Why Netflix's recommendations feel worse every year

The complaint is now so common it has become background noise. You open Netflix, scroll four rows, recognise nothing you want, and put on something you have already seen twice.

The instinctive explanation, that the algorithm got dumber, is almost certainly wrong. Recommender systems have improved enormously over the last decade. The problem is not capability. It is where Netflix points that capability.

A recommender optimises whatever you ask it to

Every recommendation system has an objective function, a number it is trying to make bigger. Choose the number and you have chosen the behaviour.

If the number is “did this person enjoy what they watched”, you get one system. If the number is “minutes watched this month”, you get a different one. If it is “minutes watched of titles whose licensing cost we are currently paying down”, you get a third, and that third system does things that look strange from the outside and make perfect sense from the inside.

None of this requires anybody to be cynical. This is simply what happens when the business that owns the catalogue also owns the recommender.

The three constraints that bend the results

1. A service can only recommend what it has. This sounds trivial and it is the single biggest factor. A neutral recommender considers every film. Netflix considers Netflix. If the best possible recommendation for you tonight is a 1974 thriller streaming on a competitor, the homepage cannot say so. Not out of malice. It simply does not hold that answer.

As the catalogue fragmented across a dozen services, the share of all films any one of them can draw on shrank. The same algorithm, working from a smaller and more idiosyncratic pool, produces worse answers.

2. Originals have to earn their cost back. Commissioning a series costs a great deal up front. The homepage is the most valuable promotional space the company owns, and using it costs nothing. So they use it. That is why a new original appears in the top row whether or not it resembles anything you have ever finished.

3. Retention lives in sessions, not satisfaction. The metric that matters commercially is whether you keep subscribing. Something you half-watched for forty minutes and abandoned looks fine against that metric. It looks terrible against “did you have a good evening”.

Why the interface makes it worse

A design layer sits on top of the economics. Designers build streaming interfaces as infinite horizontal shelves, which is an excellent format for browsing and a poor one for deciding.

Rows invite comparison, and comparison invites deferral. Each additional row raises your reference point for what an acceptable choice looks like, so the eighth row leaves you less likely to commit than the second did. Choice research documents this effect thoroughly, and streaming interfaces come close to a purpose-built machine for producing it.

A service has little incentive to fix that, because time spent browsing is still time spent in the app.

What a neutral recommender can do differently

Not magic. Just a different objective and a different set of constraints.

No catalogue to defend. A tool that licenses nothing has no reason to prefer one film over another beyond whether you will like it.

Explicit signals instead of inferred ones. Watch history is a weak signal because it conflates “loved” with “finished” with “was on while I did the washing up”. A deliberate reaction tells you far more: this yes, this no, this not tonight. A skip informs the model as much as a like does.

Short lists, not shelves. Three films picked for you is a decision. Two hundred films arranged in rows is a browsing session, and browsing sessions are where evenings go to die.

Nothing to sell. If nobody can pay for placement, there is no placement to buy.

The honest limits

A neutral recommender carries real disadvantages and they deserve naming. It does not know what you actually finished, because it cannot see your viewing. It cannot promise a film still streams where it says. It starts knowing nothing about you, where Netflix already holds years of your behaviour. And a taste model is only a model. It will confidently offer you something you dislike, and the only fix is telling it so.

But it aims at the right thing. That turns out to matter more than the sophistication of the machinery, which is the part the “algorithms got worse” theory misses entirely.

The short version

Netflix’s recommendations are not badly built. They are well built for a problem that is not yours: filling the most valuable promotional space in the company with titles it needs watched, drawn from the catalogue it happens to own.

Once you see it as an optimisation problem with the wrong objective, the whole experience stops being mysterious and starts being predictable. Including the part where you give up and rewatch The Thing again.