
+Creator economy: ~$250B globally (2024, Goldman Sachs), projected $480B by 2027. Brazil influencer marketing: R$2.18B (2024).
+
+**Income distribution (extremely unequal):**
+- Top 1% earn ~80% of total revenue
+- ~2M creators earn >$100K/year globally
+- ~46M creators are "amateur" (<$1K/year)
+- Median full-time creator income: ~$50K/year
+
+### 12.2 Platform Monetization Programs
+
+| Platform | Program | Requirements | Revenue Model |
+|----------|---------|-------------|---------------|
+| **YouTube** | YPP | 1K subs + 4K hours (or 10M Shorts views/90 days) | 55% revenue share (long-form), 45% Shorts. CPM $2-30 |
+| **Instagram/Facebook** | Subscriptions, Badges, Reels Bonus | Varies | Subscriptions for exclusive content. Meta less generous in direct share |
+| **TikTok** | Creativity Program Beta | Videos >1 min | Pays per qualified views. RPM historically low ($0.02-0.05/1K views old Creator Fund) |
+| **LinkedIn** | None (direct) | N/A | Indirect: leads, clients, consulting |
+| **X/Twitter** | Ads Revenue Sharing | Premium + 5M impressions/3 months | Subscriptions, Tips |
+
+### 12.3 Brand Deals -- Brazilian Benchmarks (2025)
+
+| Tier | Followers | Average Price per Post (Instagram) |
+|------|----------|-----------------------------------|
+| Nano | 1-10K | R$200-1,000 |
+| Micro | 10-100K | R$1,000-5,000 |
+| Mid | 100K-500K | R$5,000-20,000 |
+| Macro | 500K-1M | R$20,000-80,000 |
+| Mega | 1M+ | R$80,000-500,000+ |
+
+### 12.4 Creator Tools Ecosystem
+
+| Category | Tools |
+|----------|-------|
+| **Video editing** | CapCut (dominant), Premiere Pro, DaVinci Resolve |
+| **Design** | Canva (dominant in Brazil), Figma |
+| **Scheduling** | Later, Buffer, Hootsuite, Metricool |
+| **Analytics** | Iconosquare, Sprout Social, Metricool |
+| **Link in Bio** | Linktree, Stan Store, Beacons |
+| **Monetization** | Hotmart, Eduzz, Patreon, Ko-fi |
+| **AI Content** | ChatGPT, Claude, Opus Clip (repurposing) |
+| **CRM/Influencer** | Squid (Brazil), Influency.me, CreatorIQ |
+
+---
+
+## 13. Social Commerce
+
+### 13.1 Market Overview
+
+Global social commerce: $1.2T (2024), projected $2.9T by 2026 (Accenture). China: >15% of total e-commerce. Brazil: regional leader in Latin America.
+
+### 13.2 Platform Shopping Features
+
+| Platform | Features | Notes |
+|----------|----------|-------|
+| **Instagram Shopping** | Product tags in posts/Stories/Reels, Shop tab, Collections | Algorithm prioritizes tagged content in shopping surfaces |
+| **TikTok Shop** | Product integration in videos/lives, affiliate marketplace, native checkout | Launched in Brazil 2024. ByteDance invested billions |
+| **Pinterest Shopping** | Product Pins with prices, visual search, Shopping ads | 83% of users plan purchases on Pinterest |
+| **WhatsApp Business** | Product catalog + WhatsApp Pay (Pix) | Brazil "last mile" of conversion |
+
+### 13.3 Live Commerce
+
+Growing format. In China: $500B in sales (2023). In Brazil, accelerating:
+- Shopee Live: largest Brazilian live commerce operation
+- Instagram Lives with product tags
+- TikTok Lives (growing rapidly)
+- Mercado Livre Lives (launched 2024)
+
+**Success factors:** Urgency (limited offers), interactivity (real-time Q&A), practical demonstration, trust in presenter, exclusive discounts.
+
+---
+
+## 14. AI & Social Media
+
+### 14.1 AI Content Generation Tools
+
+| Type | Tools |
+|------|-------|
+| **Text** | ChatGPT, Claude, Gemini, Jasper, Copy.ai, native platform AI (Instagram captions, LinkedIn assistant) |
+| **Image** | Midjourney, DALL-E 3, Stable Diffusion, Canva Magic Design, Adobe Firefly |
+| **Video** | Runway ML, Pika, Opus Clip/Vidyo.ai (repurposing), Synthesia/HeyGen (AI avatars) |
+| **Audio** | ElevenLabs (voice cloning), Suno/Udio (music generation), Podcastle (AI editing) |
+
+### 14.2 AI-Powered Analytics
+
+- **Metricool AI:** Suggests optimal times based on history
+- **Lately AI:** Analyzes long content and generates social posts
+- **Sprout Social AI:** Sentiment analysis and trend detection
+- **Exploding Topics:** Identifies exponentially growing topics
+
+### 14.3 Deepfakes & Synthetic Media
+
+- Platforms invest in detection models (Meta's Video Authenticity, Google's SynthID)
+- Meta, YouTube, TikTok require AI-generated content labeling
+- EU AI Act and Brazilian AI Bill regulating deepfakes
+- Electoral deepfakes a growing risk, especially in Brazil
+
+### 14.4 AI Moderation
+
+- >95% of violating content removed by AI before human reports (Meta)
+- Limitations: difficulty with sarcasm, cultural context, and Brazilian Portuguese regional slang nuances
+- "Borderline" content can have reduced distribution without removal (shadowban)
+
+---
+
+## 15. Brazilian Context
+
+### 15.1 Social Media Landscape (2025)
+
+| Metric | Value |
+|--------|-------|
+| Population | ~216M |
+| Internet users | ~183M (86.2%) |
+| Social media identities | ~144M (67.8%) |
+| Daily social media time | 3h49min (world's highest) |
+| Gender split | 55.8% female, 44.2% male |
+
+**Platform MAU in Brazil:**
+
+| Platform | MAU (Brazil) | Notes |
+|----------|-------------|-------|
+| WhatsApp | ~169M | 99% of smartphones |
+| YouTube | ~142M | |
+| Instagram | ~141M | #1 content platform |
+| TikTok | ~131M | From 83M (2023) to 131M (2025) |
+| Facebook | ~109M | |
+| LinkedIn | ~75M | |
+| Telegram | ~75M | Grew post-X blocks (2024) |
+| Pinterest | ~38M | |
+| Threads | ~34M (est.) | |
+| X/Twitter | ~22M | Dropped after judicial block (2024) |
+
+### 15.2 WhatsApp Dominance
+
+99% smartphone penetration. Used for: personal communication, business, customer service, sales, payments (Pix via WhatsApp), news. 5M+ small businesses use WhatsApp Business.
+
+**Marketing implication:** In Brazil, WhatsApp is frequently the "last mile" of conversion. Journeys start on Instagram/TikTok and end with "chama no zap." Strategies ignoring WhatsApp are incomplete.
+
+### 15.3 Creator Economy Brasileira
+
+Brazil has the 3rd largest creator economy (after US and China):
+- 500K+ professional content creators
+- R$2.18B influencer marketing market (2024)
+- Management platforms: Squid (acquired by Locaweb), Influency.me, Airfluencers
+- Multi-platform presence is the norm (Instagram + TikTok + YouTube)
+- Strong humor and Brazilian cultural references
+- Creator events: VidCon Sao Paulo, Influencer Conference
+
+### 15.4 CONAR & #publi Regulation
+
+- CONAR requires clearly identified advertising (#publi, #ad, #patrocinado)
+- Article 36 Consumer Defense Code prohibits hidden/misleading advertising
+- Identification must be "immediate and conspicuous" (not buried in hashtags)
+- Stories/Lives need identification per slide/verbal+visual
+- Fines: CONAR up to R$250,000; Procon additional fines for misleading advertising
+- LGPD applies to follower data collection
+
+### 15.5 Brazilian Benchmarks
+
+**CPM (Cost Per Thousand Impressions) -- Ads:**
+
+| Platform | Average CPM (Brazil) |
+|----------|---------------------|
+| Instagram Feed | R$15-35 |
+| Instagram Stories | R$10-25 |
+| Instagram Reels | R$8-20 |
+| TikTok In-Feed | R$5-15 |
+| YouTube Pre-Roll | R$20-50 |
+| Facebook Feed | R$8-20 |
+| LinkedIn | R$40-100 |
+
+**Organic Engagement Rate:**
+
+| Platform | Brazil ER | Global ER |
+|----------|----------|-----------|
+| Instagram | 1.5-3% | 1-2% |
+| TikTok | 5-9% | 4-8% |
+| LinkedIn | 2-4% | 1.5-3% |
+| YouTube | 3-5% | 2-4% |
+| Facebook | 0.5-1% | 0.3-0.8% |
+
+Brazil consistently shows above-global-average engagement rates, reflecting highly social and participative user behavior.
+
+### 15.6 PT-BR Content Nuances
+
+- **Regional slang:** Algorithm NLP may not capture regional nuances. Use colloquial but nationally accessible Portuguese
+- **Emojis:** Brazilians use extensively. Posts without emojis may seem "cold"
+- **Tone:** Formal on LinkedIn, casual/authentic on Instagram and TikTok
+- **Hashtags:** PT-BR hashtags reach Brazilian audience; English reaches global. Hybrid strategy recommended
+- **Brazilian humor:** Memes, self-deprecation, comic exaggeration culturally strong and high-engagement
+- **Brazilian music:** Trending Brazilian songs on TikTok/Reels can generate local boost
+- **Key dates:** Carnaval, Festa Junina, Black Friday (extremely strong in Brazil), Dia das Maes
+
+---
+
+## 16. Key People & References
+
+### 16.1 Essential Thinkers
+
+| Person | Contribution | Key Work |
+|--------|-------------|----------|
+| **Eli Pariser** | Coined "filter bubble" | "The Filter Bubble" (2011) |
+| **Tristan Harris** | "Race to the bottom of the brainstem" | "The Social Dilemma" (Netflix, 2020) |
+| **Eugene Wei** | "Status as a Service", "TikTok and the Sorting Hat" | Essays (2019, 2020) |
+| **Nir Eyal** | Hook model (Trigger > Action > Variable Reward > Investment) | "Hooked" (2014) |
+| **Jonah Berger** | STEPPS framework for virality | "Contagious" (2013) |
+| **Gary Vaynerchuk** | "Jab Jab Jab Right Hook" framework | "Jab, Jab, Jab, Right Hook" (2013) |
+| **Li Jin** | "100 True Fans" (update of Kevin Kelly's 1,000) | "The Passion Economy" (2020) |
+| **Chris Dixon** | "Read Write Own" -- Web evolution framework | "Read Write Own" (2024) |
+| **Jaron Lanier** | Critique of ad-based social media model | "Ten Arguments..." (2018) |
+| **Casey Newton** | Leading journalist on platform decisions | Platformer newsletter |
+
+### 16.2 Essential Books
+
+| Book | Author | Year | Core Topic |
+|------|--------|------|-----------|
+| "The Filter Bubble" | Eli Pariser | 2011 | Algorithmic personalization critique |
+| "Hooked" | Nir Eyal | 2014 | Habit-forming products / social engagement |
+| "Contagious" | Jonah Berger | 2013 | Science behind virality (STEPPS) |
+| "Jab Jab Jab Right Hook" | Vaynerchuk | 2013 | Platform-specific content strategy |
+| "No Filter" | Sarah Frier | 2020 | Internal Instagram history |
+| "The Age of Surveillance Capitalism" | Shoshana Zuboff | 2019 | Data economy and platform surveillance |
+| "Read Write Own" | Chris Dixon | 2024 | Decentralized future of platforms |
+| "Influence" | Robert Cialdini | 2006 | 6 persuasion principles applied to social |
+
+### 16.3 Essential Academic Papers
+
+| Paper | Authors | Year | Contribution |
+|-------|---------|------|-------------|
+| Deep Neural Networks for YouTube Recommendations | Covington et al. (Google) | 2016 | YouTube two-stage recommendation architecture |
+| Monolith: Real Time Recommendation System | Liu et al. (ByteDance) | 2022 | TikTok recommendation architecture |
+| Instagram Explore Recommender System | Medvedev et al. (Meta) | 2019 | Two-tower model for Explore |
+| Attention Is All You Need | Vaswani et al. (Google) | 2017 | Transformer architecture -- base for modern recommendation |
+| The Spread of True and False News Online | Vosoughi et al. (MIT) | 2018 | Fake news spreads 6x faster than truth on social media |
+| Algorithmic Amplification of Politics on Twitter | Huszar et al. (Twitter) | 2021 | Internal study on political content amplification |
+
+---
+
+## 17. Actionable Checklists
+
+### Platform-Specific Content Optimization
+
+**Instagram:**
+```
+[ ] Reels: Hook in first 3 seconds (visual + text)
+[ ] Reels: Optimize for watch-through (keep viewers to end)
+[ ] Carousels: 5-10 slides with actionable content (save-worthy)
+[ ] Stories: Use interactive elements (polls, quizzes, questions)
+[ ] Post when followers are online (check Insights)
+[ ] Respond to DMs to strengthen relationship signal
+[ ] No TikTok watermarks on Reels
+[ ] Create original content (avoid pure reposting)
+```
+
+**TikTok:**
+```
+[ ] Hook in first 2 seconds (high skip rate = suppression)
+[ ] Optimize for completion rate + total watch time
+[ ] Use trending audio when relevant
+[ ] Add on-screen text (indexed for categorization)
+[ ] Skip #fyp/#foryou hashtags (irrelevant)
+[ ] Use niche hashtags for initial categorization
+[ ] Create for shares (DM shares = strongest signal)
+```
+
+**YouTube:**
+```
+[ ] Thumbnail: High contrast, expressive face, short text
+[ ] Title: Keyword at start, < 60 chars, curiosity element
+[ ] Description: 200+ words, keyword in first 100 chars, timestamps
+[ ] Optimize for both CTR AND watch time (avoid clickbait)
+[ ] First 30 seconds must retain viewers
+[ ] Cards and end screens to extend session time
+[ ] Consider thumbnail A/B testing
+```
+
+**LinkedIn:**
+```
+[ ] Write long-form text that rewards reading time (dwell time is #1 signal)
+[ ] Use document carousels for educational content
+[ ] Avoid external links in post (native content only)
+[ ] Engage meaningfully in comments (50+ characters)
+[ ] Limit to 1-3 highly relevant hashtags
+[ ] Leverage newsletter feature for maximum organic reach
+[ ] Post during business hours (7-9h, 12h BRT)
+```
+
+### Engagement Quality Assessment
+
+```
+[ ] Are saves/shares growing (not just likes)?
+[ ] Are comments substantive (not just emojis)?
+[ ] Are profile visits increasing after content view?
+[ ] Is follower-to-lead ratio improving (B2B)?
+[ ] Are DM conversations increasing?
+[ ] Is engagement velocity strong in first 30-60 minutes?
+[ ] Is content being referenced/cited by others?
+```
+
+### Cross-Platform Repurposing Workflow
+
+```
+1. Create master content piece (long video, podcast, or article)
+2. [ ] Extract 5-10 clips for Reels/TikTok/Shorts (add native captions)
+3. [ ] Transform key insights into Instagram/LinkedIn carousels
+4. [ ] Create text threads for X/Twitter
+5. [ ] Write newsletter edition from key takeaways
+6. [ ] Design 5-10 quote cards for Stories
+7. [ ] Adapt tone/format for each platform culture
+8. [ ] Schedule across platforms at optimal times per analytics
+```
+
+### Brazilian Market Checklist
+
+```
+[ ] Content in Brazilian Portuguese (not Portugal Portuguese)
+[ ] WhatsApp strategy integrated (last mile of conversion)
+[ ] #publi disclosure on ALL sponsored content
+[ ] LGPD compliance for data capture
+[ ] Cultural calendar planned (Carnaval, Black Friday, Dia das Maes)
+[ ] Emojis used appropriately for Brazilian audience expectations
+[ ] Brazilian trending audio/music considered for Reels/TikTok
+[ ] Regional nuances tested (humor style, slang level)
+[ ] PIX integration for micro-transactions if applicable
+```
+
+### Algorithm Health Monitoring
+
+```
+Monthly:
+[ ] Review reach trends per content type
+[ ] Compare engagement rate vs platform benchmarks
+[ ] Check for distribution anomalies (possible restriction)
+[ ] Audit content against known penalties (watermarks, engagement bait)
+[ ] Review posting times vs follower activity patterns
+
+Quarterly:
+[ ] Check for algorithm updates from official platform sources
+[ ] Review format performance shifts (carousels vs Reels vs static)
+[ ] Assess signal weight changes (saves vs likes vs shares)
+[ ] Update content strategy based on platform direction
+```
+
+---
+
+> **Source research:** MS-007 Social Algorithms Master System | 42 sources | April 2026
+> **Compiled for:** squad-content knowledge-base | SINAPSE AI