Case · M4 × Pay (e-commerce)
Product recommendations: the 35% claim vs a controlled test
A widely repeated estimate says recommendations drive 35% of Amazon's purchases. When researchers switched recommendations on and off in a large experiment on a major fashion retailer's site, overall sales rose about 11%: a real effect, far smaller than the claim (Kumar & Hosanagar, 2018).
The outcome · 2018
Recommendations raised sales of recommended items 9% while sales of the originally viewed items fell 1.9%; overall product sales rose 11% vs a no-recommendation control, across 2M+ product page views (Kumar & Hosanagar, 2018). The 35% Amazon figure is a disputed estimate.
Stage cell: M4 × Pay (e-commerce).
What they did. Researchers compared shoppers who saw product recommendations with a control group on the same site who saw none, over 2M+ product page views from nearly a million sessions.
The mechanism. Chou credits Amazon's personalization with the Alfred Effect: recommendations good enough that switching feels unthinkable GT#83 (Actionable Gamification Ch.8, p.190). Hooked's version is personalization powered by stored behavior HK-INV08 (Hooked Ch.5). The test shows why attribution needs a holdout F26.
The outcome. +11% overall sales (2018) Case grade A. From a regulator, a public filing or a peer-reviewed study.: +9% for recommended items, -1.9% for the items originally viewed. Chou's book cites a ~30% share of Amazon sales (a 2012 press claim), and the commonly cited 35% is an estimate the study's authors called hard to believe.
What it means at your stage. Measure recommendation impact with a holdout, not click-through attribution, and count cannibalization of the items people came for.
Caveats. One retailer, not Amazon; fashion category. Mechanism link Evidence grade B. Moderate evidence: solid studies with some limits..
Would you have called it?
This case is also a round of Predict the Outcome: read the situation, pick what you think happened, then see the evidence.
Play this roundTry this
The techniques behind it
Each one is a starting point, not a recipe. Gamify It turns a technique into an experiment card for your product: the change, the metric, the guardrail and how long the test needs at your traffic.
- TechniqueEvidence grade D. Practitioner framework or anecdote: plausible, untested.
Alfred Effect (GT#83)
visible personalization
Personalization so well fitted to a user that switching feels unthinkable, even when a rival is better or cheaper. Chou stresses making it noticeable: many users never realize a product has learned them before they churn.
- TechniqueEvidence grade C. Limited evidence: few or small studies.
Store value that makes the product better with use
a product that improves as you use it
Design the investments users make so they build up in forms that improve the next visit: content they collect, data about themselves, people they follow or who follow them, and skill with the tool. Say plainly when a recommendation comes from what they stored.
Limits
What we don’t claim
- The Hook Model is a practical design tool, not a validated or complete theory of behavior.Hooked, Introduction
- Octalysis is a practitioner framework built from observing games, not a validated scientific model.Actionable Gamification Ch.15, pp.399-400
- No success is attributed to a single technique: real outcomes have many causes.
- Company names are used editorially; no endorsement is implied.
Sources
Actionable Gamification · Yu-kai Chou
Ch.6, p.98 · Ch.8, p.190
Hooked · Nir Eyal
Ch.5
How valuable are online product recommendations? · UF Warrington College of Business (Kumar & Hosanagar, Information Systems Research), 2018
Case grade A. From a regulator, a public filing or a peer-reviewed study.Evidence grade B. Moderate evidence: solid studies with some limits.
Source IDs: CASE-M06GT#83HK-INV08F26
Book content is paraphrased, never quoted at length. Last verified October 24, 2026.
Your product
Find the stage your product is stuck in
Your product has its own version of this story. Find the stage it’s in and the move that fits it. Connecting is free; a Pro subscription starts it.