One of the more surprising phenomena to emerge from Nebula’s growth is discovery. When we started, the assumption was that a subscriber would come in for the creator or creators they loved. Our community was small, so you could easily browse the categories, find the creators you already knew, follow them, and move on with life. To our bemusement, people almost immediately started clicking around and just… watching stuff. Maybe the curated nature of Nebula leads to greater confidence in exploration.
We didn’t design or expect Nebula to be a discovery platform, so while we enjoyed reading emails, tweets, and reddit posts from folks who had discovered new creators they love and binged all of their videos, we weren’t really sure what to do with it. Our Featured page, modeled after Apple’s App Store more than anything else, was comprised of a hero rail, there to promote Nebula Originals, and a random assortment of loosely-defined category rails. I’ll be honest: we were just trying to fill up the page to prove we had stuff you could watch.
As time went by, Featured became more sophisticated. We added rails designed to show off exclusive content types, and our editorial team started creating topical rails to coincide with world events or big movie or game releases. These editorial rails are intended to take advantage of the evergreen nature of most of our catalog. Aside from me occasionally complaining about puns in the rail names, the team has full autonomy over what gets featured.
To keep things fresh, we also added a top rail called “Latest Videos”. This is just a waterfall feed of every video as it gets posted. We average 10-15 new videos per day, so this does a great job of giving Nebula a bit of life. It also helps the audience spot things they might enjoy from creators they may not be aware of. This works great until we start getting complaints that TLDR News fills up the rail every morning and people start asking for a mute button.
Nobody actually wants a mute button. Not really. What people want is to use that rail for discovery, and they want it to only be populated with things they’d like. Filtering the waterfall will only work for so long — eventually we’ll have enough creators releasing enough new videos that no amount of muting will bring it under control. Mute? No, what the audience wants is a recommendation algorithm.
Fine. We’ll do it. You’re getting a recommendation algorithm.
I know, I know. Believe me, I know. There are a million ways to get this wrong, and millions of people have millions of opinions about what works and what doesn’t. However, despite what some might guess, we’re not of the opinion that recommendation systems are inherently bad. Our careers are made possible by recommendation algorithms. We discover new creators through recommendation algorithms. Some of my best friends are recommendation algorithms.
The reason creators are often frustrated by recommendation systems isn’t that they’re evil, it’s that they’re opaque, and that opacity can easily feel unfair when your hard work suddenly doesn’t perform as expected and you have no way of immediately identifying why. There’s an entire cottage industry of low-effort content gurus who will gleefully charge you to give you recycled advice on how to improve your thumbnails.
What makes these systems good or bad is how they’re used. In our case, Nebula has a distinct advantage over public, user-generated content platforms: curation. The average Nebula video is of objectively higher quality than the average video on any social media platform. Nebula videos are nutritious. We have no junk food to push on you, and no advertisers for which the content needs to be ad-friendly. I’ll humbly submit that if we were to randomly choose ten videos for you from our catalog, you’d probably enjoy at least one of them.
This presents an interestingly low barrier to entry, and a solution to the real problem that many of us have with these recommendation systems: lack of transparency. What if — and I’m just spitballing here — what if we, like, just told everyone how it works? We could ship a minimum viable product version of a recommendation algorithm in a week, and collect feedback from creators and subscribers over time to make it better and more fair, and we could openly discuss how we do it so that others can chime in and offer suggestions for how to keep it balanced.
Yeah. That sounds cool. Let’s do that. Instead of spending ages building an opaque, potentially broken system for people to complain about, let’s scratch an immediate itch and make it a conversation all of the stakeholders can participate in. We’ll treat this feature as a public beta, make it visible to everyone, and invite discussion. We’ll hold regular roundtable discussions with our creators to refine the system over time, based on their needs and the feedback from the audience. Most importantly, we’ll document it. We’ll keep an up-to-date breakdown of the system logic at go.nebula.tv/algorithm for anyone to read.
At its best, a recommendation algorithm is an adorable robot puppy. It excitedly greets each new visitor with things the visitor might like. Over time, as the puppy is trained by behavior, it gets better and better at guessing what each visitor would want to watch. You liked this video, and other people who liked it also liked this other video. Would you like to see it? The puppy just wants to make you happy.
As a creator-built platform largely designed to sidestep the frustrations we felt with recommendation algorithms, we realize the only way we can feel comfortable building our own algorithm is to do so in a way that every creator can understand and contribute to. The system needs to have clearly defined goals and priorities, and safeguards to ensure that smaller creators don’t get pushed down.
We’re building this according to three simple, straightforward rules:
- The algorithm should be transparent. The creators and the audience should have a reasonable understanding of how the system works. This is a conversation, and we need all participants on the same page.
- The audience should have tools to control what they see. Over time we’ll add settings, like buttons, and other feedback systems. Recommendations shouldn’t only be based on behavior.
- Creators should take priority in recommendations. User behavior is good and interesting, but no matter how hard we try, the creators will understand those relationships better than we can. We’ll give the creators themselves tools to influence what gets recommended after or around their videos.
For the beta launch, we’re keeping it simple by adding a couple of rails at the fringes of what we think will eventually be truly useful: “Discover Something New” and “Channels You Might Have Missed”. In terms of “discovery” this is obviously a little reductive, but it’ll help us start to understand a little more about user behavior and — importantly — the impact on site performance. Going deeper will require a little more internal tracking of user behavior, which we want to approach very, very cautiously. Over the coming weeks and months we’ll apply what we learn to more experimental rails, based on the heuristics laid out in our algorithm definition, and informed by feedback from subscribers and creators.
It won’t be perfect to start. That’s the point. Let’s train this puppy together.
