Cracking YouTube in 2017 with Matt Gielen – TubeTalk #141

YouTube Algorithm Matt Gielen

Last updated on June 14th, 2024

YouTube Algorithm Matt Gielen

YouTube famously holds their algorithm close to their chests, and the only way to really understand what works is to gauge the results of thousands of videos and channels. Luckily, Matt Gielen of Little Monster Media has done just that and released the findings in a report titled Cracking YouTube in 2017.

In this episode of Tube Talk, Dane Golden sits in for Jeremy Vest and talks to Matt about his research and what we as creators and brands can take away from it to be more successful on the platform.

Enjoy, and please share if you find this episode valuable!

Transcript:

Dane Golden: This is Dane Golden of VidActionTV, sitting in for Jeremy Vest of Vidpow. Today we have Matt Gielen, G-I-E-L-E-N, of Little Monster Media Co. When it comes to how YouTube works, Matt is just the guy that sees the matrix. I’ve said this before, and I’ll say it again.

Matt Gielen: Well thanks Dane.

Dane Golden: Yeah. And Matt released his third version of reverse engineering the YouTube algorithm, or whatever it’s called, and it’s Vidcon. What’s it called?

Matt Gielen: The report we released was Cracking YouTube in 2017, and the presentation was just reverse engineering the YouTube algorithm, and it kind of carried off of the previous blog posts that I’ve written, Reverse Engineering the YouTube Algorithm, Part 1 and Part 2, on Tubefilter.

Dane Golden: Yeah. So you can get that info on Tubefilter or on LittleMonsterMediaCo.com. Is that correct?

Matt Gielen: That is correct.

Dane Golden: Matt, I want to jump right in, and ask you … Maybe it’s a basic question, but in some of the stuff you just released, you’re talking a lot about video planning and structure really. You’re talking about how long videos should be, and how often videos should be posted. And frankly, a lot of people aren’t gonna be listening, or posting YouTube, and say, “Listen. I just want to upload my videos. It’s your job to tell me how to get people to watch them. Why do I need all this math, and can’t you just get me better distribution? Why do you have to tell me how long videos need to be, or how long that they should be posted?” How would you respond to that?

Matt Gielen: Oh, coming right out of the gate firing there, Dane.

Dane Golden: Boom.

Matt Gielen: Well, I would take a deep breath, and a sip of water. And then I would probably say, “I get that, and I understand, but it’s a bit of a symbiotic relationship, where you have to design programming, and your programming strategy around the platform on which you’re distributing that content. And what we’ve seen from the platform, and from the audience on that platform, in this case YouTube, that what they want are videos that appear to be, or what it appears that they want, is videos between seven and sixteen minutes in length. What we’ve also seen is that there appears to be a direct correlation between volume of videos you’re posting, and the amount of viewership that you’re getting. In that regard, if those are two of the biggest driving factors of viewership, where we see the strongest correlations, then my advice would be to create content within that framework, and if you don’t want to create content within that framework, then you just have to accept the fact that you’ll probably, or potentially generate lower performance as a result.

Dane Golden: Let’s talk about some of the new numbers you’ve come up with. You said there’s a certain length of videos that do better, and a certain frequency that they come out. So what are those numbers?

Matt Gielen: Yeah. I’ll answer your second question first, which is about frequency of post. Essentially, what we saw was that, for videos that generate between five million and twenty million views, those videos are being posted to channels that, on average, are uploading 91 videos every six months. For videos that do over 20 million views, the average volume of distribution from those channels, are 115 uploads per month. Generally speaking, you can say something along the lines of, “Well, if there’s 26 weeks in half a year, and half a year, you see 91 uploads, then they’re posting a little over three videos per week, for videos that get between five million and twenty million views. So that’s a pretty high frequency. And then when you go to 20 million plus, the average there was about 115 videos per six months, which equates to nearly four and a half to five videos per week. So the greater the frequency of upload onto your channel, we see there’s a strong correlation between the average views those videos get.

Now, that doesn’t mean necessarily, that if you just start pumping out three, four videos a week, your views are gonna be between five million and twenty million. It’s just the average that we saw. And we should also footnote that, by saying that we looked at channels that had 50,000 subs or more. So that’s kind of the threshold that we were looking at.

Dane Golden: So there wasn’t a lot of … I’m guessing there’s not a lot of brand videos you’re looking at. These are mostly YouTubers, because YouTubers … Some brands get more subs than that, but mostly they’re YouTubers and media companies, or do you know?

Matt Gielen: Well, all the data was provided to us, anonymized. There are not a lot of consumer package good brands, that have more than 50,000 subscribers on YouTube. There are definitely some, but for the most part, there’s not. But, that being said, I would say that I look at every YouTube channel as a brand. And whether you’re an individual YouTuber, a media company, consumer packaged goods, software as a service. Any field you go into, as it pertains to YouTube, you are a brand. So in that regard …

Dane Golden: The rules still apply the same. Is that what you’re saying? Regardless of what the video’s about?

Matt Gielen: Yeah. I think when people hear brand, they think of someone who’s gonna advertise on my videos, or run a lot of promotion against their videos, to juice the views, or jack up the views. And it’s perfectly possible that some of those channels, and some of those videos were taken into account, within these metrics, but I think that was kind of the benefit of doing 1,100 channels, and 70,000 videos, where any one channel, or any one video isn’t going to totally screw everything, in terms of giving us good data.

Dane Golden: And did you say the length of the video?

Matt Gielen: Oh, in terms of the length of the video, what we found was videos that were between seven and sixteen minutes got up to 50% more views, especially in relation to videos that were shorter than that. And when you looked at different time frames … The time frames we looked at was one day, two days, seven days, and thirty days. And the biggest difference was, at the seven day mark, so videos that were seven to sixteen minutes at the seven day mark were getting about 20% of the number of subscribers on that channel to watch in raw viewership, whereas videos that were shorter were only getting about 14%. Videos that were longer weren’t too far behind. They were at about 19%, but it was still lower. But, what we saw was when we got out to the thirty day mark, videos that were seven to sixteen minutes in length were getting about 32%, whereas videos that were longer than that were only getting about 27%, and videos that were shorter were only getting about 26%. So the difference really got more significant as time went on.

I want to back up just one second, and talk about the methodology here, because it’s very, very difficult to compare-

Dane Golden: Just so I … Just to interrupt you Matt. The number, generally, that you said was the best, was between seven and sixteen minutes, generally. But now you’re gonna tell me why.

Matt Gielen: Yeah. It’s a curve. It’s definitely a curve, but the methodology behind why we look at the viewership as a percentage of subscribers, to each individual channel, is so that we can weight the videos more accurately, based on the potential reach of any individual channel, within its subscriber base. Now, there’s all sorts of differences in there, but it’s really the only other way, other than potentially, monthly views, or six months views, to really do apples to apples comparisons on individual videos, and those videos’ performance. If you had 50,000 subscribers, and you post a video, and on the first day, it does 100,000 views, that’s far more impressive than if you have 200,000 subscribers, and on the first day, it does 100,000 views. The percentages are massively different. So, by weighting them, based on the number of subscribers you have at the time of release, it gives us a much more accurate picture as to the potential success of a video across a wide array of channels.

Dane Golden: But you’re not … I think you’re probably not saying that I could just shoot a video of birds at the park for 10 minutes, and post that three times a week, and I would be a YouTube star.

Matt Gielen: No, I’m definitely not saying that, because I would highly doubt that your subscribers would come back time and time again, to watch that video. There’s a lot more in there. What I equated it to in the presentation was, you can’t really sit down and tell an editor, “Make me a video that gets a five to eight minute average view duration.” What you can do is you can do some back of the napkin math, and determine roughly how long your video needs to be, in order to get that five to eight minute average view duration, which is what we saw, the pretty big sweet spot, in terms of viewership for any individual video, both on a one, two, seven, and thirty day time frame. Five to eight minutes is kind of a sweet spot for average view duration, you can’t really set out to craft a video to do that, until you’ve posted some stuff. And then, once you do expand the length, assuming you had to expand it, or shorten the length, assuming you had to shorten it, until you’ve posted some of those videos, and seen what the average view duration was.

Now, that’s not to say that that’s the right solution for everyone. There are curves, and there are plenty of outliers, especially when you get to the 30 day mark. What we’re saying is that we see this pattern, broadly.

Dane Golden: Could you talk … Matt, you’ve spoken about before, and not necessarily this time, but some of the things I see, because I work with brand channels a bit. And by that, I mean, sometimes B to Bs, sometimes B to Cs. And sometimes they’re posting a ton of videos, great great frequency, but they seem like they’re sort of locked in at a very low number of views, and sometimes the views get worse, the more videos they post. You’ve spoken about, in the past, about how, if your video gets somehow below average, future videos on that channel, maybe on a totally different topic, are generally gonna do worse. Is that, am I saying that correctly?

Matt Gielen: Excuse me. Yes, generally speaking. In Reverse Engineering the YouTube Algorithm 2, which was based on the Google paper, Deep Neural Networks for YouTube Recommendations, and in the presentation I give to Vidcon, essentially what you’re discussing is a concept of collaborative filtering, or essentially just individual user preference. What ends up happening is, let’s say you have a channel that has like, 10,000 subscribers, and you’ve been posting regularly, once or twice a week, and all of a sudden, you start uploading 10, 15, 20 videos a week. Well, of those 10,000 subscribers, if you were getting 5,000 of them to watch every video you put up, and all of a sudden, there starts to be tens and twenties, and hundreds of videos going up, that they can’t watch, but YouTube is still recommending those videos to them, essentially over time, with the machine learning going on at YouTube, empowering the algorithm, they’re going to stop serving those videos to people that have not expressed an interest, through viewership, your videos. That’s essentially the potential outcome of just increasing the frequency of posts, without increasing the value of video, or at least keeping the value of the video the same, to the audience.

Dane Golden: Would you rank the value … Would you just say the value in this case is simply, the audience has determined they want to watch it for a significant duration? Would that be a good shorthand of saying what would be value, in this situation?

Matt Gielen: Potentially. It’s that, combined with their click through rate on your videos, and how that compares in relation to other videos on that topic, and how that compares in relation to other videos, not on that topic, that YouTube thinks that person might be interested in.

Dane Golden: But, in short though, a poorly watched video, can it affect future videos?

Matt Gielen: Potentially. I think we’re seeing less of that, generally, across YouTube, where, let’s say you have B content strategy, where you upload behind the scenes, or this and that. We’re seeing that have far less of an impact on viewership than it used to have. And what it does mean, is that YouTube is getting better, I think, better at categorizing videos, and showing potential videos to your subscribers, that they’ve shown an interest in, a particular series or something like that, and other ones that they’ve shown they’re not interesting in, not showing them those videos.

Dane Golden: Okay.

Matt Gielen: YouTube likes to say that the algorithm is audience driven, which is a very convenient way for them to not take responsibility for it, while simultaneously keeping complete ownership over it, if that makes sense. What we’re seeing more and more is that, if you think about the algorithm in terms of an equation that has a goal of generating watch time, and then you think of the individual metrics in that equation, or variables in that equation, as all the potential data and metrics YouTube is learning about individual viewers, and about videos, and about channels, the weighting on, W-E-I-G-H-T-I-N-G, weighting on the various elements can be dialed up and down. My assumption is that there’s very little of that being manually done, and the vast majority of that is being done [inaudible 00:16:24] machine learning, and the deep neural network.

Dane Golden: Well what about … You’ve also mentioned that day one viewership … Basically, you can get the most number of views, or the highest percentage of views in relation to subs. Basically day one, you’re saying, is super important. And clearly, I’m assuming you can’t just run the video as a pre-roll, and get views that way, or pay robots to watch it. How do you … This is sort of the whole thing of audience development, but how do you get them to watch on that first day?

Matt Gielen: Yeah. That’s the biggest first part, where your title and thumbnail have a very out-sized importance on your first day performance. But I’d say, even beyond that, both of those aspects of your video, stem from the programming of choice, what you choose to make, how you choose to make it, what format it’s in, what style it’s in, what topic it covers, is gonna have the greatest impact on your overall viewership, because that, in large part, is gonna determine your title and your thumbnail, assuming you’re trying to accurately title and make a thumbnail.

Dane Golden: So you’re saying just talk about video games.

Matt Gielen: If it were only that simple. But all of that demands a deep understanding of your audience, so if you don’t understand your audience, you’re not gonna understand what those answers should be, in terms of, what should we make, how should we make it, what should we make it about, when should we post it, et cetera, because if you don’t know what your audience is into, people who have watched your previous content, people that have subscribed to you, people who have opted in for notifications, you’re just gonna be shooting in the dark, and saying, “Well, I think they like this. I think they like that. I’ll make more like that.”

Dane Golden: I asked you a very unfair question, basically. How do you do what you do, in one sentence. But I think you did mention that the, at Vidcon, that optimization begins at the drawing board, or did you say that?

Matt Gielen: Optimization begins on the whiteboard. Yeah.

Dane Golden: The whiteboard. Yeah, you did say that.

Matt Gielen: Mm-hmm.

Dane Golden: And-

Matt Gielen: In a lot of ways, that’s great news for producers. There’s never been a better time to be a producer on YouTube where you can actually understand less about the algorithm, and be better off. Now, that doesn’t mean that there isn’t a tremendous amount of value in a deep knowledge of the platform, and a deep knowledge of the audience. It just means that there’s probably a bit less weighting to people who are growth hackers, who are just kind of doing heartless or soulless content, that is designed just to drive viewership, but it does mean that they have a greater chance of prospering on the platform.

Dane Golden: I wanted to ask you one last question, Matt. You talked a lot about freshness, versus relevancy, in the paper. Does this mean people are not gonna see my old videos anymore?

Matt Gielen: No, definitely not. There’s a couple things in there. First of all, YouTube is placing more emphasis, especially as it pertains to suggested videos, and the browse page, on fresh videos, or new videos. One of the rumors floating around YouTube is that this is because they sped up their processing time significantly, and it was just a logical outcome of the processing time, because people prefer fresh content, stuff that they haven’t seen before. But no, that definitely doesn’t mean that the library of content won’t continue to be extremely valuable. How to Build a Chair, if you have the best how to build a chair video, will always be relevant to someone searching for how to build a chair, unless something crazy happens in the woodworking industry, that I’m not aware of, where how to make a chair becomes irrelevant, because we all have 3-D printers in our space capsules, or whatever it is.

Dane Golden: So old content can still get surfaced, if it’s good.

Matt Gielen: Yeah. 100%. Absolutely. If a video is going to drive “watch time on site,” it will still be served. Now, what we are seeing is there appears to be a greater impetus to promote new videos on new videos, and to promote older videos on older videos. So, if you do evergreen content, then chances are you’ll have a higher percentage of older videos in your suggested videos from yourself, than you will for newer, more trendy, topical type stuff.

Dane Golden: Okay. And, Matt Gielen, where again can they find this study, and where can they find you?

Matt Gielen: You can find it, and me, on LittleMosterMediaCo.com.

Dane Golden: That’s Matt Gielen. He sees the matrix, when it comes to YouTube. My name is Dane Golden, from VidActionTV. Find me there, or at Dane Golden Everywhere, and also find Jeremy Vest, of Vidpow, at Vidpow Bam on Twitter, and a Vidpow.com. Until next week, keep talking tube.

VidAction
Scroll to Top