Amazon KDP Algorithm Deep Dive: How the Recommendation System Actually Works

Beyond Keywords: Amazon's Machine Learning Revolution

    Most KDP publishers still think Amazon's algorithm is about keywords and sales velocity. That was true in 2015.





    In 2025, Amazon's recommendation engine uses **machine learning models** that analyze dozens of signals you cannot directly see. Understanding these signals is the difference between books that Amazon actively promotes and books that die in obscurity.

The Core Insight

      Amazon does not just track what people buy. It tracks what people read, how they read it, when they abandon it, and what they buy next. The algorithm rewards books that create satisfied readers, not just sales.

The 180-Day Rolling Average System

    Amazon's algorithm does not treat all sales equally. It uses **time-decay weighting** where recent performance matters more than old performance.

How It Works

    Amazon tracks sales and engagement on a 180-day rolling window:



    - **Days 1-30:** Maximum weight (recent performance)
    - **Days 31-90:** Moderate weight (establishing patterns)
    - **Days 91-180:** Lower weight (historical baseline)
    - **Beyond 180 days:** Minimal influence on current recommendations

Strategic Implication

      A book with declining sales over 6 months will see Amazon recommendations slow down, even if it has strong lifetime sales. Momentum matters more than total volume.

Velocity vs. Volume

    **Scenario A:** Book sells 1,000 copies in Month 1, then 50/month for 5 months


    **Scenario B:** Book sells 200 copies per month consistently for 6 months





    Which performs better in recommendations? **Scenario B**, despite having the same total sales (1,250 vs 1,250). Consistent velocity signals sustained reader interest.

Completion Rate Tracking: The Hidden Metric

    This is where most publishers fail. Amazon tracks whether readers finish your book.

What Amazon Measures

    - **Kindle Unlimited page reads:** Exact page-by-page progress
    - **Time-to-completion ratios:** How long it takes readers to finish
    - **Abandon rates by position:** Where readers quit
    - **Session length patterns:** How long readers stay engaged per session
    - **Re-reading behavior:** Do readers return to sections?




    **Industry Average Completion Rate:** 40-50%


    **Target for Algorithmic Promotion:** 60%+


    **Elite Performance:** 70%+

Why Completion Rate Matters More Than Sales

    Amazon's ML models detect satisfaction through behavior, not reviews.

The Math of Consumption Weighting

      **Book A:** 2,000 sales, 30% completion rate = 600 satisfied readers


      **Book B:** 1,000 sales, 70% completion rate = 700 satisfied readers




      Amazon promotes Book B more aggressively despite lower sales volume. Satisfied readers buy more books.

Engagement Quality Signals

    Amazon's ML models analyze reading patterns to detect engagement quality:

1. Reading Velocity Patterns

    **Steady Progression:** Reader moves through book at consistent pace = High engagement signal


    **Erratic Patterns:** Long pauses, skipping chapters = Confusion or boredom signal


    **Rapid Skimming:** Very fast page turns without pauses = Low value signal

2. Session Length Analysis

    Amazon tracks how long readers stay in each session:



    - **Short sessions (under 5 min):** Low engagement
    - **Medium sessions (15-30 min):** Standard engagement
    - **Long sessions (60+ min):** High engagement (binge reading)

3. Highlight and Note Frequency

    Readers who highlight passages or take notes signal high value:



    - Books with 5+ highlights per reader = Actionable, valuable content
    - Books with 0-1 highlights = Entertainment or low practical value

The Also-Bought Network Effect

    Amazon's **item-to-item collaborative filtering** creates recommendation networks based on purchase patterns.

How Also-Boughts Form

    - Reader buys Book A
    - Within 30 days, reader buys Book B
    - Amazon creates association: A → B
    - After 50+ repeated patterns, B appears in "Customers also bought" for A

Strategic Opportunity: Series Architecture

      If you publish Book 1, 2, 3 in a series with strong read-through (60%+ of Book 1 readers buy Book 2), Amazon's algorithm creates automatic cross-promotion. Each new reader of Book 1 generates algorithmic momentum for the entire series.

Book Series Math

    **Scenario:** 3-book series with 65% read-through rate



    - 100 readers buy Book 1
    - 65 readers buy Book 2 (also-bought connection strengthens)
    - 42 readers buy Book 3 (trilogy completion signal)
    - Amazon now recommends all 3 books together in bundles




    **Result:** Amazon promotes your entire series as a package, multiplying discovery potential.

Author/Brand Reputation Scoring

    Amazon tracks author-level patterns across your entire catalog:

Positive Reputation Signals

    - **Consistent quality:** All books maintain 4.0+ average ratings
    - **Publication cadence:** Regular releases (1-2 books per quarter)
    - **Series completion:** Readers who finish series buy your other series
    - **Low return rate:** Few refund requests
    - **Cross-catalog purchases:** Readers buy multiple unrelated books from you

Negative Reputation Signals

    - **Inconsistent quality:** Mix of 4.5-star and 2-star books
    - **Abandoned series:** Book 1 published, but no Book 2 for 12+ months
    - **High return rate:** Readers request refunds frequently
    - **One-time readers:** People buy one book and never return

The Reputation Multiplier

      Once Amazon identifies you as a "quality author" (consistent 4.0+ ratings, strong completion rates, regular releases), new book launches receive algorithmic promotion from day 1. Your reputation becomes a launch multiplier.

Category and Keyword Optimization

    While ML signals dominate recommendations, categories still determine initial discoverability.

The Category Strategy

    Amazon allows **10 categories** per book (2 visible on product page, 8 backend):



    - **Primary category:** Your main competitive category (aim for Top 100)
    - **Secondary category:** Niche subcategory (aim for #1-10)
    - **Backend categories:** 8 additional keyword-driven categories for broader reach

Keyword Matching Evolution

    Amazon's keyword matching now uses **semantic search**, not just exact matches:



    - **Old system:** "weight loss" only matched "weight loss"
    - **New system:** "weight loss" also matches "fat burning," "body transformation," "lose pounds"




    **Implication:** Write naturally relevant keywords, not keyword-stuffed lists. Amazon's ML understands context.

The Golden Features: What Amazon Sees That We Cannot

    Machine learning models extract "golden features" — patterns invisible to humans but predictive of success:




    - **Title sentiment analysis:** Emotional tone of your title
    - **Description readability score:** Complexity of your book description
    - **Cover visual clustering:** How your cover compares to top sellers in category
    - **Price positioning:** Where your price falls in category distribution
    - **Reading level analysis:** Complexity of sample pages




    These features feed into Amazon's recommendation models. You cannot directly optimize for them, but understanding they exist shapes strategic decisions.

Optimization Strategies

1. Design for Completion

    - **Shorter is better:** 25,000-40,000 words have higher completion rates than 80,000+ words
    - **Front-load value:** Deliver actionable insights in first 3 chapters to hook readers
    - **Use clear structure:** Numbered frameworks, checklists, step-by-step guides
    - **Chapter cliffhangers:** End chapters with questions or teasers for next section

2. Build Series Momentum

    - **Complementary sequencing:** Each book solves next logical problem
    - **Back-matter CTAs:** Strong call-to-action at end of Book 1 for Book 2
    - **Rapid release:** Publish books 30-60 days apart to maintain momentum
    - **Series branding:** Consistent covers, titles, and keywords across series

3. Maintain Velocity

    - **Consistent release schedule:** Amazon rewards regular publishers
    - **Avoid long gaps:** 6+ months between releases hurts author reputation score
    - **Launch coordination:** Time releases to align with category trends

Common Mistakes

Mistake 1: Obsessing Over Reviews

    **Problem:** Focusing on review count instead of completion rate


    **Reality:** Amazon weighs completion signals more than review quantity


    **Solution:** Optimize for reader satisfaction (completion), reviews follow naturally

Mistake 2: Keyword Stuffing

    **Problem:** Using irrelevant keywords to game search


    **Reality:** ML models detect keyword-content mismatch and penalize rankings


    **Solution:** Use genuinely relevant keywords that match actual book content

Mistake 3: Ignoring Author Brand

    **Problem:** Publishing unrelated books with inconsistent quality


    **Reality:** Author reputation score affects all future launches


    **Solution:** Build coherent brand with consistent quality standards

The Future: AI Search and Rufus 2.0

    Amazon is rolling out **Rufus**, an AI shopping assistant that uses natural language to recommend products.

What Changes

    - **Conversational discovery:** "Find me books about X for Y audience" replaces keyword search
    - **Semantic matching:** AI understands book content, not just metadata
    - **Quality filtering:** AI prioritizes books with strong completion and satisfaction signals

Winning Strategy for AI Search

      Books with clear value propositions, strong completion rates, and genuine reader satisfaction will dominate AI recommendations. Gaming tactics will fail. Quality becomes the only sustainable strategy.

Action Plan

    - Audit current catalog for completion rate proxies (KU page reads vs sales)
    - Identify books with high completion and double down on similar content
    - Design new books for 60%+ completion (shorter, clearer, front-loaded value)
    - Build series with complementary sequencing for strong read-through
    - Maintain consistent publication cadence (1-2 books per quarter)
    - Monitor velocity trends and adjust pricing/marketing to maintain momentum

Next Steps

    - [Build an algorithmic brand](/learn/algorithmic-brand-guide) designed for Amazon's ML systems
    - [Master series architecture](/learn/book-series-success) for maximum read-through
    - [Generate books](/brand-builder) optimized for completion and engagement