Feed Ranking Algorithm Design Template

What This Template Is For

A feed ranking algorithm decides which items appear first when a user opens your product. It is the most consequential piece of code in any content-driven application. The ranking directly controls engagement, retention, and user satisfaction. A bad feed buries relevant content under noise. A good feed surfaces the right thing at the right time without the user having to search for it.

This template provides a structured approach to designing feed ranking systems for SaaS products, internal tools, social platforms, and collaboration apps. It covers the core ranking formula, signal taxonomy, diversity constraints, recency decay functions, abuse countermeasures, and A/B testing methodology. If you need to design the content discovery layer that sits alongside the feed, the Content Discovery Template covers recommendation surfaces and cold-start strategies. For understanding the retrieval layer that feeds candidates into the ranker, the Search Ranking Template covers scoring and relevance models.

The Technical PM Handbook offers guidance on working with ML engineers to scope ranking projects and define success metrics that balance engagement with user wellbeing.

How to Use This Template

1. Define the feed's purpose. Is this a social feed (maximize engagement), a work feed (maximize productivity), a news feed (maximize informed decisions), or a notification feed (maximize action completion)? The purpose shapes every ranking decision.
2. List the candidate item types. A feed often blends multiple content types: posts, comments, updates, announcements, tasks, recommendations. Each type needs its own scoring considerations.
3. Define the signals. Separate them into engagement signals (clicks, likes, comments), quality signals (author reputation, content completeness), relevance signals (topic match, relationship strength), and contextual signals (time of day, device, user state).
4. Design the scoring function. Start simple (weighted linear combination of signals) and add complexity only when measurement shows diminishing returns.
5. Set diversity and fairness constraints. Without constraints, the feed devolves into showing the same popular content types repeatedly. Define rules for type diversity, author diversity, and freshness.
6. Plan the A/B testing framework. Every ranking change should be testable. Define your North Star metric, guardrail metrics, and minimum detectable effect before shipping experiments.

The Template

Feed Context

Current state. [How is the feed currently ranked? Reverse chronological? Simple popularity? No ranking?]

Problems.

• [Problem 1: e.g., Users miss important updates buried under low-value posts]
• [Problem 2: e.g., Engagement concentrated on top 10% of content creators]
• [Problem 3: e.g., Feed feels stale within 2 hours of last visit]

Goals.

• [Goal 1: e.g., Increase daily feed sessions from X to Y]
• [Goal 2: e.g., Increase content breadth (unique authors engaged per session)]
• [Goal 3: e.g., Reduce time-to-first-meaningful-action to under X seconds]

Ranking Architecture

Pipeline stages.

1. Candidate Generation
   Source pool: all items from followed entities, teams, topics
   Window: last [N] hours or [N] items per source
   Output: ~[N] candidates

2. Pre-Scoring Filter
   Remove: blocked users, muted topics, already-seen items, policy violations
   Output: ~[N] candidates

3. Scoring
   Apply ranking formula to each candidate
   Output: scored and sorted list

4. Post-Scoring Rules
   Apply diversity constraints, deduplication, pinned items
   Output: final ranked feed

5. Delivery
   Paginate (page size [N])
   Track impressions for feedback loop

Signal Taxonomy

Engagement signals (how users interact):

Quality signals (content attributes):

Relevance signals (user-item fit):

Scoring Function

Approach. [Choose one]

• ☐ Weighted linear: score = w1engagement + w2quality + w3relevance + w4recency
• ☐ Two-stage: pointwise ML model (logistic regression / gradient-boosted tree) for P(engagement), then apply business rules
• ☐ Multi-objective: separate models for P(click), P(like), P(comment), combined with scalarization weights
• ☐ Learning-to-rank: listwise model (LambdaMART) trained on user sessions

Formula (if weighted linear):

score = (engagement_score * 0.3)
      + (quality_score * 0.2)
      + (relevance_score * 0.3)
      + (recency_score * 0.2)

Recency decay function.

Chosen function. [Which one and why. Include the specific parameter values.]

Diversity and Fairness Rules

Pinned items.

• [System announcements: always position 1]
• [Admin pinned: positions 1-3, max 1 pinned at a time]
• [User pinned: not applicable / bookmarked items section]

Abuse and Gaming Prevention

A/B Testing Framework

North Star metric. [The single metric that best captures feed health. e.g., "Meaningful interactions per DAU" or "Tasks completed from feed per week"]

Guardrail metrics (must not regress):

• [User satisfaction survey score]
• [DAU retention (D7, D28)]
• [Content creation rate (items per creator per week)]
• [Report/hide rate per 100 impressions]

Experiment methodology.

Experiment backlog.

Monitoring and Observability

Filled Example: B2B Team Collaboration Feed

Feed Context

Current state. Reverse chronological feed. Users report missing important updates because they are buried under low-value automated notifications. 67% of feed items are "status changed" events that users ignore.

Goals.

• Reduce missed actionable items from 23% to under 5%
• Increase feed engagement rate from 12% to 25%
• Decrease time-to-action on task mentions from 4.2 hours to under 1 hour

Scoring Function

score = (action_urgency * 0.35)
      + (relationship_strength * 0.25)
      + (content_quality * 0.15)
      + (engagement_velocity * 0.10)
      + (recency * 0.15)

Where:

• action_urgency: 1.0 if user is @mentioned or assigned, 0.7 if direct team project, 0.3 if followed project, 0.1 if org-wide
• relationship_strength: based on collaboration frequency with author (messages, shared tasks, meetings)
• content_quality: human-authored > automated status change. Posts with context (> 50 words) scored higher
• engagement_velocity: how quickly others on the team are engaging with this item
• recency: exponential decay, lambda = 0.05 (half-life ~14 hours)

Diversity Rules

• No more than 3 consecutive automated status updates
• At least 1 actionable item (mention, assignment, approval request) in first 5 items
• Announcements from leadership pinned for 48 hours
• Same project: max 5 of first 15 items (prevents one active project from flooding)

Common Mistakes to Avoid

• Starting with ML when rules would suffice. A well-tuned weighted linear formula with 5-8 signals outperforms a poorly trained ML model. Start simple, measure, and add complexity when the simple model plateaus. Use the RICE framework to prioritize which ranking improvements deliver the most impact.
• Optimizing for engagement without guardrails. Pure engagement optimization leads to addictive patterns that eventually hurt retention. Always pair engagement metrics with satisfaction surveys and negative-signal tracking (hides, unfollows, report rates).
• Ignoring feed freshness. A feed that shows the same items for 3 visits in a row feels dead. Users stop checking. Apply recency decay aggressively enough that the feed feels different every 4-6 hours.
• Not testing with a holdout. If you iterate on ranking without a permanent holdout group on the previous version, you lose the ability to measure cumulative long-term effects. Keep 5% of users on the baseline ranking indefinitely.
• Shipping without an A/B test. Every ranking change, no matter how small, should be tested. "We are pretty sure this is better" is not evidence. Run the experiment, wait for significance, then ship or revert.

Key Takeaways

• Start with a weighted linear scoring function before investing in ML models
• Define diversity rules to prevent content type and author concentration
• Use exponential recency decay to keep the feed feeling fresh
• Always A/B test ranking changes with a permanent holdout group
• Pair engagement metrics with satisfaction surveys and negative-signal tracking

About This Template

Created by: Tim Adair

Last Updated: 3/5/2026

Version: 1.0.0

License: Free for personal and commercial use