Game Theory - How To Create A VIRAL YouTube Trend

Episode Date: May 23, 2024

Join former Game Theory Host MatPat as he breaks down the specific's of what makes something go VIRAL!

Big Payday, Big Payday, Big Payday, 10 minute of video. Dude, before we get it, I will see, if you can subscribe to YouTube, on YouTube, turn out, On order to show everyone in my opinion, like every other and leave, four comments, Two, a comment, Let's go!

Enough noise, today I'm exposing the facts about how YouTube's Black Box actually works. Hello internet, welcome to Game Theory, the show that occasionally throws the game part of the title out the window to expose the secrets of YouTube. And today is one of those days. Now a few episodes ago I talked about how the YouTube algorithm Unintentionally helped to explode Minecraft and all the YouTube channels covering the game one diamond mine card at a time And the fact that the YouTube algorithm can make or break not just creators but entire businesses

Cod Cawthin Fidget Spinners made me want to go a lot deeper and get a whole lot more specific for all you creators who watch this show Now at this point most of you know that Steph and I help other you YouTubers optimize their channels and we get literally thousands of requests from you to help over email in-person at conventions Even creepy notes left in our mailbox Which legitimately was a thing that happened one time in the past Please don't ever do that again. That was kind of weird. It was also unsigned which just added to the creepiness factor Overall, it was a very thoughtful note but also just very disturbing to receive in your mail

I know you had the best of intentions, but it was unmarked unaddressed written in code and had literally been left physically in my mailbox So meant that you found out where we lived and that you kind of creepily hung out in front Let's try to avoid that in the future thanks Needless to say a lot of you are looking for help and you're asking for it all over the place And while we'd physically love to it just takes so much time that there's no way that we could possibly help everyone I'd never get actual theories about games done So because of that we're always looking for ways to give the largest number of you the secret tips that we've learned over the years

Since at the end of the day it empowers all of you to create better, smarter videos. To anyone who's ever wanted to learn more about YouTube, how it works, or how their favorite creators got to be where they are, and how you can be the next Mark, Felix, Jack, Jake, Logan, Liza, Lily, or Lou. Just a, just not Matt Pat. Because there can only be one mad pat. First, let's get one thing straight. This is not speculation. So no more myths, no more hearsay, not even... Dare I say it? Not even theories. These are the secrets of the algorithm delivered by the people who made it themselves.

Google. That's right. A few months ago while trolling through the bowels of the Google research website, which is what we do for fun, we stumbled across a research paper published last year that actually explains the real YouTube recommended algorithm, uncovering the code that unlocks all that stuff that shows up on your home screen when you log in. Yeah! This is the real deal delivered in the absolute driest way possible. An engineering research paper. It's like the plain rice cake. of the literary world. This stuff is worse than reading Hemingway, but I'm here to decode it for you in a way that I haven't seen done anywhere else. And I mean that. I've been talking about the YouTube algorithm since 2012. And in 2014, get this, a real YouTube employee told me that there are so many pieces to the YouTube algorithm that, quote, no single person knows how the entire algorithm works. End quote. That is a real line from a real guy who should know the real stuff. Real scary, right?

Seems like the YouTube people should probably know what's going on on their own platform. It'd be like if Henry Ford was like, oh yeah, there's a lot of doohickeys in that car engine, but who are we to say what they really do? But understanding the engineering behind the algorithm shows why this is possible. No one person can predict everything about the algorithm, and it's on purpose. YouTube has become way too big, and YouTube user behavior is so unique to each individual viewer that no human can handle all that data. So now the YouTube algorithm works,

by what's known as machine learning, which, at its core, is basically a kind of artificial intelligence. You see, this whole time we've been talking about it wrong. I think when anyone on YouTube says the YouTube algorithm, people just assume it works like a formula. If X, then Y. If YouTube's algorithm is optimizing for likes, I put in a video that begs for likes and I get more views out. But it's not nearly that simple. In reality, it's taking a bunch of different data points, X's and Y's and Z's, to calculate what gets more views. So we need to stop talking about it like it's a formula and instead start thinking about it like what it truly is Two neural networks. You see a neural network, at least as it matters here, is just a filter with a bunch of different layers

Like how Rhett and Link pee into a jar then drink each other's urine through a straw filled with a bunch of different filtration layers to get the warm slightly salty still kind of urine-flavored water on the other end The urine in this example are all the videos on YouTube There's an analogy for ya. Them sucking into the straw is someone looking for a video to watch, and YouTube's neural networks are the little hollow filtration fibers inside that straw. The fibers get progressively smaller and smaller, filtering out first big particles like dirt and salts, and then finer things like bacteria and parasites, until you get the awkwardly warm, clean water on the other end.

That's you, Danky Memes 47. That warm water retches spit out of his mouth? That is your video! You want them to be retching on your slight, salty, warm urine-soaked video. So where are these two filters and how can you, and the next video that you make, get your way through them? While the first neural network is called YouTube's candidate generator, which is happily not a machine that pumps out Trump and Clinton clones. It's just a fancy word for the filter that takes all the zillions of videos floating around in YouTube's urine jar and narrows them down to just a few hundred. The second one is called the ranking filter, and it's like your own personal YouTube scorecard. It's the one that takes those few hundred videos

that YouTube thinks you might like and turns them into the top one or two YouTube is really sure that you will like and then they'll stick those on the recommended list at the top of your page. So what do you need to do to get through these filters and make it into Link's mouth? Well, let's start with the candidate generator. How to beat out the other billion possible videos to make it into the top 100. Well first you should know some history. According to Google, this filter used to be shallow. It used to only look at one thing and one thing only. What else? had that viewer watched. So if you watched five Orbees videos this week, your recommended feed would be more full of Orbees than Guava Juice's bathtub. And, I'm guessing, some of other places. Just saying those balls are small, getting all up in there. The problem is that people on YouTube end up watching a lot of one-off videos.

I mean, sure, I'll watch a single $1 versus $10,000 fidget spinner video, learn that they're all crappy clickbait because no one has the actual Cajonaz to spend that kind of money on a fidget spinner, and then never click on a nub spinner, another one again. On YouTube, people tend to watch one video out of curiosity, like the Diamond spinner or B-movie meme videos, but then never want to watch another of that type again. But under the old system, all they would see would be more of that sort of video, more videos about expensive fidget spinners. As such, YouTube needed to take more into account than just search history.

And now it does. Just take it straight from Google. Quote, a key advantage of using deep neural networks as a generalization of matrix factorization is that arbitrary, continuous, and categorical features can be easily added to the model. Search history is treated similarly to watch history. Each query is tokenized into unigrams and bygrams and each token is embedded. Once average, the user's tokenized embedded queries represent a summarized dense search history. End quote! Ha ha ha!

Yep, that gobbledygook is what we read for fun around here. In English, what all that garbage means is that it prioritizes two main factors. What you've searched for recently and where you've spent the most watch. But what's any of this mean for you, Danky Memes 47? Well, it's why trends like fidget spinners or B-movie or We Are Number One memes are hitting so much harder than past years. If you watch one video out of curiosity, YouTube will send you a bunch more because you showed interest in the subject recently. And when someone creates a fidget spinner video looking like this, one that everyone clicks on because it looks so cool, it reactivates everyone in YouTube's algorithm for the phrase, fidget spinners, thus prolonging the trend another weaker.

So as such don't title your video good times or let's play part 62 use searchable words Proper names catch phrases in other words stuff that people will be searching for and will be in other video titles Because if what people are searching for match what's in your title and tags you have a much better shot at popping up in someone's recommended place That's the power of recency but there's also the factor of time Create something longer that two-minute Skyrim remix ain't gonna get you recommended come on danky There's not enough watch time there, especially when you're competing with stuff like 30-minute let's-play videos and 21-minute hairstyle compilations or one-hour vine compilations. To get through that filter, the most important thing is watch time. So create things that people are gonna want to watch for a long amount of time.

Some of the other factors in the candidate generator are demographics. Let's go back to that plain rice cake and hear what it has to say. Quote, demographic features are important for providing priors so that the recommendation behave reasonably for new users. The user's geographic region and device are embedded and concatenated. Simple binary and continuous features such as the user's gender, logged-in state, and age are input directly into the network as real values normalized to it. End quote. Don't you dare take that gibberish lightly? There's a reason why toy unboxing videos have 19 million views apiece. They're not fooling anyone about who their audience is.

When channels we consult for say that their audience is every Heart stars and horseshoes were so inclusive. Oh no, you are so wrong! That is a big red flag to us when the YouTube algorithm specifically filters by all your audience demographics. By trying to appeal to everyone, you're a lot more likely to reach no one. And last, but certainly not least, to make it through this first filter, and still, only the first filter, sorry to say that you're gonna have to get in while the getting's good. Once again, two quotes.

Freshness. YouTube has a very dynamic dynamic corpus with many hours of video that are uploaded per second. The recommendation system should be responsive enough to model newly uploaded Newly uploaded Recreasing this recently uploaded fresh content is extremely important for YouTube as a product We consistently observe that users prefer fresh content Blah blah blah blah blah blah blah blah blah blah blah blah YouTube uses the word freshness a whole lot when it's talking about content Which personally makes me feel a bit uncomfortable makes me feel like YouTube is

going through my videos and squeezing them to judge them, like mangoes or other ripening fruit. Get your hands off my melons, YouTube! But think about it. YouTube wants to show you what's relevant right now. They've actually started caring even more about that recently. Which is why a lot of creators have noticed that they don't get video traffic for more than 48 hours, or at most a week. This is especially true for popular search terms. The window for trends like fidget spinners shuts fast.

So, uh, don't pull a game that. theory and do a video using fidget spinners as a case study for how the algorithm works a month after the trend is already moved on. That would be a bad idea. These videos take a long time to make, okay? It's not my fault, and fidget spinners are actually a very accurate example for all the points that we're trying to make in this episode. Just work! Just work! I'm just slow, okay? I'm just... just takes a lot of time to write these videos. It would have been relevant! Do as we say, not as we do. Which means that if you want to stay relevant to the YouTube algorithm, you need to be able to create a format of video that allows you to comment on popular things quickly. Again, not game theory.

Absolutely not game theory. Worst possible video style I could have ever chosen, but hey, I'm proud of what we make and we've been doing okay. Man, that's a lot of stuff and it's still only the first filter. Long, focused videos with searchable titles that need to perform well in the first couple of days. Holy cow! Well, here's some good news. There are a couple of things you can pretty much ignore,

namely stuff like likes and comments. That's right to put it in Google's own words, quote, Although explicit feedback mechanisms exist on YouTube, thumbs up, down, etc. We use implicit feedback. That means stuff like watch minutes and searches in the model. The choice is based on orders of magnitude, more implicit user history available, where explicit feedback is extremely sparse. End quote.

That means that compared to the sheer number of views, searches, and watch time, stuff like likes and comments don't stack up, and don't actually give a good indication of what is good and what is bad. So Wolfie, save your voice. When you ask for likes, you might be getting the warm and fuzzies, but you sure as heck ain't skyrocketing up the recommended charts. So congratulations, your video has made it through the candidate generator because it's new, it's long, it's searchable, and it targets a specific demographic. Man, and people think this job is easy. But wait, you are still not done.

Now you're competing for those top few spots. How can you bring home the win, or in this case, the click? The ranking neural network takes all the possible videos from the candidate generator and scores them based on what it thinks you're most likely to not just click on, but to watch for the longest. It does that by asking how long have other people watched it? Whether you're already subscribed to that channel and how many of that channel's previous videos you've watched? Those last two factors are where this gets really interesting for us creators, and they're also the two most important parts. As Google says, consider the user's past history with a channel that uploaded the video being scored.

How many videos has the user watched from that channel? When was the last time the user watched a video on this topic? These continuous features describing past user actions on related items are particularly powerful because they generalize well across disparate items If you watch your favorite channel regularly, you're gonna see a lot of videos from that channel recommended for you But say there's a channel where you skip a few videos or you just haven't watched a few videos on their topics in a while All of a sudden you'll stop seeing that channel in your recommended feeds Which for a lot of creators means the difference of millions of views But there's even a little bit more here, danky, my friend.

All these videos at the top are in hot competition, so YouTube is always A-B testing the top 10 or so recommendations. I might get Infinite Diet Coke Machine as my top video right now, but if I refresh the page, YouTube might swap that out for Diet Coke Addicts Anonymous. Another video that I'm very likely to engage with and one that I probably should engage with. Seriously though, pull up a tab with your logged-in homepage and try it for yourself. Those first few videos are the ones out of all of YouTube that YouTube things that you'll like the most. Wait a few seconds and refresh or just pull up the same account on your phone and you'll get a slightly Rearranged list with a lot of the same videos just in a different order. That's the ranking filter at work

A be testing you in real time. So what can you do to stay in it as a creator? Well consistency is more important than ever. If you produce one type of video, don't suddenly switch to something else because your subscribers might start skipping videos Which in turn will start filtering your content off their homepage. This syndrome hit the cover artist scene really hard a couple years ago Where one channel started vlogging between songs, others bandwagon to follow suit, and as a result all the channels died Because the vlog started to water down the content that everyone was coming for the cover songs It was basically a case of Lemmings. Let me follow this person blindly. Oh wait, now we're falling off a cliff In the gaming space it was a different but related problem for channels that only cover It was all well and good when Scott Cawthon was producing a game like every other week. There was plenty of content to sustain multiple videos

But now that it's taking more time to develop games the channels that built themselves off of fnaf and fnaf alone don't have content to fill their slates and slowly those channels start to wither away Unless they've trained their audience to watch stuff that isn't fnaf related also another recommendation produce content regularly people can't keep watching your videos if you don't produce them in the first place they're a very Very few old channels out there that do a pretty good job of getting around this system, but ones like epic rap battles walk a thin and very dangerous line. The first episode of each new season has to perform if they hope to retain visibility to the audience after each of their five-month long breaks. And if you thought all that was intense, get this. The ranking network already has access to stuff like the words that are used in the video and how good the thumbnail is.

Think that kind of thing is too hard for an algorithm? Well, think again! YouTube has had to have had a thumbnail algorithm for years. It recognizes things like the resolution of the image, whether the thumbnail contains a face, color, styles, all that. It has been training this algorithm for years now. Notice how it's gotten better at auto-generating thumbnails? Like, it used to just take random screenshots at various points throughout the video. Now, it tends to select things with a face. Sure, it's not very good, but it's improved, because it's studied thousands upon thousands of examples of thumbnails created by

some of YouTube's top creators and how many clicks they've gotten in places like recommended and suggested video feeds. On top of all of that, it knows the words you're using so it can create auto captions and filter you out of things like the kids app. But don't think Big Brother YouTube stops there. If you look at what else Google Brain has in the works, which yes, is a legitimate department at Google Google! And what do you want to be when you grow up Jimmy? I want to be a member of the Google Brain. It's creepy. If you look at what else Google Brain has in the works, it becomes pretty clear.

that YouTube would like to be able to decide much more than just whether a video is good or bad. Take for instance Google research blog about quantifying comedy. That's right, teaching a machine to judge whether a video is funny or not. You think that's impossible? Well, not so fast. Google is well on its way to gauging the funniness level of a video based on the words that people are using to describe it in the comments section. Are they lolling? Lolling, lullo-l-l-l-l-l-l-l-ing, or just plain lolling. It certainly matters to go Google, which means eventually it could be mattering to us. And finally, Google's latest project that they just published back in February is about object recognition in videos.

Right now, all YouTube can really recognize are a few images and rudimentary shapes. But soon, YouTube will know that this is a cat, and this is a mat pat. And we'll also be able to follow both of those things around the screen at all the different angles that we might be facing. This gives them a massive amount of information about what's in videos and could lead to a point where our titles and tags don't even matter anymore. Well, all of this seems intense, and it certainly is. The good part is that if the system is made right, it's unbiased, and genuinely shows you what is best for you based on the data,

on what you watch most frequently and what you watch the longest. It's a whole new world out there, but at least now you know the truth. The whole truth and nothing but the truth. From the brain of Google itself, translated by My Brain and Steph's Brain, to your brain. So go out there and conquer. I can't wait to see ya on the homepage. But hey, that's just a bunch of facts that I translated from a Google engineering paper to help my fellow creators. Thanks for watching!