Key points
- A recommendation system looks at your past record and picks out, in order, what you'll probably like next.
- There are broadly two approaches: looking at items whose content resembles each other, and looking at patterns of similar taste across many people's records.
- It doesn't just use star ratings and likes. Whether you watched to the end, or skipped past in a few seconds is a far bigger ingredient.
- A new user with no history gets recommended poorly, and a new item is hard to surface through the "people like you" approach. This is called the cold start problem.
- Getting the match right isn't the whole job. Keep showing only more of the same thing, and you get boxed into a narrow fence.
The analogy
Walk into the coffee shop down the street often enough over a few months, and your order gets shorter. The owner remembers what you always drink and has it ready before you ask, and quietly suggests a new bean when it seems like your kind of thing. This — looking at your past record and picking out what you'll probably like next — is a recommendation system.
The suggesting splits into two approaches. One looks at how the drinks themselves resemble each other: a customer who always picks a nutty bean gets offered a new bean with a similar flavor. The other looks at how customers resemble each other: if a few regulars whose taste overlaps with yours loved a new item, it gets suggested to you even though you've never tried it.
Both approaches hit a wall with a first-time customer. There's no record left behind, so there's nothing to go on. So a first-time customer is offered what sells best first, or asked a few questions about taste.
In detail
The ingredient is behavior, not star ratings
Very few people bother to leave a star rating every time. So recommendation these days mostly runs on traces people leave just by going about their business, not traces they meant to leave — what they clicked, how long they stayed, whether they watched to the end or skipped away in three seconds, whether they added something to a cart and never bought it.
The advantage of these traces is that there's an enormous amount of them; the drawback is that what they mean is fuzzy. Staying on something a long time could mean it was interesting or that it was confusing, and there's no way to tell from the trace alone, while an accidental click gets counted as interest too. So each signal gets weighted differently, and several traces get layered together.
Looking at items whose content resembles each other
Turn a book's genre and author, a song's tempo and instruments, or the words used in a product description into a bundle of numbers, and you can measure the distance between one item and another. This approach hands you whatever sits close to something you already liked.
It works immediately even on something brand new — as long as there's a description, it can be placed even before anyone has bought it yet. The tradeoff is that it keeps circling back to more of the same thing. Watch one scary movie and you get fed nothing but scary movies from then on.
Looking at people whose taste resembles yours
There's another route. Find people whose choices overlap heavily with yours, and recommend whatever they liked that you haven't seen yet. It works without knowing anything about what an item actually contains — all it needs is a record.
This route's strength is turning up surprises. Things that share no genre and no description with what you liked, yet somehow fit, come out of this approach. The tradeoff is that it needs a certain amount of history built up first, and it tends to keep pushing whatever's already popular further to the top.
A real service filters in two stages
Where there are millions of items, there's no way to score every single one. So a fast, rough method narrows the field down to a few hundred candidates first. Then a more careful model ranks only those candidates.
At the very end, a set of rules takes one more pass — mixing things up so the same category doesn't run in a row, slipping in something newly added on purpose, and dropping anything you've already bought. The list that shows up on your screen has passed through several layers to get there.
How do you know it matched well
The first check usually sets aside the most recent part of the past records, shows only the rest, and tests whether the system can guess the part held back. After that, real users get split into groups shown different recommendations, and the results get compared.
There's a trap here, though. Aim only at the click rate, and sensational titles tend to rise to the top. So alongside how well it matched, a system also tracks how varied the results are, how much new material it surfaced, and whether people stayed satisfied over the long run.
More precisely
A recommendation system treats the problem as turning both a user and an item into a bundle of numbers, and calculating a score for how well the two fit together. The approach that looks for patterns of shared choices across many people's behavior records is called collaborative filtering. It can find people with similar taste, or it can measure how alike items are from the records, as in "people who bought this also bought that." The approach that uses what an item actually contains is called content-based recommendation. Real services blend the two, then layer on context like time of day or which device someone's using.
The analogy breaks down in one place. The coffee shop owner recognizes a customer and suggests something good, but a recommendation system doesn't understand taste. It only finds patterns in the records and turns them into a score, so even an accidental click gets mixed in as a trace of interest. And the owner's goal is the customer's satisfaction, while a service's recommendation is usually aimed at whatever goal the service itself has set. What gets chosen as that goal reshapes the entire list that ends up on your screen.
Try it yourself
- Embedding Projector (explore an embedding space) ailearn.space Spin around a space where similar things cluster close together, and you get a feel for recommendation as nothing more than finding what's nearby
- Semantris (word-association AI) ailearn.space The machine picks whichever word fits best with the one you typed. It's a playful way to feel what scoring a match actually means
Common misconceptions
It's easy to think a recommendation is reading your mind, but actually it's finding patterns in items close to your past records and in the records of people who behaved similarly to you, and turning that into a score.
It's easy to think a good match means it's been listening to your conversations, but actually search and click history alone can produce a match good enough to feel that way.
It's easy to think the better the match, the better the recommendation, but actually showing only more of what you already like closes off any chance to run into something new.
One-line summary
A recommendation system finds items and people that resemble your past behavior and hands you, in order, what you'll probably like next.
Spotted an error or have a better analogy? Suggest an edit · Last updated 2026-10-17