Netflix Category Genre Generator

What these categories actually are

You have seen these on a streaming home page and assumed a person wrote them, or that there are a couple of dozen of them. Neither is true. There are tens of thousands, most of which no user has ever knowingly seen, and they are assembled by a machine out of a vocabulary built by hand.

Press Generate and it hands you one. A few from the list:

• Action Sci-Fi & Fantasy for ages 11 to 12
• African-American Showbiz Dramas
• Zombie Horror Movies
• World Music Concerts
• Action Thrillers based on Books
• Adult Animation
• Cerebral Foreign War Movies
• Gritty Crime Dramas featuring a Strong Female Lead

Categories that specific do not come from a marketing department picking a few dozen buckets. They come from a system that took cinema apart, tagged every piece, and can reassemble the pieces into an almost unlimited number of oddly precise sentences. The story of how it was discovered is genuinely strange.

Drawing a category

1. Press Generate for one category.
2. Ask for a few and use them as an actual watchlist prompt, the oddly specific ones are the good ones.
3. Type into Contains to search, by genre, by mood word, by anything.
4. Copy to keep them.

Where these come from

Streaming micro-categories gathered and checked by hand, a large sample of a system that runs to tens of thousands. The tool picks from that pool in your browser.

The man who found the door by accident

In 2014, a journalist named Alexis Madrigal, writing for The Atlantic, got curious about these categories, the weirdly specific ones that surfaced on his home page and could not be found in any menu. He started collecting them in a spreadsheet.

Then somebody noticed the thing that cracked it open. Each category lived at a web address ending in a number, and the numbers ran in sequence. Which meant you did not have to wait for the system to show you a category. You could simply take the address, change the number on the end, and walk straight into whichever one that number pointed to.

So Madrigal did what any curious person with some programming would do. He wrote a small program to fetch them one after another, counting upward, and let it run. It came back having found on the order of seventy-six thousand distinct categories.

Seventy-six thousand. Not two dozen. Not a few hundred. A number so large that no human being could have written them, and no home page could ever display them, sitting there behind sequential numbered doors, most of them never once shown to a user. He had found the back of the machine.

And the categories themselves were the tell. Violent films starring one specific actor. Gory low-budget horror from one specific decade. These are not buckets somebody sat down and named. They are outputs, generated on demand, and that meant there had to be a generator.

The thirty-six-page manual for watching a film

Madrigal traced the system to an executive named Todd Yellin, and what Yellin described is the part that should change how you think about a streaming service.

The company had paid people to watch films. Not casually, watch them. And as they watched, they tagged, filling in an enormous structured questionnaire about every aspect of what they were seeing, guided by a training document that ran to thirty-six pages of instructions on how to do it consistently. Internally the whole scheme had a suitably grand name: Quantum Theory.

Consider the depth of what got recorded. Not just the genre, but the setting, tagged by location. The characters, tagged by occupation. The plot, tagged for how conclusively it ends. And a whole set of qualities rated on a sliding scale, not present or absent but graded from one to five: how romantic, how gory, how sad the ending is, how morally upright the lead is. Every film got a romance score, even the ones nobody would ever call romantic. Every ending got placed somewhere on the line from happy to devastating.

This is the crucial move, and it is easy to miss. They did not sort films into boxes. They measured films along dozens of independent dimensions, and turned each film into a long list of numbers and tags. A film was no longer a thing that belonged to a genre. It was a point in a space with dozens of axes.

Once every film is a list of numbers, you are no longer running a video shop. You are running a database, and a database can be asked questions a shelf never could.

How a sentence like this gets built

Here is where the tens of thousands come from, and it is almost embarrassingly simple once you see it.

The engineers took all those hand-made tags and wrote a grammar: a set of rules for stringing tags together into a readable phrase. An adjective of mood, then a region or a source, then a genre, then an age range, in that order, and only in combinations that make sense. Feed the grammar the tag for gritty, the tag for a strong female lead, and the tag for crime drama, and it prints Gritty Crime Dramas featuring a Strong Female Lead. Feed it different tags and it prints something else.

The number of categories is therefore not a list somebody wrote. It is a multiplication. Every mood word times every region times every genre times every qualifier, minus the combinations the rules forbid, and the total runs into the tens of thousands without anyone ever having to name them one at a time. The categories are not stored. They are generated, on request, from parts.

If that sounds familiar, it should, because it is exactly what this page is doing. A tool like this holds a pool of pieces and assembles one on demand, rather than storing every possible result. The difference is only one of scale. The streaming system has a richer grammar and a bigger vocabulary, but the principle is identical: do not keep the outputs, keep the parts and the rules for combining them.

What the categories are really for

The strange categories were never really meant for you to browse. They are a side effect of the actual purpose, which is quieter and cleverer.

Once every film is a list of numbers and every viewer is a record of what they watched, the system can do the thing it was built to do: match the two. It does not need to know that you like clever heist films with a downbeat ending set in Europe. It only needs to notice that the films you finish tend to score a certain way on a dozen hidden dimensions, and then find other films that score the same way. The oddly specific category on your home page is just a human-readable label stuck onto a cluster the machine already found.

Which means the label is a little window into how the system sees you. When it offers you something as specific as Understated Films featuring a Strong Female Lead, it is not describing a genre that exists in the world. It is describing a corner of its map where it thinks you live, translated into English so you will click.

There is something worth sitting with in that. Cinema, the whole sprawling art of it, has been measured, scored, and turned into coordinates, so that a machine can steer you around it. That is an astonishing feat of organisation and a slightly cold one, and the funny little category names are the one place where the machinery becomes visible, poking through the surface, showing you for a moment exactly how thoroughly everything has been taken apart and counted.

Questions people ask

Are there really tens of thousands of these?

Yes, though the exact number moves as the vocabulary and the rules change. The figure of around seventy-six thousand comes from one journalist's count in 2014, and the point is not the precise total but the order of magnitude: far more than a person could write, because they are generated rather than authored.

Why are most of them hidden?

Because they are raw output, and most of it is useless to any given person. The system generates every combination its rules allow, but only ever surfaces the handful that match what it has learned about you. The rest sit unseen behind their numbered addresses, valid but never shown.

Did people really watch everything by hand?

The foundational tagging was done by trained people following the manual, yes, because a machine in that era could not reliably judge how sad an ending was or how upright a character seemed. Modern systems lean more on automation, but the original vocabulary that made the whole thing possible was human judgement, written down in a very long document.

Do other services work the same way?

The specifics differ, but the underlying idea, break every item into measurable attributes and match on those rather than on broad genre, is now how essentially all recommendation works, for films, music, shopping and more. The visible micro-categories were an early and unusually candid glimpse of a technique that is now everywhere and mostly invisible.

References

1. Madrigal, A. C. (2014). How Netflix Reverse-Engineered Hollywood, The Atlantic. The original investigation and interview with Todd Yellin. https://www.theatlantic.com/technology/archive/2014/01/how-netflix-reverse-engineered-hollywood/282679/
2. FlowingData (2014). How Netflix creates movie micro-genres, on the scalar tags and the generative syntax. https://flowingdata.com/2014/01/03/how-netflix-creates-movie-micro-genres/