The problem they solve
A video service may hold tens of thousands of titles. A store may sell millions of products. No one can browse all of that, so services try to pick a short list likely to interest each person. A recommendation system is the software that makes that guess. It is not reading minds. It is estimating, from patterns in past behavior, how likely you are to click, watch, buy or read something.
What signals go in
Systems differ, but they draw on a handful of kinds of information.
- Actions you take. Watching, skipping, buying, liking, saving, searching, or how long you stay on something. These are often called implicit signals because you did not say “I like this” outright.
- Ratings and reactions. Explicit signals such as stars or thumbs up.
- Details about the items. Genre, price, topic, length, or the words in a description.
- Context. Time of day, device, or what you were just looking at.
- Details about you. Sometimes basic profile information, though many systems work mainly from behavior.
Collecting this information raises digital privacy questions, and some of it can be shared or combined with other data. Privacy policies usually describe, in general terms, what a service collects.
Idea one: people like you
One classic approach is collaborative filtering. The idea is that people who agreed in the past may agree again. Consider a hypothetical example with three viewers.
- Ana watched shows A, B and C.
- Ben watched A, B and D.
- Chen watched A and B.
Chen’s history looks like Ana’s and Ben’s. Both of them also watched something Chen has not: C and D. A system using this approach might suggest C and D to Chen. It never needed to know what the shows are about, only who watched what.
A variant looks at similar items instead: items watched by many of the same people are treated as related.
Idea two: items like what you liked
Content-based methods look at the items themselves. If you read several articles about gardening, the system can recommend more articles whose text or tags are about gardening. This works even for a brand-new item that nobody has watched yet, but it can keep showing you more of the same.
Combining methods and learning from outcomes
Large services typically blend these approaches and add machine learning on top. A common pattern is in two steps. First, a fast method narrows a huge catalog to a few hundred candidates. Then a more detailed model scores those candidates and orders them. The scoring model is trained on past results, using the supervised learning approach of comparing predictions with what people actually did.
Different services also optimize for different things: clicks, minutes watched, purchases, or return visits. The goal chosen shapes what rises to the top, and it is not always the same as what is best for the viewer.
Why suggestions sometimes miss
- The cold start problem. A new user has no history, and a new item has no audience, so there is little to go on.
- Shared accounts. If several people use one profile, the signals are mixed.
- Odd one-off behavior. Buying a gift for someone else can skew suggestions for weeks.
- Behavior is not always preference. You may click on something because it was placed prominently or because the thumbnail was striking, not because you wanted it.
- Limited or skewed data. Popular items have far more data than niche ones, so they tend to get recommended more, which then makes them even more popular.
Effects on what you see
Because the system learns from what people already click, it can create a feedback loop: it shows you more of what you have engaged with, and your next actions reinforce the pattern. Over time this may narrow what a person sees, though how strongly it does so depends on how the service is designed. Some services add variety on purpose, mixing in less obvious choices. The same pattern has been linked to unevenness in outcomes, a topic known as algorithmic bias.
In online shopping, recommendations can also steer attention toward certain products. On social platforms, they influence which posts spread. It is reasonable to remember that the order in which things appear reflects a ranking decision, not a neutral list.
Practical points for users
Most services let you adjust or clear history, rate items, or mark “not interested,” and these actions usually feed back into future suggestions. Using separate profiles helps with shared accounts. And if your feed feels repetitive, deliberately searching for something different can widen what the system learns about you.