YouTube’s recommendation engine is one of the most successful innovations Google has ever built. A staggering 70 percent of watch time on YouTube is driven by YouTube’s own recommendations.
Despite this, the SEO industry tends to focus on sayings like “YouTube is the world’s second largest search engine,” and emphasize ranking in YouTube search results or getting YouTube listings in Google search results.
Especially surprising is the fact that YouTube has actually published a paper (The YouTube Video Recommendation Engine) describing how its recommendation engine works.
Yet this paper is rarely referenced by the SEO industry.
This article will tell you what’s in that paper and how it should impact the way you approache SEO for YouTube.
To this day, metadata remains far more important for SEO on YouTube than it is for search results in Google.
While YouTube is now able to create automated closed captions for videos and its capacity to extract information from video has improved dramatically over the years, you should not rely upon these if you want YouTube to recommend your video.
YouTube’s paper on the recommendation algorithm mentions that metadata is an important source of information, although the fact that metadata is often incomplete or even entirely missing is an obstacle that their recommendation engine is designed to overcome as well.
To avoid forcing the recommendation engine to do too much work, make sure that every metadata field is populated with the right information with every video you upload:
Include your target keyword in the video title, but make sure the title also grabs attention and incites curiosity from users.
Attention-grabbing titles are arguably even more important on YouTube than traditional search, since the platform relies more heavily on recommendations than search results.
Include a full description that uses your keyword or some variation on it, and make sure it is at least 250 words long.
The more useful information you include here, the more data YouTube has to work with, allowing you to capitalize on the long tail.
Include the major points you will cover in the video and the primary questions that you will address.
Additionally, using descriptions that relate to other videos, as long as it is appropriate from the user perspective, may help you turn up in the recommendations for those videos.
Keyword tags still matter on YouTube, unlike the meta keyword tag for search engines, which is completely defunct.
Include your primary keyword and any variations, related topics that come up in the video, and other YouTubers you mention within the video.
Include your video in playlists that feature related content, and recommend your playlists at the end of your videos.
If your playlists do well, then your video can become associated with keeping users on YouTube longer, leading to your video showing up in recommendations.
Use an eye-catching thumbnail. Good thumbnails typically include some text to indicate the subject matter and an eye-catching image that creates an immediate emotional reaction.
While YouTube’s automated closed captions are reasonably accurate, they still often feature misinterpretations of your words. Whenever possible, provide a full transcript within your metadata.
Use your keyword in your filename. This likely doesn’t have as much impact as it once did, but it certainly doesn’t hurt anything.
2. Video Data
The data within the video itself is becoming more important every day.
The YouTube recommendation engine paper explicitly references the raw video stream as an important source of information.
Because YouTube is already analyzing the audio and generating automated transcripts, it’s important that you say your keyword within the video itself.
Reference the name and YouTube channel of any videos you are responding to within the video as well in order to increase the chances that you will show up in their video recommendations.
Eventually, it may become more important to rely less on the “talking head” video style. Google has a Cloud Video Intelligence API capable of identifying objects within the video.
Including videos or images within your videos referencing your keywords and related topics will likely help improve your video’s relevancy scores in the future, assuming these technologies aren’t already in motion.
Keep your videos structured well and not too “rambly” so that any algorithms at play will be more likely to analyze the semantic content and context of your video.
3. User Data
Needless to say, we don’t have direct control over user data, but we can’t understand how the recommendation engine works or how to optimize for it without understanding the role of user data.
The YouTube recommendation engine paper divides user data into two categories:
- Explicit: This includes liking videos and subscribing to video channels.
- Implicit: This includes watch time, which the paper acknowledges doesn’t necessarily imply that the user was satisfied with the video.
To optimize user data, it’s important to encourage explicit interactions such as liking and subscribing, but it’s also important to create videos that lead to good implicit user data.
Audience retention, especially relative audience retention, is something you should follow closely.
Videos that have poor relative audience retention should be analyzed to determine why, and videos with especially poor retention should be removed so that they don’t hurt your overall channel.
4. Understanding Co-Visitation
Here is where we start getting into the meat of YouTube’s recommendation engine.
The YouTube paper explains that a fundamental building block of the recommendation engine is its ability to map one video to a set of similar videos.
Importantly, similar videos are here defined as videos that the user is more likely to watch (and presumably enjoy) after seeing the initial video, rather than necessarily having anything to do with the content of the videos being all that similar.
This mapping is accomplished using a technique called co-visitation.
The co-visitation count is simply the number of times any two videos were both watched within a given time period, for example, 24 hours.
To determine how related two videos are, the co-visitation count is then divided by a normalization function, such as the popularity of the candidate video.
In other words, if two videos have a high co-visitation count, but the candidate video is relatively unpopular, the relatedness score for the candidate video is considered high.
In practice, the relatedness score needs to be adjusted by factoring in how the recommendation engine itself biases co-visitation, watch time, video metadata, and so on.
Practically speaking, what this means for us is that if you want your video to pick up traffic from recommendations, you need people who watched another video to also watch your video within a short period of time.
There are a number of ways to accomplish this:
- Creating response videos within a short time after an initial video is created.
- Publishing videos on platforms that also sent traffic to another popular video.
- Targeting keywords related to a specific video (as opposed to a broader subject matter).
- Creating videos that target a specific YouTuber.
- Encouraging your viewers to watch your other videos.
5. Factoring In-User Personalization
YouTube’s recommendation engine doesn’t simply suggest videos with a high relatedness score. The recommendations are personalized for each user, and how this is done is discussed explicitly within the paper.
To begin, a seed set of videos is selected, including videos that the user has watched, weighted by factors such as watch time and whether they thumbed-up the video, etc.
For the simplest recommendation engine, the videos with the highest relatedness score would then simply be selected and included in the recommendations.
However, YouTube discovered that these recommendations were simply too narrow. The recommendations were so similar that the user would likely have found them anyway.
Instead, YouTube expanded the recommendations to include videos which had a high relatedness score for those would-be initial recommendations, and so on within a small number of iterations.
In other words, to show up as a suggested video, you don’t necessarily need to have a high co-visitation count with the video in question. You could make do by having a high co-visitation count with a video that in-turn has a high co-visitation count with the video in question.
For this to ultimately work, however, your video will also need to rank high in the recommendations, as discussed in the next section.
6. Rankings: Video Quality, User Specificity Diversification
YouTube’s recommendation engine doesn’t simply rank videos by which videos have the highest relatedness score. Being within the top N relatedness scores is simply pass/fail. The rankings are determined using other factors.
The YouTube paper describes these factors as video quality, user specificity, and diversification.
Quality signals include:
- User ratings.
- Upload time.
- View count.
The paper doesn’t mention it, but session time has since become the driving factor here, in which videos that lead to the user spending more time on YouTube (not necessarily on that YouTube video or channel) rank better.
These signals boost videos that are a good match based on the user’s history. This is essentially a relatedness score based on the user’s history, rather than just the seed video in question.
Videos that are too similar are removed from the rankings so that users are presented with a more meaningful selection of options.
This is accomplished by limiting the number of recommendations using any particular seed video to select candidates, or by limiting the number of recommendations from a specific channel.
The YouTube recommendation engine is central to how users engage with the platform.
Understand how YouTube works will dramatically improve your chances of doing well on the world’s most popular video site.
Take in what we’ve discussed here, consider giving the paper itself a look, and incorporate this knowledge into your marketing strategy.