YouTube’s VP of Engineering explains the process to recommend videos

YouTube comes off with its most detailed explanation regarding its recommendation engine. Also, it puts some light on how it chooses videos to display on your home page.

To demystify the recommendations on YouTube, the company’s VP of Engineering published a blog post and video Q&A. It has addressed the most common questions from both the creators and videos.

The most important thing that one should keep in mind is, the recommendation system of YouTube follows the design to help each viewer find out the satisfactory videos. YouTube says that it uses personalization so that you can find out the importance of many factors for individual logged-in users.

YouTube creators often ask what the engine values more between the overall watch time and viewer retention. Also, they ask if 50% of a 5-minute video is better or rather worse than the 3 min or 10min video.

In such a case, 3 minutes is the overall longer watch time. However, it is only 30% for a 10-minute video which is much weaker in retention compared to its 50%. The data from YouTube indicates that people tend to find it satisfactory when they watch a larger percentage of a video. However, it does not necessarily mean that retention gets a higher value than watch time.

It is quite impossible to say which factor matters the most as it varies as per the user and videos. Certain types of videos have low retention rates. However, viewers can still be satisfied if they get what they want to see.

Recommendations come personalized for each of the users. It means that the watch time does not get treated equally across the board. However, for a viewer, 5 minutes of watch time can be a lot if they watch a video for 2-3 minutes.

There are no such dominant factors in the retention and watch time. How long does it take to get recommended? It sometimes happens that a video fails to gain visibility on the homepage months after its publication. YouTube can’t always help a video if it does not have enough data to understand whom to recommend.