YT YouTube2Text

YouTube2Text → Transcripts → How Social Media Algorithms Actually Work (And How to Beat Them)

How Social Media Algorithms Actually Work (And How to Beat Them) — Transcript

by Kallaway · 3,716 words · 540 segments · language en · Watch on YouTube

Open in transcriber Summarize with AI Download SRT/VTT

Full transcript

0:00Today we're talking about the social
0:01media algorithm. If you want to get more
0:03views with less effort, it's critical
0:05you understand how the algorithms
0:07actually work. And once you learn this,
0:09I guarantee you will never look at
0:11content the same way again. In this
0:13video, I'm going to break down how
0:14social algorithms work, why they pick
0:16certain videos to push over others, and
0:19the specific things you can do to make
0:21them prioritize your content. Now, this
0:23information is based on a ton of outlier
0:25data and comments made by the Instagram
0:28CEO himself. So, this is the latest and
0:30greatest for what's actually working
0:32right now. If you just follow this, your
0:34content will perform way better. By the
0:36way, if you don't know me, my name is
0:37Callaway. I have a million followers,
0:38I've done billions of views, and content
0:40is all I do all day long. All right,
0:43first, let's just start with how the
0:44algorithms actually work. And this is
0:46actually super helpful to understand.
0:48Once you hear it, it'll make a lot of
0:49sense. Social media companies only have
0:52one goal, to keep people on the platform
0:54as long as possible. When people stay
0:56longer, they watch more ads, and the
0:58companies make more money. It's as
0:59simple as that. Now, to keep you on the
1:01platform longer, they do their best to
1:03serve you the content they think you'll
1:05enjoy the most. And that means the
1:07algorithm is just one giant matchmaker.
1:10It's matching people with content. If it
1:12does a good job with this matching,
1:13you're going to keep watching and stay.
1:15But if it starts doing a bad job and it
1:17shows you stuff you don't want to watch,
1:19well then you're going to leave. It
1:20sounds simple, but this is how social
1:22algorithms work in a nutshell. Now,
1:24here's why this matters for you. If you
1:25want to hijack the algorithm and make it
1:28push your video more, all you have to do
1:30is help it make better matches with your
1:32content. [music] So, in this video, I'm
1:34going to explain exactly how to do that.
1:36And this really is the highest leverage
1:37algo hack you could ever learn. Okay,
1:39now here's how this matchmaking process
1:41actually works under the hood, so we can
1:43understand exactly what to do to
1:45manipulate it. When you post a video on
1:47social media, the very first thing the
1:49platform does is analyze what that video
1:51is about. I call this a digital
1:53fingerprint. Now, this analysis is
1:55multimodal, so it's watching your video
1:57with computer vision to understand
1:59what's going on visually. It's listening
2:02to your video with audio fingerprinting
2:04to get a better understanding of the
2:05transcript and what's actually being
2:06said, and it's also reading all the
2:08metadata, the caption, the hashtag, the
2:11creator, the location, anything else it
2:12can find. It then combines all that
2:15information together in real time to
2:17build a single contextual understanding
2:19of the video. This is called a topic
2:22mapping. Now, based on that topic
2:23mapping, the algorithm builds a fit
2:26score, which is its prediction for who
2:28it thinks will best like this video. So,
2:30at this point, you've posted it, it's
2:32analyzed it, but it hasn't been shown to
2:34anyone yet. So, we're ready to start
2:35showing the video to people. But, this
2:37is where things get interesting, because
2:38obviously the algorithm doesn't just
2:40blast your video off to millions of
2:41people right off the bat, or you'd have
2:43millions of views. This is what actually
2:45goes on under the hood. The algorithm
2:47uses its fit score to pick roughly 200
2:50people to show the video to first. This
2:52is called the initial sample test group.
2:55If it could rank all 100 million people
2:58that are on the app at one time, this is
3:00the group of 200 people it thinks will
3:02like the video the most. Now, very
3:04important, of these 200 people, most of
3:06them are non-followers, because the
3:09algorithm wants to test how well
3:10strangers react to your video. It knows
3:12followers should like it, cuz they
3:13already follow you, but if strangers
3:15like it, too, well, then that means this
3:17is a really good video. This is why when
3:19people say followers don't matter
3:21anymore, they're kind of right. They
3:22don't matter in the sampling process,
3:24because most of those 200 people are
3:26non-followers. They're strangers to you.
3:28Now, based on the metrics from this
3:29initial sample group of 200 people, the
3:31algorithm is going to get positive,
3:33neutral, or negative data back.
3:35Essentially, of those 200, how many of
3:37them liked and watched the video? What
3:39was the set of data? If the data is
3:41positive, that means the algorithm's
3:43guess of the fit score was accurate. And
3:46so, it knows exactly what type of person
3:48to push the video to further. The next
3:50time it pushes, let's say it's 2,000
3:52people. And if that's good, then it's
3:5420,000 people. And if that's good, then
3:56it's 200,000 people. And it just keeps
3:58going until the data starts coming back
4:00weaker. Now, if the original data was
4:02neutral, kind of good, kind of bad, the
4:04algorithm will redo its fit score and
4:06push the video again. But this time,
4:08only to maybe another group of 200. It
4:10doesn't go crazy to 20,000. It just
4:13resamples. If the data is negative right
4:15off the bat with those 200, well then
4:17the algo's going to tighten up and stop
4:19pushing almost immediately. And again,
4:21it slows down that push because it
4:23doesn't want to risk alienating people
4:25and pushing them off the platform
4:27because they see a bad video. So, if
4:28you're in the 200 view jail, or you post
4:31a video and it flops, what this really
4:33means is that the algorithm got bad data
4:35back from that initial sample group of
4:37200. Now, one more thing before we move
4:39on this. The reason why even million
4:41view banger videos eventually slow down
4:43is because even those run out of people
4:46that want to watch it. The data gets
4:48bigger, bigger, bigger, and then
4:49eventually it starts to fade off. So, in
4:51a nutshell, this is how social
4:52algorithms actually work under the hood.
4:55This is what happens when you go to post
4:56a video. You post it, it does the topic
4:58mapping, it samples with a group of
5:01roughly 200 people, and then it either
5:03boosts, retries, or stops immediately
5:06based on how the sample data comes back.
5:08So, what does all this actually mean for
5:10you tactically? Knowing this is
5:12happening under the hood, how can you
5:13best adjust your content strategy to
5:15take advantage and get the algorithm to
5:17push you more? That's what we're going
5:19to go through right now.
5:22If you want to hijack the algorithm to
5:24get more views, you only need to do two
5:26things. Number one is to help the
5:28algorithm build a better fit score, so
5:30that it finds the best possible sample
5:32group of 200 people to go to first. And
5:35then number two is to make sure that
5:36sample group actually likes and engages
5:38with your video. Because if the sample
5:40group is the right fit, and they
5:42actually like your video, well then
5:43you're going to be in great shape
5:44because the algo will get amazing data
5:46back and just [music] keep boosting you
5:48to more and more people. So, what I'm
5:49going to do now is break down the
5:51tactics for how to trigger both of those
5:53things. Helping find the right sample
5:55group, and then helping make sure that
5:57sample group engages well. This is
5:59essentially the systematic process for
6:01activating the algorithm and getting it
6:03to work for you. Now, if you like how
6:05I'm breaking this down, kind of from
6:06like a scientific and psychology
6:08perspective, I actually just published a
6:10free guide doing the same thing for my
6:12entire content system. It's from ideas
6:15to hooks all the way to monetization.
6:16Science-based and data-backed. This is
6:18the exact content system I ran last year
6:21to generate hundreds of millions of
6:22views and millions in profit from
6:24content. It's also the same thing I
6:26install with business owners when I work
6:28with them one-on-one. Completely free,
6:30my gift to you. You can get it below or
6:32at the link shortformsytem.co.
6:34All right, let's talk sample groups.
6:36What can you do to help the algorithm
6:38build a better fit score and find a
6:40stronger sample group for your video?
6:42Here's the answer, very simple. All you
6:44have to do is consistently make videos
6:46about the same topic for the same
6:48audience avatar over and over and over.
6:51You want to become intentionally precise
6:53and narrow with the topics you pick and
6:55how you position. Here's why. After a
6:58few similar videos in a row, the
7:00algorithm will start to understand that
7:02your channel talks about X topic for Y
7:05avatar profile. It will then have built
7:07sample groups for all those previous
7:09videos and have a very clear
7:10understanding of who to go back to from
7:13an avatar perspective. The more this
7:15avatar group is the same over time, the
7:17more confident the algorithm can be when
7:20it dials this in. Now, this process of
7:22consistently making videos about the
7:23same few topics for the same avatar is
7:26called audience matching. And if there's
7:28one content principle I swear by, it's
7:30this. When you make videos for lots of
7:32different topics for several different
7:34avatars, the sample data comes back
7:36mixed and the algorithm gets confused.
7:38When it's confused, it pushes your video
7:40less because it doesn't want to risk bad
7:42fits to bad viewers. For example,
7:45imagine you made three videos. One on
7:47tech, then the next one on health
7:48trends, and then the next one on
7:50politics. The algorithm would have no
7:52idea what your fourth video is going to
7:54be about. And because of this, it
7:56doesn't know which of your previous
7:57three videos it should model its fit
7:59score after. So, let's say your fourth
8:01video also ends up being about health
8:03trends. Chances are the algorithm's
8:04going to build a blended fit score
8:06across those first three videos. A
8:08little bit of people from tech, a little
8:10bit of people from health trends, and a
8:11little bit of people from politics. Not
8:12literally those people, but influence
8:15from who liked those videos. And when it
8:17does this, the fit score targeting for
8:19your fourth video is going to be a mix
8:21of all three. And so, when it pushes it,
8:23of course, the sample data for the
8:25health trends video that also has tech
8:27and politics type viewers is going to
8:29come back weak. This will result almost
8:32certainly in your video flopping. What
8:34this means in simple terms is that if
8:35you want to help the algorithm find the
8:38right sample group and build a better
8:40fit score, you got to keep your topics
8:42and audience selection narrow
8:43consistently. And this means sometimes
8:45saying no to ideas that seem viral, but
8:48would resonate with the wrong audience.
8:50Even one viral hit to an audience
8:52outside of your core demo will result in
8:55the next several videos having poor
8:57sample data because it confuses the
8:59algorithm. This discipline in topic and
9:01audience selection is very important,
9:03and typically beginners that are
9:04starting out are spraying and praying
9:06all over the place, and they don't have
9:07this. Okay, so that's one side of the
9:09equation. Very simply, just narrow your
9:11topic and audience, and your sample fit
9:13will go up. Now, on the other side of
9:15the equation, the two-part piece was
9:16making sure, once you have that sample,
9:18that it actually engages well with your
9:20video. And that means they watch it,
9:22they like it, they save it, they share
9:24it, they comment, they repost, all the
9:26engagement metrics, as many as we can
9:27possibly get.
9:28>> [music]
9:28>> So, what can we do on this side to make
9:30sure that initial sample data from these
9:32people comes back strong? Well, when the
9:34algorithm's gauging if it's strong or
9:37not, it's really only looking at three
9:38core metrics. The first one is average
9:41watch time. How long did someone watch
9:43in a number of seconds per video? And
9:45also, by proxy, percent completion. What
9:47percent of the video was completed on
9:49average? The second metric is engagement
9:51rate. So, this is likes plus comments
9:53plus shares divided by views. And the
9:55third, which is really important and
9:57nobody can access, is called watch time
9:59session share. In a session for a
10:01viewer, if they're on there for 60
10:02minutes, how many of those 60 were spent
10:04watching your videos? That percentage is
10:07watch time session share. You can't
10:09access this anywhere, but it's a
10:10critical metric that social algorithms
10:12use to know how influential your content
10:14is.
10:15>> [music]
10:15>> So, the million-dollar question for you
10:16is how can you improve these metrics?
10:18What can you do in your video to make
10:20sure those metrics go up so the sample
10:22data comes back clean, so that it just
10:23pushes your video to more people? Well,
10:25a short, cheeky answer is if you want
10:27the metrics to go up, you just make
10:29better videos with better ideas,
10:31stronger hooks, better storytelling, and
10:32more interesting visuals, obviously.
10:35But, that's not helpful at all. So, is
10:36there anything tactical you can do at a
10:38studs level to increase the
10:40effectiveness of those videos? And if
10:42you watch this channel a lot, you
10:44already know, of course there is.
10:45There's only four things you need to do
10:47to make your video better so that that
10:48engagement rate goes off the charts.
10:51>> [music]
10:51>> Number one is that the topic needs to be
10:52relevant for the ideal viewer. This is
10:55obvious, and it goes with the first
10:56piece I said. What you cover has to
10:58actually solve a problem that they have.
11:00If that's the case, engagement will go
11:02up. Number two, the information you
11:04share needs to be both non-obvious and
11:07tactically implementable. Is it new
11:09stuff they haven't heard before, and can
11:10they actually use it to solve that
11:12problem? If those things are true, the
11:14engagement rate will go up. Number
11:16three, the viewer has to actually have a
11:18high absorption of the information you
11:20say. It could be on target and
11:22non-obvious, but if they can't actually
11:23understand what you're saying, then they
11:25can't apply it. So, if they could apply
11:27it, the engagement rate will go up. And
11:28then number four, there needs to be a
11:30short distance to implement your
11:32recommendations. Tactically
11:33implementable means they can take a
11:35little bit of action and get a big
11:37result based on your promise. Now, you
11:39won't typically hear people frame it in
11:41this way, but if your content has those
11:42four attributes, I guarantee you're
11:45going to have higher engagement. If you
11:46have higher engagement, the data comes
11:48back more positive, they push it to more
11:50of the people, those people are people
11:51you want, and the flywheel spins. What
11:53this really means, those four things in
11:55layman's terms, you got to cover a core
11:57pain point or problem they have. You got
11:59to have something useful or interesting
12:01to say. You got to say it in a way they
12:03can actually understand, and they have
12:05to be able to take what you say and
12:07apply it on their own. Those are the
12:09four horsemen to driving good video
12:11performance. If you do this, you're set,
12:14and that's really all you need to hijack
12:15the algorithm to push you more. Pick an
12:18avatar, stick to it, narrow your topic
12:20selection, and then drive those four
12:22things home. When you do this, the
12:24sample group will stay dialed, and
12:25they'll all engage with the video at a
12:27high rate. Incidentally, these four
12:29factors are also how you turn viewers
12:31into buyers. If you want people to buy,
12:33those four components also make sense to
12:35include in the video. They're kind of
12:36like the core DNA if you're trying to
12:38build a money machine with content. Now,
12:40I'll say this, the easiest way to make
12:42sure you're picking the right topics
12:44that actually work for your avatar group
12:46is to just study the videos that are
12:49already working in your niche. It shocks
12:51me how few people actually do this, but
12:53in sandcastles.ai, you can build a group
12:56of competitor channels that are already
12:58crushing, and just filter by outlier
13:00score to see all the best performing
13:02videos. There's now this feature where
13:04if you save the video to library, you
13:07can see all the attributes. I'm talking
13:09transcript, topic, the exact hook, the
13:12exact storytelling mechanics, everything
13:14about the video that drove the
13:15curiosity. You can then take that, remix
13:17it. Everything you need is right in
13:19there. So, all you need to do if you're
13:20confused on which topics to pick for
13:23your avatar group, just go in
13:24sandcastles and use this resource. It
13:26shocks me how few people are using data
13:29to make their topic decision. This
13:31basically guarantees that you're serving
13:32the right stuff to your audience. Now,
13:34before I end this video, I just want to
13:35include one more bonus topic around the
13:38algorithm. Because I know my explanation
13:39is a little bit theoretical, hopefully
13:41it made sense, hopefully it helped you,
13:42you have action items to work on. But, I
13:44just want to include one more thing at a
13:46tactical level that you can take away
13:48and really hammer value from this video.
13:50Another way to drive algorithmic push is
13:52to increase the number of comments on
13:54your video. Most people know this. There
13:56are five things you can do tactically to
13:58increase the number of comments you're
14:00getting. Number one is to take a hard
14:02stance on your topic. People typically
14:04comment when they violently agree or
14:06disagree, mostly disagree, with whatever
14:09your stance is or perspective. If you
14:11play the middle and hedge, you're going
14:13to get fewer comments. So, I recommend
14:15you pick a side, pro or con, and that
14:17will drive comments. [music] Number two
14:19is to pick the side that is the
14:21contrarian side. Like I said, people
14:22love to comment when they disagree, when
14:24they think you're wrong. If you pick the
14:26contrarian side, that means you think
14:27the majority of people are wrong, which
14:29means they'll think you're wrong. The
14:31majority of people will want to comment
14:32because they disagree with you. If you
14:34create more enemies, you drive more
14:36comments. Tip number three is to amplify
14:38the stance you take by ratcheting up the
14:40way you frame your points. If you said
14:42something like, "This is the best way to
14:44cook pasta." versus, "This pasta is
14:47better than all the mom and pop pasta
14:49shops in the world." Which one is going
14:50to drive a more violent discussion in
14:52the comments? The more extreme version,
14:54of course, always is. So, that's how you
14:56ratchet up your stance. Tip number four
14:58is to build your topics around
14:59cult-loved brands, people, ideas, and
15:03movements. The more you talk about
15:04things people already have made up their
15:06opinion on, the faster they're willing
15:08to jump into the comments, especially if
15:10they disagree. For example, if you take
15:12a stance on Nike versus just the
15:13category of shoes, more people will have
15:16already made up their opinion whether
15:17they like or dislike Nike, and it will
15:19trigger them to comment. Tip number five
15:20is to position your take or stance to
15:22drive significant emotion. The more
15:25people feel something when they watch
15:26your video, the more they're going to
15:27feel compelled to want to comment. Now,
15:30those five things around comment, that
15:31was just a little extra sprinkle to give
15:33you more tactics on how to drive
15:35activation and engagement to make the
15:38algorithm push you. All of those feed
15:40back to picking the right topic for the
15:42right group. So, it's kind of like a sub
15:44point on what I just went through.
15:45Hopefully that's helpful and you can put
15:46that to use. All right, guys. That's all
15:48I've got for this video. As a recap, we
15:49covered a lot of ground. We really broke
15:51down the ins and outs of how social
15:52algorithms work and how you
15:54theoretically and fundamentally can
15:57hijack them to push your videos more. I
15:59tried my best to kind of demystify this
16:01black box and give you tactics that you
16:04can use to your advantage. As always,
16:05guys, I'm trying my absolute best to
16:07give you perspectives that most people
16:08don't cover in a tactical bite-size way
16:11that you can actually put to work. What
16:12you typically see about the algorithm is
16:14talking about settings, hacks, or these
16:16little caption tweaks. None of that
16:18stuff actually works. If you think that
16:20actually works, you need to watch this
16:21video again. Posting time does not
16:23matter. Hashtags in your captions don't
16:25matter. The captions themselves don't
16:27matter. The only thing that matters is
16:29making great videos for a specific
16:31avatar group across a narrow band of
16:33topics over and over and over. That's
16:35the only thing that matters. That is the
16:37cake. Everything else is the icing.
16:39Focus on the cake. As a reminder, if you
16:41want access to my full content system,
16:43this is the exact blueprint I use with
16:45my own team. How I find ideas, how I
16:47write hooks, how I validate these
16:49things, how I do research, my entire
16:51funnel to turn viewers into dollars,
16:53literally everything. I've got it linked
16:54below for free, my gift to you,
16:56shortformsytem.co.
16:58And if you have any other topics around
17:00social media growth, around content and
17:02content systems that you want me to
17:04cover or that you feel blocked on, just
17:06drop them in the comments. We use the
17:08comments to inform the database of
17:09videos that we make next. So, anything
17:11you guys are stuck on, please put it in
17:13the comments. It will help greatly to
17:14inform what we should make. All right,
17:16guys. That's all we've got. We will see
17:18you on the next video. Peace.

About this transcript

This page contains the full transcript of How Social Media Algorithms Actually Work (And How to Beat Them) by Kallaway, generated from the public captions YouTube serves with the video. The transcript has 3,716 words across 540 segments, with the original timestamps preserved so you can click any line to jump to that moment in the embedded player.

What you can do with it

Use the transcript to take notes, quote the speaker, build a study guide, generate a summary with ChatGPT or Claude via the YouTube Summary tool, or export it as a timed subtitle file with YouTube to SRT. You can also re-open it in the transcriber to choose a caption track available on the video.

Free YouTube transcript tool

YouTube2Text is a free YouTube transcript generator — no signup, no daily limit. Paste any YouTube link and get the full transcript instantly, with timestamps, click-to-jump, caption language selection, AI prompts for ChatGPT, Claude, and Gemini, and exports to TXT, SRT, VTT, or Markdown.