Technology & Computing · Digital life & society
How recommendation feeds decide what you see
Follow how a feed narrows billions of items to a screenful, what its score rewards, and how the design of a screen steers the choices you make.

What you will learn
- Distinguish selecting candidates from ranking them.
- Explain why a recommendation isn't an independent judgment of importance.
- Explain how optimizing clicks can favor a misleading headline.
- Distinguish a measurable proxy from the user benefit it's meant to represent.
Lesson 1 of 4 · 4 min
How a feed picks a few videos out of billions
YouTube holds billions of videos, and your home screen shows a handful. Something had to choose. Google says 60% of watch time on YouTube comes from recommendations, so most viewing starts with that choice.
A search begins with words you typed. A recommendation system offers things you didn't know to ask for. So the way it narrows billions down to a screenful shapes what you end up watching.
No system can study billions of videos in detail every time you open the app. So a common design works in two stages.
Candidate generation comes first. It cheaply pulls a much smaller pool from several sources, such as videos related to ones you've watched, what's popular or trending, and what your friends liked. Ranking comes second. A more careful model scores each candidate in the pool and sorts them.
Say the pool comes down to a cooking demonstration, a comedy clip and a repair tutorial. The ranking model gives each one a score. A typical score is a prediction, such as the probability that you'll watch this video, worked out from your watch history, your language and country, the time and the video's own details.
The highest score goes first. What you can't see from outside is which prediction the system was built to make. Two services could rank the same three videos differently and both be working as designed.
The sorted list still isn't final. A re-ranking stage applies extra rules. It can remove candidates, such as videos a separate model has flagged as clickbait. It can adjust scores to favor newer videos. And it can add variety, because the closest matches tend to be near-copies of each other. Google's course imagines a viewer shown nothing but owl videos, who soon loses interest.
So the order on your screen doesn't follow one score from top to bottom. People chose the rules, the goal and the signals, even where a model does the arithmetic.
Try to recallWhy doesn't appearing first in a recommendation list prove an item is the most important?
The top spot goes to whatever scores highest on the system's chosen prediction, such as how likely you are to watch, after re-ranking rules for things like freshness and variety. None of that measures importance or truth.
Lessons in this topic
How a feed picks a few videos out of billions
Follow a recommendation from the candidates a system gathers, through scoring and re-ranking, to the short list on your screen.
Why a feed that chases clicks learns clickbait
See what happens when a recommender is told to maximize clicks or watch time, and why neither number equals a satisfied viewer.
Dark patterns and the choices a screen steers you into
Learn to spot designs that hide or obstruct a choice, and see why the position of an item on a screen distorts what a feed learns about you.
Two ways a recommender guesses what you'll like
Compare recommendations based on similar items with ones based on similar people, and see why a brand-new video is hard for a feed to place.
Three quick stories from this topic
The mischievous genie
A feed told only to maximize clicks can end up promoting clickbait
Google's course for engineers compares machine learning to a mischievous genie that grants exactly the wish you state. Reward clicks and it may serve clickbait. Reward watch time and it may favor very long videos. Neither number means the viewer was glad they watched.
The measure is a substitute
Clicks are easy to observe. Satisfaction and learning are harder to measure. A proxy helps you count something, but it never becomes the human benefit it stands for.
Ask what the user wanted
A short answer that lets someone finish a task may beat a long session. Staying engaged and getting something done are different outcomes.
Most YouTube viewing is recommended
Google says recommendations account for 60% of the time people spend watching YouTube
The figure comes from Google's own course for engineers, which adds that 40% of app installs on Google Play come from recommendations too. So more than half of all YouTube viewing starts with a choice a ranking system made.
Billions in, a screenful out
YouTube holds billions of videos. A recommender first gathers a small pool of candidates, then scores and sorts them, then adjusts the list for things like freshness and variety.
Where an item sits matters too
Items higher on a screen get clicked more than items lower down. Engineers call it positional bias, and it means a click count partly measures placement.
Fake countdown timers
The FTC counts fake countdown timers among the design tricks it calls dark patterns
The timer suggests an offer is about to end when it isn't. The agency's 2022 report also lists ads disguised as independent articles, fees revealed only late in a purchase and subscriptions made hard to cancel.
A case the FTC brought
The FTC alleged that the online learning site ABCmouse promised 'Easy Cancellation' while making it extremely difficult to cancel free trials and subscriptions.
Why it works
Extra effort on one path changes what people do. So a count of customers who stayed subscribed reflects the design of the screen as well as what they wanted.
Sources
- Recommendations: what and why? | Machine Learning | Google for Developers, Google for Developers
- Scoring | Machine Learning | Google for Developers, Google for Developers
- Re-ranking | Machine Learning | Google for Developers, Google for Developers
- FTC Report Shows Rise in Sophisticated Dark Patterns Designed to Trick and Trap Consumers | Federal Trade Commission, US Federal Trade Commission





