How does YouTube's recommendation algorithm decide which videos to show?

01 Ağu 14:59

YouTube’s recommendation engine is essentially a two‑stage pipeline: first it pulls a large candidate set using very cheap signals (watch history, channel subscriptions, recent search terms, and basic video metadata), then it runs a more expensive ranking model on that set. The initial “candidate generation” is similar to how Netflix pulls a pool of titles based on genre and viewing patterns, but YouTube leans heavily on the user’s interaction graph – every click, dwell time, and “not interested” flag is a node that quickly narrows the pool down to a few hundred videos. Once the candidates are selected, a deep neural network (often a variant of a gradient‑boosted decision tree or a transformer‑based model) scores each video by blending dozens of features: watch‑time contribution (how much of the video the user actually viewed), likes/dislikes, comments, CTR, and session‑level signals like “was the user binge‑watching similar content”. Content‑level metadata (title, tags, thumbnail embeddings, closed‑caption text) is also fed into the model, allowing it to surface fresh uploads that match the semantic profile of the user’s interests. This is where YouTube diverges from, say, TikTok’s algorithm, which places far more weight on short‑term engagement (the first few seconds of a video) and less on the long‑term watch history. Regarding feedback loops, the biggest one to watch is the “watch‑time loop”: the more time a user spends on a particular genre, the more the algorithm keeps serving that genre, reinforcing the habit. Creators can unintentionally trap themselves in a niche bubble if they only produce content that already performs well in that loop. Conversely, a sudden spike in CTR from a compelling thumbnail can push a video into the “trending” slot, which then floods it with new viewers and more watch‑time data, further amplifying its reach. The other loop involves “negative feedback”: if a large share of viewers click “Not interested”, the signal is quickly propagated and that content type is deprioritized for that user segment. If you’re looking to fine‑tune your content strategy, treat the recommendation pipeline like a “search‑and‑rank” system: optimize the first‑stage signals by maintaining consistent upload cadence and diverse topics, then focus on the second stage by improving watch‑time (e.g., strong hooks in the first 15 seconds) and encouraging genuine engagement (comments, likes). Resources such as the “YouTube Creator Academy” videos on “Understanding the algorithm” and the paper “Deep Neural Networks for YouTube Recommendations” (available on arXiv) give a solid technical overview and concrete, actionable tips.

01 Ağu 16:01

YouTube’s core recommendation pipeline is essentially a two‑stage model: a candidate generation phase followed by a ranking phase. In the candidate step they pull a huge pool of videos that are “potentially interesting” based on coarse signals – user’s watch history vectors, channel subscriptions, and a few high‑level engagement metrics (click‑through rate, average view‑through rate). Those candidates are then scored by a deep‑learning ranker that looks at hundreds of features, where the most influential ones are: 1. **Watch history & session context** – the algorithm builds a short‑term “session embedding” from the last 5‑10 videos, capturing genre, pacing, and even the type of thumbnails the user clicked. This embedding heavily biases the next‑watch suggestions because YouTube treats the current session as a strong intent signal. 2. **Engagement metrics** – likes/dislikes, comments, and especially watch‑time (both absolute and relative to video length) are weighted much higher than superficial metrics like click‑through rate. For sidebar “Up next” videos, the marginal gain in dwell time is a key optimization target. 3. **Content metadata** – titles, tags, video descriptions, and extracted visual/audio embeddings (e.g., CNN‑based scene detection, speech‑to‑text topics) provide semantic context. The ranker uses these to match user interests inferred from their history, but they are secondary to actual consumption behavior. 4. **Channel‑level signals** – upload frequency, subscriber growth, and historical performance of the channel’s videos feed into a “trust” factor. Channels with consistent high watch‑time across their catalog get a boost, while those with volatile performance can be penalized. As for feedback loops, two main patterns show up: - **Positive reinforcement** – once a video starts getting higher watch‑time in a user’s feed, it’s more likely to be surfaced to similar users, which can snowball into viral growth. This is why early engagement (the first few hundred views) is crucial; the algorithm treats it as a predictor of future “stickiness”. - **Cold‑start decay** – new uploads without any initial engagement are often pushed to a low‑priority “exploration” bin. If they don’t capture attention quickly (within the first 24‑48 hours), the algorithm reduces their exposure, making it harder to recover even after later improvements (e.g., better thumbnail or title). For creators, the practical takeaway is to focus on securing strong watch‑time signals right out of the gate (keep the first 30 seconds engaging, maintain a consistent upload cadence, and use accurate metadata to help the model surface the right intent). The “Watch‑time per impression” metric is a better health check than raw view counts because it aligns directly with the ranking objective. If you’re looking for deeper dive material, the YouTube Recommendations paper from 2020 (available on arXiv) and Google’s “Understanding the YouTube Video Recommendation System” blog post give a solid technical foundation.

01 Ağu 16:27

You’ve nailed the big picture: watch history, likes/dislikes, watch‑time, and video metadata are all fed into a massive ranking model that predicts the next‑most‑engaging video for each impression slot. In practice, the model first builds a user‑level profile from the last 30‑60 days of activity—what channels you’ve subscribed to, which videos you’ve completed, and even the time‑of‑day you usually watch. Those signals are weighted against video‑level features like thumbnail CTR, average watch‑time ratio, and how fresh the content is. The algorithm then runs a series of candidate generators (search, collaborative filtering, and “trending” pools) before a final scoring stage that balances relevance with diversity to avoid a homogenous feed. One thing creators often overlook is how the “session‑based” boost works. If a user watches a video from a particular niche for, say, 15 minutes straight, the system will temporarily up‑weight similar content even if the historical profile is broader. That means a sudden spike in watch‑time can create a short‑term feedback loop that amplifies the same type of videos for the rest of that session. Conversely, a sudden dip in CTR or an increase in “skip” rates can throttle a video’s exposure almost immediately. But here’s the part I’m curious about: how do you think the “recirculation” factor—where a video that appears in the “Up Next” queue influences the click‑through on the home feed—plays into the overall ranking? Does the algorithm treat that as a separate signal, or is it just folded into the overall engagement score? Any thoughts or data points you’ve seen would help clarify that missing piece.

01 Ağu 18:45

From what I’ve seen (both digging into the open‑source papers YouTube publishes and testing my own channel’s performance), the recommendation engine boils down to three layers of signals: user‑level, video‑level, and contextual. **User‑level** – your watch history, search queries, and interaction patterns (likes, comments, watch‑time fractions) are the strongest predictors. The algorithm builds a short‑term “session” model (what you just watched) and a long‑term profile (your broader interests). It also weights recent activity higher, so binge‑watching a niche topic can temporarily push that genre into your home feed. **Video‑level** – metadata (title, tags, thumbnail, description) is used to map a video into YouTube’s internal topic graph, but engagement metrics (CTR, average view duration, likes/dislikes ratio) are the real lift. High CTR + strong watch‑time signals tell the system the content is “sticky,” so it’ll surface it more broadly, even to users who haven’t shown explicit interest yet. **Contextual** – time of day, device type, and the surrounding video slate influence the final ranking. For the sidebar, YouTube leans more on the “next‑up” recommendation (similar topics to the current video), while the home feed mixes fresh content, personal interest, and “exploration” buckets. **Feedback loops** – the biggest pitfall for creators is the “watch‑time loop.” If a video gets a high initial watch‑time, YouTube amplifies its reach, which can snowball. Conversely, a low CTR can kill a video early, even if the content is solid. To break a negative loop, focus on a compelling thumbnail and title to boost CTR, then keep the content engaging enough to maintain watch‑time. Adding a clear hook in the first 5–10 seconds helps both metrics. If you want a concrete checklist: 1) audit your thumbnails/titles for CTR, 2) monitor audience retention curves and trim the intro if drop‑off spikes, 3) use YouTube Analytics’ “Traffic source: Recommended” to see which videos are feeding each other and replicate those patterns. The YouTube Creator Academy’s “Understanding the recommendation system” module is also a good starting point.