How does a social media feed algorithm work?

Social media feed algorithms use a multi-stage pipeline called the CandidatePipeline. It starts by retrieving candidate posts from both accounts you follow (In-Network via Thunder) and discovery content (Out-of-Network via Phoenix), then filters for safety and freshness, scores engagement probability across 18 signals, applies diversity rules, and selects top-ranked posts for your feed.

What is the CandidatePipeline in social media algorithms?

The CandidatePipeline is the modular foundation of a social media recommendation system. It defines specific stages—from content discovery through retrieval engines, safety classification, pre-filtering, multi-step engagement scoring, and final selection—that every piece of content must pass through before appearing in a user's feed.

Why do some posts go viral while others don't?

Viral posts overcome the Out-of-Network Penalty (0.7 weight factor) through massive positive engagement signals. The algorithm scores 18 engagement signals including likes, replies, retweets, shares, and dwell time. Posts with overwhelming positive engagement can mathematically overcome negative signals like blocks or reports, while author diversity decay prevents any single account from dominating feeds.

Does posting more often help with social media algorithms?

No. Due to Author Diversity Exponential Decay, flooding the platform with posts yields diminishing returns. The algorithm applies score decay to multiple posts from the same author, with a floor parameter of 0.3 ensuring later posts retain only 30% of their score. The system prioritizes a few high-impact posts over high-volume spam.

How does quote tweeting affect content distribution?

Quote tweets are counted as a positive engagement signal by the algorithm regardless of intent. The model cannot distinguish between quoting to criticize and quoting to endorse, so outrage sharing and angry quote tweets effectively boost a post's distribution score, creating a feedback loop that amplifies controversial content.

Digital Literacy 12 Min Read 2,850 Words By Robert McCullock

The Journey of a Post:
A Behind-the-Scenes Tour of Your Feed

Every time you refresh your feed, a high-stakes digital audition occurs in milliseconds. Here's exactly how the machine decides what you see.

⬡ Millions of Posts — Global Content Corpus

▼

⬡ Thousands Retrieved — Thunder + Phoenix

▼

⬡ Hundreds Filtered — Safety + Freshness

▼

⬡ Scored + Ranked — 18 Engagement Signals

▼

✦ Your Feed — Top-K Selected

The Digital Funnel

Introduction — CandidatePipeline

Behind the screen, a sophisticated system called the CandidatePipeline is working to sort through hundreds of millions of potential posts to find the few dozen that actually reach your eyes. Think of this process as a massive funnel: it starts with a vast ocean of global content and narrows it down through layers of retrieval, safety checks, and complex mathematical predictions.

Key Concept

CandidatePipeline — The modular foundation of the recommendation system. It defines the specific stages—from discovery to final ranking—that a piece of content must pass through before it is served to a user. Understanding this pipeline is the key to seeing "under the hood" of your digital experience.

Finding the Candidates

Stage 1 — Retrieval

The first step is gathering the raw materials. The system uses two distinct discovery engines to pull posts from a corpus of millions, narrowing them down to a manageable set of thousands.

Engine	Focus	Technology
⚡ Thunder	In-Network — Posts from accounts you follow	High-speed `DashMap` (concurrent HashMap) indexing the last 48 hours of content for instant `O(1)` lookups
🔥 Phoenix	Discovery — Relevant posts from the entire platform	Two-Tower neural network using L2-normalized embeddings and Dot Product Similarity with `ANN` (Approximate Nearest Neighbor) searches

⬡ The Insight

The primary benefit of this stage is massive reduction. These engines transform an impossible-to-process corpus of millions into a manageable set of thousands that are semantically relevant to your interests.

Checking the Guest List

Stage 2 — Classification

Once those thousands of candidates are gathered, the system must determine the relationship between you and the author. This is handled by the InNetworkCandidateHydrator.

In-Network Label

Applied if the post is from someone you follow or if it is your own post.

Out-of-Network Label

Applied to discovery content found by Phoenix from accounts you do not follow.

⬡ The "So What?"

By labeling posts here, the system ensures that "followed" accounts are treated correctly regardless of which engine found them. This allows the scoring rules to evolve independently of the discovery engines, ensuring your connections always get the "followed" status they deserve.

The Quality Control Filter

Stage 3 — Pre-Filtering

Before the system invests heavy computing power into ranking, it performs a Safety and Freshness Checklist. If a post fails any of these sequential filters, it is dropped immediately. Click each checkbox to trace through the pipeline.

Sequential Filter Chain

The Talent Show

Stage 4 — Multi-Step Scoring

The remaining candidates now enter the most competitive phase: a four-step sequence where they are scored based on predicted engagement. The Phoenix Scorer uses a Grok-based transformer to predict the probability of 18 specific engagement signals.

14 Positive Signals

Favorite (Like) Probability
Reply Probability
Retweet (Repost) Probability
Photo Expand Probability
Click Probability
Profile Click Probability
Video Quality View (Completion)
Share Probability
Share via DM Probability
Share via Copy Link Probability
Dwell (Stop Scrolling) Probability
Quote Tweet Probability
Quoted Click Probability
Follow Author Probability

4 Negative Signals

Not Interested Click
Block Author Probability
Mute Author Probability
Report Post Probability

Plus a continuous metric: Dwell Time (total seconds viewed)

⬡ Author Diversity — Exponential Decay

If one author has multiple posts in your top list, the system applies a decay. However, a floor parameter of 0.3 ensures that even a prolific author's later posts retain 30% of their score, balancing variety with content relevance.

⬡ Out-of-Network Penalty

Posts from accounts you don't follow are multiplied by an OON_WEIGHT_FACTOR of 0.7. This gives In-Network content a home-field advantage; discovery content must be significantly more engaging to beat out your follows.

⬡ Batch Independence — Attention Mask

To keep scoring fair, the system uses an "attention mask." Imagine a teacher grading a test in a vacuum: the mask ensures a post's score is based solely on your history and the post's quality, prevented from being influenced by other posts in the same batch during calculation.

The Final Cut

Stage 5 — Selection & Visibility

The TopKScoreSelector selects the highest-scoring winners. Before they reach you, two final safety checks occur.

Visibility Filtering (VFFilter)

Queries an external service to check for spam or policy violations. The system runs safety checks for "Home" (In-Network) and "Recommendations" (OON) in parallel to minimize latency.

Conversation Deduplication

Identifies "threads" and keeps only the highest-scored post from any single conversation, preventing your feed from being cluttered by a single long debate.

⬡ Pro Tip

Because Phoenix re-scores content on every request using current engagement stats, a post's rank can shift dramatically as it gains traction in real-time.

▣

Why Content Wins or Loses

Summary — Critical Takeaways

Understanding the pipeline reveals critical takeaways for navigating and creating content. Your feed is not a random stream—it's a meticulously filtered, ranked, and deduplicated reflection of the platform's mathematical predictions for your attention.

Takeaway 01

Quality > Quantity

Due to Author Diversity decay, flooding the platform with posts yields diminishing returns. The system prioritizes a few high-impact posts over high-volume spam.

Takeaway 02

The Math of "Outrage"

A post with 10,000 likes (weighted at 30) easily overcomes 100 blocks (weighted at -100). Sheer volume of engagement can drown out negative signals.

Takeaway 03

The Quote-Tweet Loop

The algorithm counts an "angry" quote-tweet as a positive Quote Score. The model cannot distinguish criticism from endorsement—outrage sharing boosts distribution.

Takeaway 04

No "Velocity" Multiplier

Contrary to popular belief, there is no speed multiplier for likes. Distribution is driven by Recency (Thunder's 48-hour index) and Semantic Matching (Phoenix).