How the YouTube algorithm REALLY works in 2023
Written by: Dexxter ClarkYouTube works like any other social media network with a computer algorithm to curate content.
YouTube is big, there are literally millions of videos on YouTube.
The YouTube algorithm is there to get the right video in front of the right viewer.
The YouTube algorithm is literally with the goal: get more watchtime.
Because more watch time means that viewers stay longer on the platform, so youtube earns more money with ads.
But how does youtube decide which videos to show?
YouTube looks at viewer behavior
• in which videos do you click?
• Which videos do you skip?
• How long do you watch a video on a topic?
• On what topics did you search previously?
• Which channels did you watch previously
• Do you comment, like, subscribe on which types of videos?
Based in this information it shows you the videos on the YouTube homescreen, the suggestions on the right side and the YouTube search results.
For every suggestion YouTube does it calculates the likelyhood of a viewer watching that video: expected watch time.
Meaning that YouTube Suggestions are highly personalised.
Not is single person on this planet has the same YouTube Home screen, or exactly the same search results.
This is why you should always use an privacy tab where you are logged out on YouTube, when you do keyword research, so you always get the most objective results.
Hallo, thanks for tuning in, my name is Dexxter Clark from the Netherlands, and I have actually a degree in software engineering.
So I look at bit different then most at the YouTube algorithm.
Misconception: good video
YouTube is all about viewer satisfaction and engagement.
The youtube algorithm It does not judge if the quality of the video is good or if the information of the video is good, it doesn’t know that.
It only knows how viewers respond to a video.
This is why YouTube has such a hard time to keep false information like conspiracy theories out of the YouTube recommendations, because people with a certain mindset enjoy those videos.
YouTube manually has to program exceptions to the watch time rule.
When I researched this video, I came across so much misinformation about the YouTube algorithm out there, but viewers still enjoy the misinformation video, so YouTube keeps pushing them.
If don’t get traction as a small channel you have to abide by 4 basic rules:
• Make viewers click – so make sure the title and thumbnail are top notch
• Make viewers watch longer – so make your video interesting from beginning to end. See the graph on the screen for reference.
• Make viewers react – ask for for engagement or say something controversial
• Make viewers return – for example by making videos around the same topics
Multiple algorithms “when they say YouTube algorithm they mean ...”
To be clear, there is not 1 algorithm, there are hundreds if not thousands, depending on how you count.
From TLS algorithm to encrypt browser traffic, to the insertion-sort algorithm for sorting comments on popularity.
But what most people mean when they are talking about “the algorithm” are the articial intelligence recommender algorithms, which are described in this paper, which are used for YouTube Home, Search and Suggested.
I will make more videos about techniques to trigger these individual algorithms, so make sure you are subscribed for that.
How these algorithms exactly work is a mistery, besides that YouTube is not very open about the workings of the algorithms, even YouTube engineers themselves don’t know how they work, because that is one of the inherit downsides of A.I.
You give it an input and if the outcome is not desireable you tweak the parameters and give it the input again.
You need to keep doing that until it gives a large amount reliable results.
But, as creators we can observe how some of these algorithms behave and learn how we can promote our videos.
Some we can just see on YouTube.
But Google actually has some demos of these algorithms on their website.
Let’s start with the speech recognition algorithm.
This one is actually pretty easy to see in action on YouTube itself.
Go to one of your videos and click on the 3 dots and click on “open transcript”.
This is why it is so important that you select the right language for a video so the algorithm can recognize what you say in the video.
The next one is the Natural language algorithm, which is a text recognition algorithm.
You can see a demo of this algorithm on Googles website [ https://cloud.google.com/natural-language ]
When I put in the script of this video.
You can see how it picks up on entities, sentiment, syntax and categories.
It is called natural language for a reason, because it picks up on keyword stuffing for example.
This is basically the algorithm that Google uses to index websites.
But also the algorithm that YouTube uses for:
• your video title
• your video description
• and the transcript of your video
Based on this information YouTube generates
• internal search labels for video.
• A topic of a video
Like I explained in my ranking video, YouTube starts to test these search labels in youtube search and topical relationships on Home and Suggested.
And based on the results of the viewer behavior it promotes the video further or not.
This algorithm explains why YouTube tags don’t work anymore, because tags can be abused by the creator and these internal search labels not.
You can literally see these search labels in the entities tab.
This algorithm also explains also why the “categories”-section is not important anymore.
You can literally see the category on the categories tab.
One interesting thing I also see is that is seems to put more weight on the beginning of the text.
One pretty famous algorithm that YouTube uses is Google translate.
Which is able to translate text into other languages.
YouTube uses the speech recognition algorithm to know what is being said and then Google translate to translate it automatically to other languages.
However there is a huge bug in YouTube for years, that it recognizes the wrong language on a video, while the video language is set correctly.
In my case it is often Dutch, but I have had it in Portuguese as well.
I’ve seen some of Cathrin Mannings videos being recognized as Japanese.
When that happens, YouTube has absolutely no clue what is said in the video and the only information it has is the title and description.
It is very wise to use your video description to describe the contents of your video.
Cloud vision AI – image recognition
Then the cloud vision AI which lets YouTube understand images.
This is used read your thumbnail for example.
It looks at objects and texts and assigns search labels for everything it ‘sees’ in the image.
It also recognizes faces, expressions and emotions.
But it also detects spoof and violence.
I’ve heard that Google is expanding this to read street signs and famous objects like the Eiffel Tower, so it knows even the locations of images.
A part of this YouTube algorithm is used in the next algorithm: video intelligence algorithm.
Google video intelligence algorithm
Regretfully Google pulled the demo of the video intelligence algorithm, so I’ll show some archive footage here.
It does the same as with images, but every image is a frame in the video.
Also here it assigns search labels to everything it sees: dinosaur, vehicle, tree
And it even recognizes explicit material.
I can imagine how that monday meeting must have gone: a manager: “john you do porn detection this week”.
John had the feed the algorithm porn samples to train the algorithm.
And then when collegues went by johns office looking at his screen “but this is for the algorithm” , “sure, [aha]”.
Clustering viewer interests
With machine learning, the algorithm looks at patterns at a large collection of data.
The YouTube algorithm constantly tries to predict what viewers want to watch.
So the youtube algorithm that looks at patterns of viewer behavior.
If a viewer watches this video, what video do they most click on next?
Is that your video? or do they constantly click on a particular video of another creator?
And YouTube algorithm tries to find patterns in viewers interests.
If a lot of male viewers around 40 with 2 children that watch videos about 90’s retro video gaming also watch a lot of 90’s movies.
So let’s suggest other 40 year old males with 2 children that watch retro video game videos videos about video games of 90’s movies.
A.I.s are normally trained with 1000s-10 thousands of samples to give reliable results.
However for a small channel, you have so little views and the algorithm can not recognize patterns.
It is actually a testament to YouTube’s fantastic engineering skills that you actually get any views on a channel with less than 1000 videos.
Share this articleIf you took value from this article, please share it on social media.
Music Producer / YouTuber
Read more about the author
Music Producer / YouTuber
Read more about the author
Read also ...
Continue reading >>
Continue reading >>
Continue reading >>