The Adele Gauntlet: Decoding the Youtube Autoplay Function

odstrie

So you’re with a friend. They’re having a bad day and need cheering up. You pull up your trusty bookmarks and open “Cat In A Shark Costume Chases A Duck While Riding A Roomba” for them. The video is a success! You and your friend start swapping crazy animal and/or Roomba stories, and you get so into the conversation that you forget that Youtube is still open. All of a sudden, something else starts blaring through your speakers. Your entire world crumbles around you. As your friend runs for cover, they shout an unintelligible stream of obscenities at you and are never heard from again.

What? You’ve never had this type of experience before? That must be why nobody’s written about this algorithm before. While it is at most a minor inconvenience among Google’s many attempts to wring video views (and accompanying ad revenue) from the general public, it can very quickly affect what gets views driven to it (even views that only initiate the stream and play a tiny segment of the video count toward the total). And those views affect even bigger things, like the Billboard Hot 100. While unlikely, a particularly strong bent toward or away from certain songs may actually make the difference between a single reaching #1 or not (or perhaps charting at all or not).

So how to actually gain more information about a mechanism like this? The Youtube blurb on the autoplay is incredibly unhelpful:

The autoplay feature on YouTube makes it easier to decide what to watch next. After you watch a YouTube video on your computer, we’ll automatically play another related video based on your viewing history.

So obviously, there’s only one other way to find out. I pulled up a completely random song on Youtube and decided to let things flow from there. Originally, my goal was just to continue watching these videos, without any skips, until I reached something weirdly different, but 522 videos later, I can honestly say that I have learned more about myself, about Drake, about Teen Beach 2: The Movie, and about life than I could have ever imagined.

My liveblog of the 522 videos I watched at the mercy of the autoplay function can be found here.

While my findings are not easy to break down into categories, I’ve tried to sort them into four overlapping trends here.

1. The autoplay function tends to take you down familiar pathways. Over 522 songs, you are bound to get a fair number of repeats if you’re letting an algorithm pick a closely related video each time. However, the regularity with which certain songs would show up in a certain order was uncanny. Take the following four sequences of songs:

76    Hello    Adele
77    Someone Like You    Adele
78    Rolling In The Deep    Adele
79    Set Fire To The Rain (Live at the Royal Albert)    Adele
80    Turning Tables (Live at the Royal Albert)    Adele
81    Make You Feel My Love (Live on Letterman)    Adele
82    Don’t You Remember (Live on Largo)    Adele
83    I Can’t Make You Love Me (Live)    Adele
84    Chasing Pavements    Adele
85    Hometown Glory    Adele
86    Rumor Has It    Adele
87    One and Only    Adele
88    Skyfall    Adele
89    Diamonds    Rihanna
90    Chandelier    Sia
91    Elastic Heart    Sia
92    Love Me Like You Do    Ellie Goulding

126    Hello    Adele
127    Someone Like You    Adele
128    Rolling In The Deep    Adele
129    Set Fire To The Rain (Live at the Royal Albert)    Adele
130    Turning Tables (Live at the Royal Albert)    Adele
131    Make You Feel My Love (Live on Letterman)    Adele
132    Don’t You Remember (Live on Largo)    Adele
133    I Can’t Make You Love Me (Live)    Adele
134    Chasing Pavements    Adele
135    Hometown Glory    Adele
136    Chasing Pavements    Adele
137    Don’t You Remember (Live on Largo)    Adele
138    Make You Feel My Love (Live on Letterman)    Adele
139    I Can’t Make You Love Me (Live)    Adele
140    One and Only    Adele
141    Hello    Adele
142    Someone Like You    Adele
143    Rolling In The Deep    Adele
144    Set Fire To The Rain (Live at the Royal Albert)    Adele
145    Turning Tables (Live at the Royal Albert)    Adele
146    Skyfall    Adele
147    Diamonds    Rihanna
148    Chandelier    Sia
149    Elastic Heart    Sia
150    Love Me Like You Do    Ellie Goulding

171    Hello    Adele
172    Someone Like You    Adele
173    Rolling In The Deep    Adele
174    Set Fire To The Rain (Live at the Royal Albert)    Adele
175    Turning Tables (Live at the Royal Albert)    Adele
176    Make You Feel My Love (Live on Letterman)    Adele
177    Don’t You Remember (Live on Largo)    Adele
178    I Can’t Make You Love Me (Live)    Adele
179    Chasing Pavements    Adele
180    Hometown Glory    Adele
181    Rumour Has It    Adele
182    One and Only    Adele
183    Skyfall    Adele
184    Diamonds    Rihanna
185    Chandelier    Sia
186    Elastic Heart    Sia
187    Love Me Like You Do    Ellie Goulding

462    Hello    Adele
463    Someone Like You    Adele
464    Rolling In The Deep    Adele
465    Set Fire To The Rain (Live At The Royal Albert)    Adele
466    Skyfall    Adele
467    Diamonds    Rihanna
468    Chandelier    Sia
469    Elastic Heart    Sia
470    Love Me Like You Do    Ellie Goulding

I have named this the Adele Gauntlet. The sequence is reminiscent of every attempt I made to learn programming: computers stuck in endless feedback loops. While the autoplay algorithm is sophisticated in many other ways, it still is unable to anticipate, from past watch history, a pattern of listening that most people would probably find aggravating or at least boring. (Certainly, when the loop was made up of Disney Channel videos, it was completely unacceptable.)

2. However, the autoplay function and its pathways are also incredibly sensitive to your watch history. There are two ways I can know this:

  • a. My pathway is unique. If you pull up “Hello” while not signed into Youtube, the first autoplay video (both when I started this experiment and at the time of posting) is “Focus” by Ariana Grande. The fact that my account did not go there is likely attributable, at least in part, to some very late nights a few years ago trying to fine-tune my kick-ass lower harmony to Someone Like You, right around the time when I first registered my current Youtube account. (Seriously, it’s a bomb-ass lower harmony. For real.) It may also be influenced by the fact that before I started this experiment, I had never listened to an Ariana Grande song on Youtube. (Problem?)
  • b. My pathway changed, twice. Notice that “Adele I” and “Adele III” above are exactly identical. However, “Adele II” and “Adele IV” are not. What happened? Well, in the middle of Song 134 (Chasing Pavements), I left the computer on which I was running this experiment, and I didn’t come back to it for two days. On a separate computer, I watched a ton of videos in the intervening time that had generally nothing to do with Adele or any of the artists listed above. As a result, when I came back and queued up Song 135 (Hometown Glory), the autoplay recognized that I hadn’t played Chasing Pavements recently and immediately marked it as my next autoplay video.
    • Oddly enough, after this switch flipped for the algorithm, I still got each song in the Gauntlet once (songs 135-150). Of course, since I was starting in a very different place in the order, my order was quite different…until Hello appeared again, at which point the autoplay went straight through the normal Gauntlet, skipping the songs I had just seen, until Skyfall appeared and I moved on to other artists.
    • Recognizing this, I tried to use this knowledge to my advantage to try to escape Disney Channel UK videos, although I can’t say for sure that it was why I did ultimately escape. However, this makes the next conclusion a logical step:

3. Your watch history can dictate not only pathways but also songs to both preference and avoid. Two examples:

  • a. Disney Channel creep. In this 522-video saga was a 219-video chunk from my new least favorite channel, DisneyChannelUK. I legitimately don’t even know why I took the time to link to them and give them any extra clicks. Please just don’t click on that link. While I did ultimately make my way out of Disney Channel territory, the algorithm had been clearly affected by the fact that I had just watched 200+ videos for whom the median viewer age was likely under 14. Here are the five songs, outside the Adele Gauntlet, that I saw twice before the Disney Channel saga and did not see a single time afterward:
    • Loyal—Chris Brown ft. Lil Wayne, Tyga:
      Oh these hoes ain’t loyal (Oh no)
      Whoa these hoes ain’t loyal
    • Post To Be—Omarion ft. Jhene Aiko, Chris Brown:
      I might let your boy chauffeur me
      But he gotta eat the booty like groceries
    • All Eyes On You—Meek Mill ft. Nicki Minaj, Chris Brown:
      (What’s your name?) My name Nick
      (Where you from?) New York in this bitch
      (Choose and pick) You got the right one
      All them hoes, ain’t nothin’ like them
      N**** you know you’d never wife them
    • Only—Nicki Minaj ft. Drake, Lil Wayne, Chris Brown:
      Yo, I never fucked Wayne, I never fucked Drake
      On my life, man, fuck’s sake
      If I did I’d Minaj wid’ him and let ’em eat my ass like a cupcake
      My man full, he just ate, I don’t duck nobody but tape
      […]
      I never fucked Nicki cause she got a man
      But when that’s over then I’m first in line
      And the other day in her Maybach
      I thought God damn, this is the perfect time
      […]
      I never fucked Nic and that’s fucked up
      If I did fuck she’d be fucked up
      Whoever is hittin’ ain’t hittin’ it right
      Cause she act like she need dick in her life
    • 7/11—Beyoncé:
      Hold that cup like alcohol, oh let go like alcohol
      Hold that cup like alcohol, don’t you drop that alcohol
      Never drop that alcohol, never drop that alcohol
      I know you thinkin’ about alcohol, I know I’m thinkin’ bout that alcohol
  • Contrast these five songs with the four songs that did show up again:
    • Lean On—Major Lazer, DJ Snake ft. Mø:
      Blow a kiss, fire a gun
      We need someone to lean on
      Blow a kiss, fire a gun
      All we need is somebody to lean on
    • Uptown Funk—Mark Ronson ft. Bruno Mars:
      ‘Cause uptown funk gon’ give it to you
      ‘Cause uptown funk gon’ give it to you
      ‘Cause uptown funk gon’ give it to you
      Saturday night and we in the spot
      Don’t believe me just watch (come on)
    • See You Again—Wiz Khalifa ft. Charlie Puth:
      How could we not talk about family when family’s all that we got?
      Everything I went through you were standing there by my side
      And now you gonna be with me for the last ride
    • Sugar—Maroon 5: 
      When I’m without ya
      I’m so insecure
      You are the one thing, one thing
      I’m living for
  • All I’m saying is that if my account behaved like a twelve-year-old girl took it over suddenly, skipping Chris Brown tracks would be a thing I completely understand happening. (Although if I were entrusted with the Youtube supervision of a twelve-year-old, I’d probably still want them to watch this video. Which is why it’s probably good that I don’t have a twelve-year-old kid.)
  • b. Taylor Swift/Katy Perry/Meghan Trainor. I exited the Adele Gauntlet three times before ending up in the Disney Channel feedback loop. The first time, I was flung to the realm of electronic dance music before making my way back through some other very white types of music. The second time, a steady diet of David Guetta and Pitbull was a slight change. The third time, however, I was sent to Blank Space by Taylor Swift. This left turn then took me through All About That Bass and an odd mishmash of Katy Perry before descending into Disney madness.
    • Of course, on the back end, Adele IV cut out early in order to meander its way through nonoffensive pop music and once again return to Taylor Swift…and wouldn’t you know? The second ten-song sequence of Katy Perry was identical to the first.
    • Regardless of whether the Disney Channel specifically strengthened the propensity toward Taylor Swift and other nonoffensive pop or merely weakened the propensity toward club music, the end result was that that was (and in several later tests, still is) the direction Youtube pushed me in after I must have cried a thousand times, hate to turn up out of the blue uninvited, plan to swing from the chandelier, etc.

4. Perhaps most significant for real-life consequences, the autoplay tends to be biased toward current mainstream pop hits. This is not a one-off conclusion; I have separately reached the top 40 on new accounts from both the Beatles and the Everly Brothers (although both took longer than the ~50 songs it took from Chumbawamba). There are a number of factors at work here, but what this might mean for calculating streams (and, therefore, the Billboard charts) is important. So how does this work?

First: The autoplay rarely, if ever, evidences a long-term trend to take you significantly backwards in time. Here is the list of songs that were released at least four years before the previous song on the list:

  • Two Princes (Spin Doctors): 1993 vs. 1997
  • U Can’t Touch This (MC Hammer): 1990 vs. 1999
  • Just The Two Of Us (Will Smith): 1998 vs. 2005
  • Hey Ya! (Outkast): 2003 vs. 2007
  • Someone Like You (Adele): 2011 vs. 2015
  • California Gurls (Katy Perry ft. Snoop Dogg): 2010 vs. 2014
  • Hot N Cold (Katy Perry): 2008 vs. 2012

Of these seven, I’d previously played six of them more than once on my current Youtube account (with the exception of Just The Two Of Us), and the only multi-year jump adjacent to any of these songs was for Outkast’s Ms. Jackson (2000 vs. 2003), which I have previously left on repeat at work for forever (forever ever?).

But does the algorithm allow you to step back in smaller chunks? Well, yes and no:

releases
Blue represents video upload date; red represents song release date. Uploads tend to bottom out in 2009 because this was when most mainstream artists switched to a VEVO-sponsored account, something I still have no goddamn understanding of seven years later.

The video I started with was released in 1997. While I was bounced around pretty abruptly to other music easily identifiable as ’90s music, I was pushed forward in time primarily by two artists whose newer material correlates with higher viewcounts, even if it is less popular or less highly regarded. Consider the following sequences:

22    Summertime    DJ Jazzy Jeff/Fresh Prince    10,141,894 views    May 20, 1991
23    Gettin’ Jiggy Wit It    Will Smith    14,486,792 views    Jan 27, 1998
24    Miami    Will Smith    18,158,590 views    Nov 23, 1998
25    Switch    Will Smith    23,098,276 views    Feb 15, 2005

30    My Name Is    Eminem    69,902,544 views    Jan 25, 1999
31    Stan    Eminem    77,691,137 views    Dec 9, 2000
32    Cleanin Out My Closet    Eminem    106,668,174 views    Sep 17, 2002
33    The Way I Am    Eminem    67,558,477 views    Sep 7, 2000
34    Like Toy Soldiers    Eminem    203,551,128 views    Jan 24, 2005
35    Beautiful    Eminem    252,368,687 views    Aug 11, 2009

No, seriously, what is “Switch.” I listened to that song once for this project and again putting this hyperlink in, and I still don’t remember it. And Stan not having even close to the largest number of views in the second sequence is completely unforgivable (well, kind of, since the version I viewed on Youtube is censored to hell and hardly gives you the full impact of the song). This brings me to my second point:

Second (and perhaps the cause of the first): The autoplay tends to preference songs with higher viewcounts. Outside of the Disney Channel frolics, the average video view numbers gravitated toward the ceiling of currently contemplated video views (conflicting figures of speech 100% intended):

views.png
The line represents a rolling five-video average of views, smoothed to minimize visual clutter. Is fashion YOUR kryptonite?

It makes some sense as to why this is the case. A video with more views is more popular; therefore, more people will probably want to get there from their current video. Either the algorithm takes this into account directly, by looking for high-viewcount videos in the immediate related area, or indirectly, by determining what else other people have watched after watching the current video.

Even in taking the Disney Channel universe as its own beast, the algorithm definitely kept steering me to the most viewed videos. DisneyChannelUK has over 1500 videos (and adding more every day, as I heard over 200 times), and yet even though I viewed fewer than 100 distinct videos, this is what their “most popular” page looks like for me:

mostpopular.png

That’s nineteen of the top 25 videos on the channel that I’ve watched. And I swear none of them are the videos I watched after I actually got invested in one of the terrible shows I kept getting clips from. (I mean, uh, that didn’t happen.) And looking at the most viewed Youtube videos at the time I went through this ordeal, seven of the eight most viewed videos released after January 1, 2013 (Blank Space, See You Again, Uptown Funk, Shake It Off, Dark Horse, Roar, and All About That Bass) and seven of the ten most viewed overall came up, often more than once.

What were the exceptions, though? Two were Gangnam Style and Bailando, which I will address in the next point. The third was Baby, a video now nearly six years old, which makes it practically ancient by the standards of the release date chart above. Furthermore, it has more or less stopped gaining views at any appreciable rate at this point, having averaged fewer than 500,000 views per day for the last two years (by comparison, Hello has continued to average over ten million views per day):

viewsperdayfinal
I definitely only circled the Adele Gauntlet each time it happened to point it out and not because it makes the scatter plot look more legitimately like there is a strong upward trend.

If you’re still not convinced, I also ran a very quick test of Maroon 5 songs (and it wasn’t just because I was sad and wanted to listen to Songs About Jane). I took the first four singles from their first album, released in 2003, and tried to see what I would pull up as the first related video. You would probably expect that other singles from that album would show up first, right? Here are the results:

  • Harder to Breathe: Makes Me Wonder (2007)
  • She Will Be Loved: Misery (2010)
  • Sunday Morning: Sugar (2014)
  • This Love: Won’t Go Home Without You (2007)

Third: The algorithm recognizes formalized “streams” and “universes” of videos that are considered closely related due to demographic data, and it seems difficult to move between these streams unless you either a) exhaust everything within your current stream, or b) deliberately calibrate your watch history to bring two streams closer together.

What do I mean by that? For example, this is why I can start at a video of a mashup and never get anywhere other than other mashups. As much as you would expect that a Snoop Dogg—AC/DC mashup would be related to videos from each artist individually, most people who are looking for this kind of mashup are already broadly familiar with popular music and are craving absurdity, commentary, or what have you. Alternatively, consider the cases of both Gangnam Style and Bailando. Both are viewcount monsters, and I should have been pushed toward them eventually if I were going to leave the autoplay running for this long. However, K-pop is not a particularly common stream (or, heh heh, a MAIN stream) in American music, and neither is Enrique Iglesias’ style of Latin music either. (Chart performance suggests that Bailando’s success may be largely due to views from outside the United States.)

Strangely, pathways can create and enforce divides within a single artist or channel. In order for me to start at any of the ten Katy Perry videos in my list and get to the rest of her singles, I generally have to watch a couple videos near the bottom of the list (“exhausting” that stream), and then cut back to a video near the top of the list (deliberately inducing a break in my pathway). Even still, a good number of Perry’s earlier singles are completely missing from any attempt I make to bring them up on autoplay (although maybe that’s because of the problematic ones).

At the same time, pathways may lend themselves to movement between streams. My DisneyChannelUK escape was accomplished partially, as far as I can tell, because of a set of rap/hip-hop videos from a movie that likely did not demographically fit in with the rest of the DisneyChannelUK viewership. I do believe that the last 15 years of mainstream music has primarily developed by blurring the lines between culturally Black and culturally white popular music (which is a completely separate blog post…or ten). However, there are still readily apparent divides, and the places where that divide is bridged (or not, leading to a shift in genre or style) is useful to explore demographic shifts and trends regarding popular music (some of this rambling occurs in the live blog).

So what does this mean? If high viewcounts are actually relevant to the number of streams that get directed to a video through autoplay, what Google has done, in effect, with the autoplay function, is create a mechanism that drags you toward Top 40 radio. As a result, a significant autoplay push toward songs that already have high viewcounts will skew the Billboard numbers toward the tail end of a song’s true popularity. Although the effect is probably small (Youtube makes up one portion of the ~25% weight streaming services are given, and most streams are probably not autoplay-generated), if Youtube streams continue to rise and the autoplay function is embraced as normal by future generations, it could lead to all sorts of havoc on the Billboard charts. Songs that would frequently arise in autoplay could have much longer tails (and thus take longer to exit the Hot 100 once they have peaked). On the other hand, a song like Gangnam Style is going to be working at a disadvantage because even though it is wildly popular on Youtube, it is not being frequently autoplayed on the magnitude that a song like Hello might, for example, ultimately underselling its true popularity.

But what do I even mean by true popularity? Was that something the Billboard charts were ever measuring? Should it be something we actively encourage the charts to measure? Or maybe it don’t matter, it clearly doesn’t tear you apart.

adele-hello-video-2

For a liveblog of the 523 videos I watched to create this article, click here.

Advertisements

Leave a Reply

Fill in your details below or click an icon to log in:

WordPress.com Logo

You are commenting using your WordPress.com account. Log Out / Change )

Twitter picture

You are commenting using your Twitter account. Log Out / Change )

Facebook photo

You are commenting using your Facebook account. Log Out / Change )

Google+ photo

You are commenting using your Google+ account. Log Out / Change )

Connecting to %s