Netflix
personalization of the artwork via contextual bandits
Netflix chooses each title's artwork member by member with contextual bandits, a layer distinct from title ranking that serves over 100 million accounts at a peak of 20 million requests per second, with a significant A/B gain that is stronger for unfamiliar titles.
Key points
- Personalization of each title's artwork member by member.
- Proprietary contextual bandits serving over 100 million accounts, peak of 20 million requests per second.
- Significant A/B gain vs non-personalized, stronger for unfamiliar titles.
- Evidence B, confirmed status.
Objective
Convert attention into viewing by showing each member the image most likely to trigger the click at the moment of consideration, and help them discover unknown titles.
The deployment
This component chooses which image to show for a given title, member by member. It is distinct from the recommendation engine that decides which title to offer: it sits on top of that ranking and personalizes the visual representation of the already-selected title. Netflix describes a contextual bandit problem: among one to several dozen candidate images per title, the model predicts the one that maximizes the probability of a play for that specific member, then keeps the highest-scoring image. The system serves over 100 million accounts, with a peak of 20 million requests per second for rendering the personalized images. A/B tests showed a significant gain over a non-personalized approach, the effect being stronger for members who do not yet know the title. Documented in 2017, the setup is still in production: established press confirms in late 2025 that Netflix's promotion technology remains driven by machine learning, and the research team published work in 2025 on artwork personalization via LLM post-training.
Results Proof B
Primary Netflix engineering blog documenting the system, its scale, and an A/B gain, plus established press and a 2025 research publication confirming continuity in production. The gain is described as significant with no public percentage, hence B.
How it works
Documented architectureThe stack in detail
- outil Bandits contextuels custom Netflix Proprietary ML that chooses, among a title's candidate images, the one with the highest play probability for each member.
- outil Modeles de prediction d'engagement Prediction of the play probability per member and per image, learned from first-party engagement signals.
- infra Infrastructure de serving d'images Netflix Rendering of personalized images at scale: over 100 million accounts, peak of 20 million requests per second.
- llm Post-training de LLM (evolution 2025) Work published by Netflix Research in 2025 evolving artwork personalization via LLM post-training; exact model not named.
How it runs, concretely
For ops teams-
1Title recommendation AI (recommendation)
The recommendation engine decides which titles to show the member (an upstream, distinct step).
-
2Image selection AI (artwork personalization)
Among the title's candidate images, the contextual bandit predicts the one with the highest play probability for that member.
-
3Personalized rendering AI / infrastructure
The selected image is rendered in the interface, at a scale of tens of millions of requests per second.
-
4Continuous learning ML team
Observed clicks and plays refine the predictions through A/B tests and exploration.
The predicted play probability per member and per title, learned from first-party engagement signals. Without a bank of candidate images per title or an engagement signal, the system has nothing to arbitrate.
How your customers perceive this type of use
Sourced studiesLe paradoxe est documente des deux cotes : 71% des consommateurs attendent des interactions personnalisees et 76% sont frustres quand elles manquent (McKinsey, 2021), mais 75% declarent ne pas acheter aupres d'organisations auxquelles ils ne confient pas leurs donnees (Cisco, 2024). La « creepy line » est localisee : messages recus quelques secondes apres une recherche et suivi de localisation sont les pratiques qui mettent le plus mal a l'aise (Periscope by McKinsey, 2019).
Acceptance conditions
- La confiance dans le traitement des donnees precede l'achat : 75% ne achetent pas sans elle (Cisco 2024)
- Un cadre legal protecteur rassure : 59% des consommateurs disent que des lois fortes sur la vie privee les rendent plus a l'aise pour partager des informations dans des applications IA (Cisco 2024)
- La personnalisation elle-meme est attendue quand elle est consentie : environ la moitie des consommateurs (US 55%, UK 52%) disent s'inscrire souvent ou parfois a des services personnalises (Periscope by McKinsey 2019)
Red lines
- Le message declenche quelques secondes apres une recherche ou un achat : deuxieme ou troisieme cause de malaise selon les pays (Periscope by McKinsey 2019)
- Le suivi de localisation percu comme de la surveillance : 40% de malaise en Allemagne et au Royaume-Uni (Periscope by McKinsey 2019)
- Le mesusage des donnees personnelles par l'IA, devenu la premiere inquietude des consommateurs, a 53% et en hausse (Qualtrics 2025)
Sources: McKinsey & Company 2021 · Periscope by McKinsey 2019 · Cisco 2024 · Qualtrics 2025
How to replicate
Inference, not sourcedData prerequisites
- several candidate images per content item
- first-party engagement signals (click, play)
- stable member identifier
Org prerequisites
- an ML team for contextual bandits
- an asset production pipeline per title
- high-load rendering infrastructure
Possible stack
- contextual bandits framework
- engagement prediction model
- A/B testing and exploration pipeline
The plan, step by step
- Step 1Produce several candidate visuals per content item on a high-traffic sub-catalog.Deliverable: Multi-variant image bank ready to serve.
- Step 2Instrument clicks and plays per displayed visual, then build a first engagement prediction model.Deliverable: Clean dataset and a model evaluated offline.
- Step 3Launch a simple bandit (epsilon-greedy or Thompson sampling) on a high-traffic surface, on a fraction of traffic.Deliverable: Bandit in production on a limited scope.
- Step 4Measure incremental clicks and plays in A/B against the single image.Deliverable: Readout of the incremental gain, or a stop decision.
- Step 5Industrialize: continuous exploration, serving scale-up, drift monitoring.Deliverable: System in production with a dashboard and alerting.
First step: Produce several visuals per content item and launch a simple bandit test on a high-traffic surface, measuring the incremental click.
The same pattern, proven in other industries
Sources
- S1 Artwork Personalization at Netflix Primary archive pending
- S2 Netflix bets on AI to power ads, search, and storytelling Established press archive pending
- S3 Netflix Artwork Personalization via LLM Post-training Primary archive pending
An error, newer info, a source?
This page lives on its accuracy. If a figure has moved, if the deployment has changed, or if you have a higher-quality source, tell us. Every sourced correction is verified before publication.