How YouTube Decides What You See
YouTube is a vital aspect of both the American and world cultural milieu — people go there for news about the world, for informed social commentary, and to learn about everything else from cooking to music, electronics, automotive repair and home improvement. Parents use it as a babysitter, students use it to study, and nearly everyone visits youtube for entertainment content.
As of 2024, YouTube has nearly 3 billion monthly active users. It is the only streaming service with over 90% market penetration in the United States, which is the country with the second most YouTube users behind India. About 246 million of a total estimated 340 million Americans are active YouTube users.
Few of those people understand how YouTube makes its sausage. In 2019, YouTube actually removed part of its website in an effort to obscure how that sausage is made, after a group of creators discovered that publicly available source code revealed internal information about how YouTube’s content distribution scheme operates. More on that in a minute. First, the basics.
YouTube is basically a giant search engine (in fact, the second largest search engine by traffic) that hosts videos, but users are mostly watching YouTube’s algorithmic recommendations, which are designed to keep you engaged with the site and its content. They have a video hosted by their YouTube Creators channel on the algorithm.
The selection of videos that appear on your YouTube HomePage are based on metrics like performance, defined as how much a video has interested similar users, and also watch and search history, how often you watch a channel or search for a topic, as well as how many times you’ve already already been offered a video to watch.
According to YouTube, “We track what viewers watch, how long they watch, what they skip over, and more. This helps us figure out what kind of videos they like best and what we can recommend to them next.”
YouTube Suggested Videos are where the algorithm really comes into play; these are the videos recommended on the right side of the screen alongside the video users are watching. These are meant “to offer you videos that you are ‘most likely to watch next,’” YouTube says.
The Home page and Suggested Videos algorithm bases this on many different factors, including -
Engagement, like watch time, video views, likes, dislikes, and shares, which are all measured to decide what kind of content you’ll see more of
User actions like adding videos to queue, saving videos to watch later, share, are used to serve you more videos from either that creator or from similar channels. Choosing feedback like not interested will make it so you won’t see that creator or that type of content.
Surveys, where YouTube prompts users with a quick survey that will impact performance, meaning how often the video is offered to others.
Now back to the source code discovery mentioned above. The information that was uncovered included YouTube’s internal ratings system, details about the ways that YouTube throttles videos, and a rating scheme for individual channels called a P-Score.
If you’ve taken a statistics course, you know a P-Score as a measure of probability. YouTube uses this in much the same way; they are a “Preference Score” used as an internal, proprietary metric which the YouTube algorithm relies on to decide which content might resonate with other users.
YouTube’s P-Scores are based on five underlying metrics: popularity, or watch time; passion or engagement, meaning comments, likes, shares, etc; protection, or how advertiser-friendly the content is; platform, a measure of how often the content is watched on which kinds of devices, for instance tablets, phones, PCs, televisions; and production quality.
A channel with a high P-Score will get into Google Preferred, which sells premium advertising on what YouTube considers the top 5% of it’s most advertiser-friendly channels. If a channel has a high enough P-Score, it will be enrolled in Google Preferred automatically, generating a larger share of AdSense revenue by running premier ads.
If a channel has a P-Score lower than YouTube’s top 5%, it’s not getting into Google Preferred. Like YouTube’s recommendation system, the P-Score has been heavily criticized since its discovery, mostly by creators who feel YouTube is unfairly suppressing their content.
This is what led YouTubers Bowblax, Nicholas DeOrio, and Optimus, along with reporter Josh Pescatore, to look at YouTube’s public source code to see what they might find.
Dirty Delete the Details
“Creators who are doing great with popularity, platform, and passion [engagement] might never have the opportunity to fix their protection scores because they’d never know it’s an issue,” the creators write in their findings, which were reported by TubeFilter and Reclaim The Net in 2019.
Murder Pop has archived the original document by the creators who did the research as a PDF here on our site in case Google eventually decides to censor it, as it resides on Google Docs.
One hour after the creators’ report was originally published, offering other YouTubers step-by-step instructions on how to find their own P-Scores, YouTube removed these sections from its public facing code entirely.
YouTube “does not alert creators or give them an opportunity to access this data without finding some sort of weird exploit to get to it,” so the group used the inspect element feature of their browsers to trawl through public code.
They found that every video showed something called a P-Score under a section called ContentLabelRating. Along with the video uploaders P-Score number, they found ratings for individual videos, including Y, G, PG, Teen, Mature, and X. Based on their research, the group thinks X-rated content is demonetized and age-restricted by default, and “deemed problematic altogether.”
A number of videos they looked at also had sections of code labelled to describe the video being throttled, a deliberate action to slow down a video’s spread.
“REASON_BRAND_SAFETY_ADS_THROTTLING,” is one such label; “FEATURE_VIDEO_IS_THROTTLED_FOR_OVERLAY” is another.
They found these on nearly every video rated Teen or above. The group’s research lead them to conclude that YouTube’s throttling falls under several different categories.
Brand safety throttling appears to refer to videos that are determined not to be advertiser-friendly, and feature video throttling seems refer to whether or not YouTube algorithms recommend a video to users.
“If the code here is to be taken literally, the webpage seems to quite literally be throttling advertisements on videos rated above PG mostly due to brand safety issues,” they wrote in their report.
“It will not fully demonetize, but it will strictly only advertise enough to make the video enough money to seem like it’s doing well.”
The report also noted that P-Scores differ between different geographical regions, and that scores tended to be lower outside of an uploaders home country.
The research also revealed that YouTube likes huge media conglomerates… a lot. Of the 229 channels the investigators looked into, the top 5 were all late-night shows, and the top 10 expanded that to news outlets owned by the same media conglomerates, with the lone exception of Linus Tech Tips.
In fact, outlets owned by large media conglomerates dominate the Top 30, which is the only visibility we have, as it seems that Google censored this document from the Google Sheets where it was stored.
If you exclude gamers like Mr Beast, who operates the largest YouTube channel with over 200 million subscribers where he makes questionable content for children with impunity, and other creators making or aggregating content for children, LinusTechTips, PewDiePie, and Joe Rogan seem to be the only independent YouTubers who have cracked the top 30.
The creators who did this research say that despite their research, they remain “completely in the dark like usual with how YouTube has developed all of this and implemented this.”
They added, “We know that this information alone is extremely valuable to the YouTube community, and so figuring this out has been a potential revelation that can benefit creators for a long time to come.”
YouTube gave TubeFilter an official statement about all of this which Murder Pop has re-published below. As of the time of this stories publication, YouTube has refused to give Murder Pop any direct comment at all, or even answer us regarding the censorship of our own content, beyond asking us for our channel before going radio silent.
“In order for videos to be eligible for advertising, they must comply with our Advertiser-Friendly Guidelines, which are public and published in detail. We have developed internal systems to mirror these guidelines so we can evaluate the millions of videos that are published to YouTube and determine which ones comply with these guidelines. To make sure we apply this process fairly, we also give creators the chance to appeal if they feel any of their videos have been unfairly demonetized. A set of monetizing channels are also eligible to be part of Google Preferred Lineups, which allows brands to appear among the most engaged and popular channels on YouTube.”
Interesting overview & indepth understanding of how YouTube algorithms prioritize what a viewer’s offered to watch. What about YouTube channels that a viewer has previously subscribed? Wouldn’t this also affect what they’re offered to view? Thanks for another well researched topic! Hopefully MurderPop will continue to grow in subscribers! Great job!