YouTube’s algorithm serves the most relevant, personalized videos to their users on five different sections of their platform: YouTube search, home page, suggested videos, trending, and subscriptions. By helping users find the videos they’re most likely to watch and enjoy, YouTube can keep viewers on the platform for as long as possible and get them to visit their site regularly.
So how does the YouTube algorithm work?
The short answer is: nobody knows the details, not even YouTube, as the algorithm is constantly changing. Besides, YouTube wouldn’t communicate about it anyway, as that would lead to people exploiting them.
That said, there are many different factors known that can help your content being suggested by YouTube algorithm.
1. Content: Watch Time, Engagement.
First of all : ” Content is king “. And even if you applied all SEO best practices, the main element to build a successful YouTube strategy remains the video content.
Apart from the content itself, one important metric YouTube uses is watch time— how long users stay on the video and channel. Indeed, it indicates that the content is relevant for the users and will keep them longer on the platform. The algorithm takes into account many different factors and ranks them accordingly: viewer retention, impressions to clicks, viewer engagement (views, share, likes, comments) and some other factors that will remain unknown. YouTube then tailors these factors to your profile so that it can suggest videos you’re more likely to click.
Whenever a video is suggested by YouTube, it is called an “impression”. An impression can lead to a views if the users click on it (we talk about “click-through rate”) and the watch time is then linked to the average view duration.
Impressions will lead to views, which will increase your overall watch time as you can see below :
FYI all these metrics are available on YouTube Analytics
The challenge here is there to increase the click-through rate (and indirectly the watch time) by having an appealing thumbnail and drive the user to the video. Warning : Your video thumbnail and title must relate to the content of the video, otherwise it won’t meet user’s expectation and the viewer retention will be low, so as your SEO.
Indeed, the title and metadata are as well one important factor in YouTube algorithm.
Titles, description and tags are a key part of video SEO on YouTube and help get your video served in the correct searches and related videos. Please find below SEO basics for YouTube metadata :
- Keep your title short and catchy:
Artist Name – Release Name (Official Video / Freestyle/ Making of. )
- Your description could contain clear calls to action for subscriptions and further listening/viewing. It should also be keyword heavy to optimise for SEO.
- 1st line: call to action to subscribe to your channel
- Then: Artist Name, Video Name, Buy Links, Social Media, other video/playlists links
Maximum : 5000 characters
- Use 5-10 high quality and relevant tags on your video :
Artist name, video title, album title, music genre, album name, other top tracks, featuring and key-words linked to external promotion
- Be succinct! If you only had few words to tell a friend about your video, what would you say. Title: about 60 characters
- Be descriptive! Tell the viewer what kind of video they will see (official audio, lyrics, official video. ). Put the most important video information up front.
- Be Smart! Use YouTube autosuggestion to see what is the most relevant to your content
After revising all the metadata, the next step to help your content being pushed in suggestions is to use YouTube features and shows to the algorithm that your channel is active.
3. YouTube Features
YouTube algorithm will push in priority the channels that have activity. The most important point being the frequency of uploads. This is why artists need to think about the content they can upload to the platform (Official Music Video, Official Lyrics Video, Video Cover, Making Of, FAQ, Behind the Scene etc.) and think about a timeline where they have regular uploads over the long-term.
If the artist don’t have video content to upload, another way to keep a channel active is using YouTube features such as :
– Comments (on the artist channel or 3rd party ones)
More generally, each time YouTube launch a new feature, they will tend to recommend the channels using it.
4. User History
Finally, the user history is also being part of the algorithm. Every user will have different suggestions based on their activity on YouTube, in order to have a customized experience and stay longer on the platform. That is why you might not seen the same top searches for the same keyword on two different sessions for instance.
YouTube will analyze the videos you watched to suggest you similar videos or channels you might like, due to keywords, similar audience or other users behaviors on the platform. The non-watched videos are also an indicator. Indeed, if you have not clicked on a specific content despite a high number of impressions, it gives the information that you might not be interested in this type of content. Likewise, YouTube will reduce the number of impressions for a video you have just seen entirely for instance. The comments are also being part of the algorithm and a positive or negative comment on a video can influence your future video suggestions.
And of course, as stated in the beginning, some other factors remain unknown but all of the above should enable any content owner to optimize his/her YouTube channel SEO.
Sep 23, 2016 2 min read
In a recent paper published by Google, YouTube engineers analyzed in greater detail the inner workings of YouTube’s recommendation algorithm. The paper was presented on the 10th ACM Conference on Recommender Systems last week in Boston.
YouTube recommendations are driven by Google Brain, which was recently opensourced as TensorFlow. By using TensorFlow one can experiment with different deep neural network architectures using distributed training. The system consists of two neural networks. The first one, candidate generation, takes as input user’s watch history and using collaborative filtering selects videos in the range of hundreds. An important distinction between development and final deployment to production is that during development Google uses offline metrics for the performance of algorithms but the final decision comes from live A/B testing between the best performing algorithms.
Candidate generation uses the implicit feedback of video watches by users to train the model. Explicit feedback such as a thumbs up or a thumbs down of a video are in general rare compared to implicit and this is an even bigger issue with long-tail videos that are not popular. To accelerate training of the model for newly uploaded videos, the age of each training example is fed in as a feature. Another key aspect for discovering and surfacing new content is to use all YouTube videos watched, even on partner sites, for training of the algorithm. This way collaborative filtering can pick up viral videos right away. Finally, by adding more features and depth like searches and age of video other than the actual watches, YouTube was able to improve offline holdout precision results.
The second neural network is used for Ranking the few hundreds of videos in order. This is much simpler as a problem to candidate generation as the number of videos is smaller and more information is available for each video and its relationship with the user. This system uses logistic regression to score each video and then A/B testing is continuously used for further improvement. The metric used here is expected watch time, as expected click can promote clickbait. To train it on watch time rather than clickthrough rate, the system uses a weighted variation of logistic regression with watch time as the weight for positive interactions and a unit weight for negative ones. This works out partly because the fraction of positive impressions is small compared to total.
YouTube’s recommendation system is one of the most sophisticated and heavily used recommendation systems in industry. The paper just scratches the surface but nonetheless gives several useful insights regarding engineering deep learning systems.