YouTube
content recommendation engine
About 70% of watch time on YouTube is driven by its deep learning recommendations, according to Neal Mohan (CES 2018), with the platform placing recommendation above subscriptions and search.
Objective
Maximize useful watch time by automatically chaining relevant videos, to retain the audience on the platform rather than depending on search or subscriptions.
The deployment
The system generates for each user a feed of suggestions (home and right-hand column) from their history, watch time, likes, shares, and survey responses. It relies on two deep learning networks, one to select candidates, the other to rank them. It processes more than 80 billion signals and compares viewing habits to those of similar profiles.
Results Proof C
The 70% figure was announced publicly in 2018 by Neal Mohan (then Chief Product Officer of YouTube) and reported by the specialized press. Qualitatively confirmed by the official YouTube blog (2021), which places recommendation above subscriptions and search.
How it works
Documented architectureThe stack in detail
- outil Systeme de recommandation two-stage YouTube Two in-house deep learning networks: candidate generation (selection among billions of videos) then ranking (ordering by useful watch time).
- infra Google Brain Google's deep learning team and technologies that support the recommendation networks.
- infra Logs de signaux d'engagement YouTube More than 80 billion signals (history, watch time, likes, shares, surveys) logged per user, the raw material of the two networks.
How it runs, concretely
For ops teams-
1Signal collection site_app / data team
History, watch time, likes, shares, and survey responses are logged per user.
-
2Candidate generation AI
A first deep learning network selects a subset of relevant videos among billions.
-
3Ranking AI
A second network orders the candidates by the probability of useful watch time for this user.
-
4Rendering and loop site_app / human
Home and suggestions are composed; the viewing that follows feeds the signals back.
Watch time and engagement signals (clicks, duration, likes, shares, survey responses). If the system optimizes clicks rather than useful watch time, it drifts toward clickbait.
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
- per-user viewing logs
- indexed content catalog
- explicit and implicit engagement signals
Org prerequisites
- dedicated ML team
- real-time serving infrastructure at scale
- governance over the optimized objective
Possible stack
- custom/in-house
- deep learning frameworks
- two-tower candidate generation + ranking system
The plan, step by step
- Step 1Define the optimized objective (useful watch time, not raw clicks) and instrument the logs.Deliverable: Documented target metric + reliable per-user logging pipeline.
- Step 2Build the candidate generation model.Deliverable: Subset of relevant candidates evaluated offline against the existing system.
- Step 3Build the ranking model and evaluate it offline.Deliverable: Ranking that beats the baseline on historical logs.
- Step 4Launch the A/B test in production on a fraction of traffic.Deliverable: Useful watch time and retention readout vs control group.
- Step 5Industrialize real-time serving and periodic retraining.Deliverable: Production system with drift and quality monitoring.
First step: Define the optimized objective (useful time, not raw clicks) before training any model, because it drives the entire behavior.
The same pattern, proven in other industries
Boots
machine learning scoring of offer-usage propensity, applied to mail and the app
Michaels
genAI personalization of marketing messaging (copy) at scale, tuned to the brand voice
Nike
personalized exclusive offers and access based on member engagement and history
Sources
- S1 YouTube Says 70% Of All Watch Time Is Driven By Its Own Recommendations (Tubefilter, citant Neal Mohan au CES 2018) Secondary archive pending
- S2 On YouTube's recommendation system (YouTube Blog, Cristos Goodrow, VP Engineering) Interested party 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.