Cracking the Youtube Algorithm

Melchor Tatlonghari

The Double Blind Test

In the field of science, a double blind test is used to validate a newly created medicine. In this method, the doctors conducting the test are unaware of which pills are placebos and which patients receive them. Similarly, the patients themselves do not know if they are taking placebos or the actual experimental drug. This ensures an unbiased evaluation of the medicine’s effectiveness.

To fully understand how the YouTube algorithm works, I conducted my own double blind test. By removing certain parameters that could introduce bias or doubt, such as gender or video excitement level, I aimed to uncover the underlying patterns and draw accurate conclusions. Over a period of more than a month, I ran multiple channels with different parameters, tweaking them every 2–3 days.

Channel Parameters

I had a total of three channels:

Gaming Channel: This channel consisted of archived videos from my wife’s gaming sessions. She is an avid gamer who plays a variety of games ranging from horror indie games to MOBA games. Travel Channel: This channel was a combination of both our experiences as we traveled around while I built applications and she maintained her streaming channel. It showcased challenges faced while moving from one place to another and living off our backpacks. Technical Channel: This channel focused on software engineering trends, being a digital nomad, relevant books, and anything technology-related. It also included sporadic lofi videos. Timing Parameters Given that I had over 1 terabyte of videos and new videos coming in regularly from our travels and technical recordings on-the-go, I had room to experiment with different release timings. Here are the methods I tried

Same Time Release: Videos were released every day at 12 midnight. Sporadic Time Release: Randomly scheduled video release dates throughout the day (e.g., 2 videos at 10 AM, 3 videos at 4 PM, 1 video at 9 PM). Bulk Release: Releasing 12 videos all at once. Single Release: Releasing one video at a time. Random Time Release: Releasing videos randomly whenever they became available, regardless of time. Video Parameters I experimented with various parameters related to the videos themselves:

Length of Video: Given the variety in the channels, the length of each video naturally had randomness to it. Title of Video: I tested three types of titles: — AI Generated Titles: I used ChatGPT to input a video description and target audience information, and asked Open AI to generate titles based on that. — Manual Trending Titles: I conducted research to identify what the target audience was searching for and created titles accordingly. — Manual Titles: I came up with titles myself based on my own ideas. Thumbnail of Video: I experimented with both allowing YouTube to select the thumbnail it deemed most impactful and creating thumbnails using Canva, following patterns from well-known YouTubers like Mr. Beast. Engagement of Videos: I tested both scenarios of not replying to any comments on the channels and replying and liking all comments on each video. Tags Names: I also experimented with forcing consistent tags based on trending tags in the technology realm, as well as randomly coming up with specific tags for each video. It is important to note that all videos were marked as not made for kids due to legal requirements, so this parameter could not be manipulated. Shorts: I started releasing shorts as well, while experimenting with the same parameters mentioned above.

Experiment Findings

Based on my experimentation across all these parameters for a month, here is how the YouTube algorithm works.

Metaphorically, the YouTube platform can be compared to a massive highway. Technically speaking, it functions as an event stream or an event bus, with videos constantly flowing like traffic. Each user uploading a video becomes an entryway to the highway.

To put things into perspective, every minute, there are two days (48 hours) worth of video being uploaded. This visualization helps grasp the vastness of the highway and the amount of data flowing through it.

The position of a video on the highway is influenced by factors such as geography, tags, title, category, and other input parameters specified when uploading the video. For example, a video uploaded in Thailand about traveling with specific tags like ‘#nomad’ will always appear at a certain point on the highway. Imagine this highway having bleachers on the side from which viewers can only see a portion of the traffic. It is similar to a racetrack in Formula One, but with videos flowing in one direction instead of going around in circles. As a viewer, you have only one chance to see a newly released video. Your view from the bleachers depends on your preferences. For instance, if you watch gaming and productivity videos, those are more likely to be recommended to you based on what you prefer.

Now here’s the key insight: Videos gain momentum and travel further in the highway based on how much engagement they receive from viewers during their limited spotlight time (usually 1–3 minutes only). The more likes, comments, shares, or views a video gets during this brief period, the longer it will stay on the highway and accumulate more views from other viewers in the bleachers. If a video fails to garner enough interaction within that time frame, it will fall off the highway and never appear again in users’ feeds. At that point, sharing it on other platforms like Twitter and Facebook becomes necessary for further exposure. People who view it through direct links or share it themselves can contribute to its view count externally; however, it won’t appear again on YouTube’s recommended videos homepage or the so called highway.

An example release of one video on the highway but not getting further traction after the initial release Some may argue, “But some of my videos continue to receive views after the initial release.” Well, there’s more to it than just the initial release. The level of interaction a viewer has with a channel plays a role. Subscribing to a channel carries more weight than simply liking or commenting on a specific video. The YouTube algorithm curates your view of the highway based on the aggregated interactions you’ve had with different channels. In other words, if I have heavily engaged with Channel A by liking, subscribing, and commenting on their videos, all their previous videos will take precedence in my recommended videos over those from Channel B, where I only watched one video. This is why content creators constantly urge viewers to subscribe, like, and leave comments — to increase the chances of their channel appearing in users’ recommended videos.

When another user with similar preferences joins YouTube, the algorithm assumes they will have similar preferences as well until they make different interactions with categories of videos. The compounding comes from when you release a new video after getting a significant number of initial viewers and they become curious enough to look at your channel, they then start to consume most of all the previous videos you created and thus adding more views to older videos.

Key Takeaways

For channels that are new and haven’t established an identity or target audience yet, titles, tags, and thumbnails play crucial roles in attracting viewers. Aim for strong hooks that resonate with your intended audience. Timing does not significantly impact viewership. Regardless of when you release a video, you only have a small window of opportunity to gain as much viewership as possible. Prioritize retention and engagement instead. This is why advice from larger YouTubers often revolves around consistently creating content. Creating a variety channel that covers multiple categories may initially be more challenging than focusing on a specialized niche. However, it can lead to long-term benefits. It may seem obvious but striving for both quantity and quality content is essential. The impact of this combination becomes more evident once you reach a tipping point, which in my theory is around 500 videos.