YouTube is the one platform where it's still possible to grow in 2021, if you play your cards right. And we're here to break it down. In this article, you'll learn how to create really good content that grows your user base. We've succeeded multiple times on YouTube ourselves. So, this isn't commentary — we've lived this.
This is a slightly technical article, but we've added an "insights" bit in every section with key take-aways. Even if you skip the algorithm parts, make sure you read the insights.
Digging deep with research
We dove into multiple Google research papers on the YT algorithm. We also spoke to creators in India to figure out how to beat the algorithm. It’s a mixture of math, science, and insights that will help us figure out what works best on YouTube in 2022.
Through all this research, we're pretty sure of how the core YouTube algorithm works. This is because Google published two papers in 2016 and 2019 on the YouTube algorithm. And both papers had the exact same core engine in action. However, there have been tweaks over the years to reduce bias and improve time on site. We’ll cover the tweaks and what they mean for creators too in this article.
The north star metric
All companies have a north star metric. Some would call this the one specific metric that the entire company is trying to improve. It’s quite obvious that for most companies this would be revenue, and this is true in YouTube's case too. But the proxy metric that is most tied in to revenue for YouTube is watch time.
The more you watch videos on the platform, the better it is for them as a company. YouTube is constantly optimizing and tweaking its algorithm. This is to ensure that you stay on the platform for as long as you can.
The primary method to increase watch time is recommendation. As boredom starts setting in with any video, the right set of new videos should be suggested on the sidebar.
We start with the 2016 paper titled Deep neural Networks for YouTube recommendations. It admits that recommending videos have three different problems:
Algorithms that work at a small scale don’t work at large scale.
The challenge is with balancing recommendations of new uploaded videos with old established videos that work well.
A user’s watch history can sometimes send wrong signals. You can’t tell when a user has really enjoyed a video, because a large majority of users who really like a video don’t press the Like button..
The Core Algorithm
The core algorithm of YouTube has a 2-layer process. The video corpus in the image is all the videos on YouTube that have ever been uploaded.
Layer 1: Candidate Generation
The candidate generation network takes events from the user’s YouTube activity history as input. It then retrieves a small subset (hundreds) of videos from a large corpus.
This means that the first layer maps out what the user has already watched and subscribed to. It will generally try to show users videos from their own network.
Insight for creators:
- For the creator this means that the main way to get recommended to a user is to get them to subscribe.
- You should probably delete the videos that have poor watch time. This is to improve the chances of your videos being shown to a particular user. Remember, if they watch one video (with high watch time) and don’t sub, then the candidate generation system might show them another video of yours. If this second video has low watch time then you will disappear from their recommendations
The candidate generation network also provides broad personalization via collaborative filtering. Collaborative filtering (CF) is a technique used by recommender systems. This is best explained by an example:
Suppose Jim is a user that likes watching Spider-Man videos on YouTube. Let’s assume that Jim also watches Messi videos because he likes soccer.
Now let’s assume Rick is a user that has created a fresh account and has watched three Spider-Man videos. Now the CF recommender system has created an association between Spider-Man and Messi through Jim’s watch history.
Based on this, it attempts to send Rick a Messi video to see if he likes it. Basically collaborative filtering recommends a video to user Rick based on the interests of a similar user Jim.
YouTube ranks the results of it’s CF algorithm and shows multiple such videos based on their relevance rank. The order of videos shown on the sidebar reflects the "rank" of each video's relevance to the user.
This also works via search queries. YouTube can create an association between two people who searched for particular keywords in the search bar and the types of videos they watch.
Insight for creators:
- Collaborative filtering can be hijacked through collaborations. Imagine that Varun Mayya collaborates with YouTuber Sid Warrier. Watchers of Sid Warrier's channel will see Varun and many of them will go watch at least one video from his channel. This tricks the YouTube algorithm into creating a CF recommendation between our channels.
- This means that collaborating with “How-To” type channels might be more valuable than personalities. This is because users who discovered a creator via a certain search query will get recommendations for your content.
- Dissing another creator also engages the CF algorithm. Here, existing watchers of A will almost always watch a diss/roast of A on another channel. This creates a really good situation for CF algorithms to begin their work creating an association.
- Likes, Down votes, In-Product surveys, and even views don’t matter at all. You can get downvoted to hell and still become highly recommended.
Layer 2: Ranking
This system takes the hundreds of videos returned at candidate generation and serves just a few to the user. It orders the videos in candidate generation by highest likelihood of expected watch time. In very short, ranking is determined by click through rate (CTR) and existing watch time of the video. Watch time is still more important here as using the CTR alone makes people clickbait too often, which YouTube doesn’t like.
It’s kind of obvious that the lower down a recommended video is shown, the lower the likelihood of it being clicked.
The 2019 algorithm accounts for this bias. If you clicked on a recommendation that was very low down the list, then that video will be recommended higher from that point on in other user’s recommended lists. The user’s past history with the channel that uploaded the video being scored is also useful.
The last thing that I didn’t mention was that YouTube is constantly always looking at the freshness of content. The algorithm tries to recommend recent content rather than old stuff.
Insights for creators:
- CTR is quite important to earn a click. Which of these thumbnails below do you think has higher CTR? The one on top has 2x the CTR of the bottom one. The learning? Get a good thumbnail designer.
- Clickbait with a great thumbnail and title, but also keep retention (watch time) high.
If you look at the engagement tab of any video’s analytics, you’ll notice where drop-offs happen. Maximum drop-off happens in the first 20 seconds.
- There’s a difference between Watch Time and % of Video Watched. If you are putting up a 4 minute video and have 90% retention, that’s probably less important than a 10 minute video with 60% retention. YouTube cares more about the total amount of time watched than the percentage of the video watched. This is the main reason live streams aren’t the greatest in the beginning, if your goal is to be recommended often by YouTube. When live, it works well, but after the video is done, you have to mercilessly cut out the boring parts. However, once you become a large creator, there is no better way to do retention than to go live.
- Creators are afraid of taking breaks because they think that consistency drives the algorithm and they must create frequently in order to be recommended more to their own audience. This is not true — YouTubers can take breaks. It’s just that their channels might not grow as fast during the break. Susan Wojcicki, the CEO of YouTube, has gone on record very recently to say that creators aren’t penalized for taking breaks. Watch the video here.
- Recently, one change YouTube made with their algorithm is to show more content from authoritative sources. If YouTube “trusts” you (this doesn’t depend on your channel size, it depends on your Google Knowledge Graph), then it is more likely to show your content.
I personally believe the YouTube ecosystem is becoming the next Hollywood. Not because of the fame of creators, but because getting in is becoming very difficult. Five years ago, producing good content with a clickbait thumbnail was good enough to get views and subs. To get in today, you have to collaborate with a bigger YouTuber and get the CF algorithm and their audience on your side. Just like Hollywood, if you know someone, you have a higher chance of getting in.
Many creators are looking beyond social media. They have started creating communities to cater to a filtered subset of followers. Creators have designed it to be personal & more engaging, and in turn, more monetizable.
Traditional social media platforms are built from a content demand & supply aggregation perspective. They don't focus on creator-led community-building.
They lack specialized tools for interactive community functions. This includes two-way messaging, P2P discussions, courses, gated events, and so on.
Followers on social media generally pay creators via their attention. For example, ad revenue on Youtube or promotional posts on Instagram. Moreover, since the platform owns the demand side, creators find it difficult to negotiate earnings. Creators hence are moving their true fans elsewhere. This will allow them to monetize their followership better.
We hope this article has shed some light on how the YouTube algorithm operates.
If you're looking to start your personal-brand community, we're more than happy to solve the biggest problem while starting; what platform to host it on? Fill in this form and we'll be in touch.