Netflix doesn't show you my poster
Why Netflix shows every member a different poster for the same title, and the one rule the system obeys
Beat by beat.
| t (s) | Voice-over | On screen | Caption |
|---|---|---|---|
| 0–3 | Netflix doesn't show you the poster it shows me. | Two phones, same title, two artworks | SAME SHOW · DIFFERENT POSTER |
| 3–7 | I read Netflix's own engineering posts. The system obeys one rule. Guess it before the end. | Header of S1; “GUESS ↓” sticker | ONE RULE |
| 7–15 | Netflix's research found artwork was the biggest influence on what members choose, and people spent about 1.8 seconds on each title. [S2] | Quote card; 1.8 s timer | 1.8 SECONDS |
| 15–23 | So each title has a pool of artworks, and a contextual bandit picks one for you. It bets, watches, and bets again. [S1] | Slot machine whose arms are posters | IT BETS ON YOU |
| 23–30 | Watch a lot of romance? Good Will Hunting shows you Matt Damon and Minnie Driver. Watch comedies? You get Robin Williams. [S1] | Side-by-side, recreated, labelled ILLUSTRATION | YOUR HISTORY PICKS IT |
| 30–38 | It records, for every member, title and image, whether that image led to a play, and keeps the image stable so it knows what to credit. [S1] | Table: member · title · image → play | IT KEEPS SCORE |
| 38–46 | Here's the rule: no clickbait. Success is labelled by the quality of what you watch, not the click, and each artwork must represent the title. [S1] | “CLICKBAIT” struck through | NO CLICKBAIT |
| 46–54 | Brew principle: personalize the door, never the room. Optimize the second minute, not the first click. | Monochrome card + ProductBrew object | PERSONALIZE THE DOOR |
| 54–60 | Next: Airbnb learned your taste from a single session. Did you guess the rule? Comment it. | Cold-open flash of EP.08 map | NEXT: AIRBNB |
On a phone, swipe the table sideways. Every factual line carries a source chip.
How it works.
Mechanism: a contextual bandit chooses one artwork from a per-title pool, using the member's viewing history as context. It is not a one-time A/B test; it keeps betting and learning.
The credit problem: to know which image earned a play, Netflix logs member × title × image → play and keeps the chosen image stable for a member.
The guardrail: the reward is the quality of engagement after the click, and every candidate artwork has to represent the title. That is what stops the optimiser from drifting into clickbait.
What a great PM would do: define the reward one step past the click, and write the honesty rule into the candidate pool, not into a review meeting afterwards.
Steelman: Personalization only works if every poster is honest to the show. The constraint protects the viewer.
Personalize the door, never the room.
Sources
[S1] Chandrashekar, Amat, Basilico & Jebara, “Artwork Personalization at Netflix”, Netflix TechBlog, Dec 2017 · https://netflixtechblog.com/artwork-personalization-c589f074ad76
[S2] N. Nelson, “The Power of a Picture”, Netflix newsroom, 3 May 2016 · https://about.netflix.com/en/news/the-power-of-a-picture