So What? Marketing Analytics and Insights Live
airs every Thursday at 1 pm EST.
You can watch on YouTube Live. Be sure to subscribe and follow so you never miss an episode!
In this week’s episode of So What? we focus on understanding social media algorithms. We walk through the academic papers, how to pull out the right information, and what actions to take. Catch the replay here:
In this episode you’ll learn:
- Where to find information about how social media algorithms work
- How to read through the papers and pick out important information
- A tour through LinkedIn and Twitter
Have a question or topic you’d like to see us cover? Reach out here: https://www.trustinsights.ai/resources/so-what-the-marketing-analytics-and-insights-show/
Katie Robbert 0:23
Well, hey friends, happy Thursday. Welcome to so what the marketing analytics and insights live show. I am Katie joined by Chris and John. How’s it going, guys? Good, Lou. On this week’s show, we are talking about demystifying social media algorithms. LinkedIn is one of the most quote unquote locked down systems in terms of being able to get data out of it. Twitter is forever evolving. There was even a big announcement this morning, which I know Chris will get into. But believe it or not, these social media platforms actually do tell you how they work. They just don’t do it in a way that most of us, myself included, can understand. They publish deeply technical, academic and development papers with formulas and words that have six and seven syllables, and you need a dictionary and a thesaurus. And so it’s not for everybody, not everybody can just pick up this paper go, oh, okay, now I know that I need to post it two o’clock instead of three o’clock. It definitely isn’t an instruction manual. That way. So on today’s episode, we’re going to cover how we go about finding out the information so that we can then share it with you. So where to find the information, how to decode it, demystify it, and pick through it to find the relevant information, and then what the heck to do about it. And we’re going to focus specifically on Twitter and LinkedIn. So Chris, where there’s been a lot going on with these platforms, where would you like to start?
Christopher Penn 2:02
I mean, that’s a very good question. So a lot of the stuff revolves around people asking the question, you know, does this work? Or does that work? And this topic came up, because I was chatting with a group of PR folks, actually, who were dispensing all kinds of advice, like, Oh, you shouldn’t have more than three hashtags for this. But you shouldn’t post more than this many times in a row. And I’m like, where did you get this information like? And it’s funny, it’s almost like urban legends and myths that just get passed down. That, you know, those things might have been true at one point in time, just like old SEO myths, like, oh, you should put like all your work keywords in white text at the bottom of your page, you know, stuff that worked great in 2006, but not so much 17 years later. And so that’s where the this this topic got started, was listening to the advice people are passing around. That sounds credible. What does doesn’t have any basis in fact? So with all these networks, the the number one question sort of the purpose, if you think about the five P’s, the purpose is, we need to know what’s what works, right? If you pull together user story, as a social media marketer, I need to know how Twitter or LinkedIn or Facebook, whatever ranks and scores content so that my content gets seen by more people or my contact is engaged with by more people. I think that would probably be the place to start now. Katie, and John, I’ll ask you both this. Besides pretend you didn’t work for Trust Insights, and you didn’t have a nerd, just sitting around all day, willing to answer your questions. If you didn’t have that, maybe think back to our agency days, we worked at agencies, how would you go about getting this information of what works?
John Wall 3:56
It’s always been the past for me is you do go to all those articles, and you grab all the urban legends, but then you test everything, because you don’t believe the urban legends. You know, like, they’re a good point to start. But there’s no proof that any of them are real. And yeah, I’d say it’s 5050. Or worse, where you try some of these things. And you’re like, yeah, no, actually, that makes things worse off.
Katie Robbert 4:16
Yeah, I my instinct would be to do some sort of a web search on, you know, best practices for, you know, getting high engagement on LinkedIn or Twitter, and see if there are, you know, noted experts on the subject, who’ve written about it and see what they have to say, and just using that as a starting point, like, you know, if, if I didn’t work with you, but I saw, like, oh, well, what does Chris Penn say about best practices? For posting on LinkedIn? What is he saying works? And so looking at those, you know, trusted sources to see if they’ve already done the work for me
Christopher Penn 5:00
I’m curious that neither of you mentioned going to the companies themselves. And the specific place that we would think about would be the engineering blogs. So both Twitter and LinkedIn, have engineering blogs where the engineers say like, yeah, here’s the thing and how it works. Why wouldn’t use it start there?
Katie Robbert 5:20
Probably the same reason we’re building this live stream episode is that neither myself nor John, are engineers, we’re more of the layman. I mean, sorry to speak for you, John. But I can say to myself, I’m not someone who’s going to seek out an engineering blog, because a lot of times, at least in my experience, they’re not written for people who don’t understand how engineering development, and it works. So they’re not going to, they don’t necessarily go out of their way to summarize a bullet point and be like, so all of this said, here are the three things you need to do.
John Wall 5:57
Okay, no, it’s funny, too, that they do terrible SEO, like, I’ve never had some random question like this, and looked it up. And you always find 16 marketing blogs that have, you know, top 10 of whatever and you don’t find, you never see one of these engineering articles, like I’ve never stumbled stumbled upon an engineering article aside from one that you’ve shared.
Christopher Penn 6:18
And that’s a really important point because to what Katie was saying, it’s not meant for the general public, right? This is this is meant to instruct other people as to how they think about solving the problems. And one of the things that we’re going to, we’re going to notice as we kind of take a tour through these things, is that it’s almost detective work. You will note, particularly with LinkedIn, LinkedIn tells you how everything works, but across a bunch of different papers. So some of the pieces that you need to know are in one paper, some of the pieces you need to know in another paper, and you have to have the wherewithal to know that those pieces are related together. But here’s, here’s an additional question, why would we not take one of these things? So this is a beer for folks that are just listening to this. This is a goose IPA. Now, why would not take a few of these things, plus some friends or colleagues who have engineering experience and say, Hey, can we sit down for 30 minutes? And go through the these engineering blogs? I know, I don’t know, what these things are about, but you might, can you and here’s what I’m looking to find? Can you help me distill that? Why would that not be an approach that, say it agency president or agency CEO would take?
Katie Robbert 7:29
I think it really depends on the relationship. And so you know, it could be seen as, okay, tell me all of the secrets that you know, so that I can do better than you. And so before you can just, you know, sit down with someone and start, you know, picking their brain about what they know, you to have that relationship stablish not everyone’s willing to give away. You know, how they know what they know, for lack of a better term, it could be an option, you know, but again, it sort of goes back to, you know, what, the way that I was approaching it of, you know, find out who, who in your network that you trust, what have they, you know, What work have they done already? What have they written about it, and so it may sort of stand in for that, you know, sitting down with someone. But I’m also adding to my list of marketing, urban legends, you know, bringing beers to it, because I think that’s sort of an antiquated way of looking at it.
John Wall 8:28
True. beers are just from the time before, it would be great if we could get back to that. But it’s been a long time.
Christopher Penn 8:36
I think it’ll be better beer, because this one looks like it says like, it’s entirely all hops. So the starting place, then is to look at these blogs, and you’re looking for specific topics. And that’s, I guess, would be the first place to start is let’s pull up in here, just the the homepage of the Twitter engineering blog. And you’ll note as we start to scroll through it, there are different topics, there are things that are around security around infrastructure, data quality and stuff. And what we’re looking for specifically, if we want to know what works on Twitter, is to be looking for articles that talk about the Twitter network itself. Now, you’ll notice there’s some words and phrases that keep popping up over and over again, right, so you’ll see some natural language stuff, you’ll see some stuff about privacy. But you’ll keep noticing that these terms graph neural networks or graph networks just show up over and over and over again. And that’s your first indicator. That that is a core technology that is part of Twitter is something that is really important.
Katie Robbert 9:40
But I guess that sort of goes back to the these aren’t meant for non technical people because I would first you know, again, assuming I don’t work with Trust Insights, I would first need to know what a graph neural network is that it’s even an important thing. I would need to know what geometric deep learning meant. I would need to know what Catherine is you have to learn a different language. So it goes back to do I need to have a dictionary with me to understand what I’m even looking for. So it sounds like your first pro tip is try to find the terms that are repeated multiple times to see what kind of importance they carry for understanding the algorithm.
Christopher Penn 10:22
Exactly. And then the second thing to do is to start popping open some of these articles and you know, without digging too deeply in them just looking to see what other articles They all link to. One of them is this one, this one is linked to a lot within Twitter’s own blog. This is sort of a cornerstone piece of content that a lot of other pieces of content, talk about and a reference. And so even again, if you don’t necessarily know the topology of all the language and stuff, you can say, this articles got a lot of links. It’s kind of like doing a primitive human version of SEO, like, where’s all the links going? Oh, they’re all going to this article. And that’s probably I would say, your second tip to know. Yeah, this is there are some articles that are more important than others. There’s some that are Cornerstone pieces of knowledge. A good indicator that an article is really, really important is that it will also typically have a link to a published academic paper. Right. So there’s a published academic paper in this case that has been approved and peer reviewed about this core technology at Twitter.
You look puzzle gating?
Katie Robbert 11:37
It’s not it’s not so much that I’m puzzled. But it’s the as someone who has written an academic paper, again, it’s not for everyone, it’s for a very specific audience. And so is it that it’s linked to an academic paper at all? Or do you also need to go through and start to understand what’s in the academic paper? Because my sense based on my own experience, is that that paper is going to be even more dense and cumbersome than a blog post.
Christopher Penn 12:10
The first place I would start is read the Yeah, to answer question, yes, it should just have a paper period, right, like a paper exists that was peer reviewed, because that’s the as you know, that’s a lot of work. Oh, goodness, yes. You don’t do it lightly. It’s not something you burn a lot of hours unless you think it’s important. That said, the blog post probably is more accessible than the paper itself. And in that, in this case, that is 100%. True, I read the paper and the paper is not a lot of fun. The blog post is slightly more fun, but it’s still about as exciting as watching grass grow. However, this paper talks about the way that they deploy a specific kind of deep learning on Twitter to decide all of its recommendations. So when you’re on Twitter, for example, you will see recommendations of every kind, you will see, let’s pull up twitter.com Here. So on the Trust Insights, Twitter, you will see, hey, here’s some people we suggest you follow. If you go to the home timeline, here’s some tweets that we recommend that you take a look at. If you go to the Explore tab, here’s some more things that you know, here’s some trends that we think you should be looking at. All of these are driven. These are all recommendation engines at work. When once you know that, and you pair it with this paper on this specific type of neural network and how it makes its recommendations, suddenly, you can now go, Oh, I know what’s fueling each of these things. So if I want Trust Insights to be shown up in the who to follow section here, I need to understand this technology. If I want to understand why Ali Frankies tweet is showing up here. It’s I need to understand that that recommendation engine.
Katie Robbert 13:57
I would say this with the caveat. So if you go back to that article that you had up, Chris, I just want to note that the date on that, at least from my screen looks like January 2021. That’s right. And so the recency of an article, and or especially an academic paper is also a factor. So you know, it, I think it depends on the industry that you’re in, in terms of how quickly academic papers can be turned around in technology. They’re faster than clinical research, for obvious reasons. You know, but keep in mind, you know, an academic paper that’s two years old, could be completely irrelevant because of the amount of time it takes to put it together and publish it and peer review it, and then get it out there. So I would say also proceed with caution, especially with knowing how quickly a lot of this technology changes. You know, if you’re looking at this article, for example, it’s two years old now. And so I would say it’s going to give you good understanding of what it was, but may very well have changed since then.
Christopher Penn 15:06
That’s right. And for some of the smaller articles, I think that’s a good caution to take into account. When you have Cornerstone technologies like something that is a major part of someone’s infrastructure, that will probably change a little bit less. If newer articles were not also linking to this one, I would be concerned that this one’s out of date. But some of the posts as of a couple of months ago, are still linking to this one as well. So I think there’s there’s still relevance. So let’s, let’s dig into this a bit. And answer the question, what does this tell us about how Twitter works? Without going into the mathematics, but you’re welcome to go through because they it actually is fairly interesting. What a graph neural network is a type of machine learning that looks at the relationships between entities, people, were tweets, or people tweeting, for example, and tries to understand and make predictions about what an entity is likely to do next. So for example, it if I follow a Katie, and Katie follows me and John follows me. You’re logical next recommendation should be John should follow Katie to write that’s a very, very primitive example of using this type of thinking, if John follows me, but does not follow Katie, but I engage with Katie’s tweets, right, the algorithm should say, from a predictive aspect, probably show Katy sweet to John to see if John will engage with it, and then learn from that, like, yes, showing Katie’s dog pictures, John engages with that. But when Katie shows, you know, cat pictures, John does not engage with them. So you have some of these learnings that that this graph network is applying not to everybody, but to between individual people. And that’s a critical part of understanding how social networks in general work, but especially Twitter, is the recommendations in the algorithm are not global, they are unique, at some level to the individual person. And then more broadly to these clusters within networks. So there is sort of a Trust Insights world on Twitter of people, we follow people who follow us. And within that, because of the way graph neural networks work, there’s a higher probability that people are going to see stuff that’s topically related, will recommend us more, but somebody who is maybe just as many degrees away from us, but not in that world will not see us as often. So that’s an important part of understanding how these graph neural networks work.
Katie Robbert 17:40
It sounds like and I’m glad you use the word degrees, it sounds very much like that goofy six degrees of set of Kevin Bacon game where the closer you are to the person who’s right in the middle of the circle, the more likely you are to be able to interact. And then the farther as you get to be like fifth degree, six degrees seventh degree, once you’re 760, at least in that game, like you don’t even count anymore. So you exact closer by degrees. Exactly. Thinking about it in that construct, I can wrap my head around it.
Christopher Penn 18:12
Exactly. Now, here’s what’s novel about Twitter’s implementation is they have gone from just a regular graphic network, which is Six Degrees of Kevin Bacon. And it started with embedded time. So it’s not just John engaged with me, or Katie engaged with John or John follows Katie. It’s also how long ago that happened. So as time passes between messages, between follows retweets, likes, etc. The more time that passes, the more that data gets embedded into each node. And then the recommendations change. So if John and Katie are constantly talking to each other in rapid frequency, if I follow them, I’m probably going to see the other person higher probability because of the time embeddings in this temporal graph network, as opposed to if they had engaged with each other a year ago and hadn’t engaged since that that recommendation engine because the time factor is so much further away. I will go okay, well, the engine will say this is not as relevant to you. So the key takeaway from understanding a temporal graph now because yes, you have to follow and be followed and engage and be engaged with, but it has to be recent. Twitter is an engine of recency. And so if people do not understand like, oh, well, we’re just gonna tweet once a day. Great, then you are behind that account, that annoying account that tweets 30 times a day, because the temporal embeddings that go into these accounts, so you are still you can’t tweet garbage, right, because the engagement still matter. But you need to be engaging and frequent. So that’s what this paper tells us.
Katie Robbert 19:49
Got it? Well, and it’s interesting because I feel like people would sort of selective listen and hear that what you’re saying is, I just need to tweet more, and I’ll be fine. But it’s deeper than that there’s more layers to it, there needs to be a certain level of engagement with those frequent posts that you’re putting up, you know, it needs to be the right kind of engagement. And the engagement needs to be recent or routes, the connection starts to really kind of degrade. And then you get put down to the bottom of the list again, in terms of the algorithm.
Christopher Penn 20:23
Exactly, because part of what this paper also talks about is that the embeddings occur, the time embeddings co occur with other embeddings. And an embedding is just a fancy word for a variable likes, comments, retweets, etc. reshares replies and said, all those things are are statistical things that are embedded into the relationship that any two nodes, any two people on Twitter have with each other. So yeah, if you were just talking to the air, like, you know, amusing to the, to the ether, about whatever topic you is on your mind these days, and nobody engages with it. Yes, the temporal embeddings will be fresh, but there’s no engagement. So there’s nothing that will tell the network, oh, this person stuff is relevant to these people, let’s show this person to these people, because there’s a relationship there. So that’s, that’s a critical part, too, is you can’t just be fast, you have to be fast and good.
Katie Robbert 21:23
So let me ask you this question, though. And maybe you’ll get into this as part of, you know, the takeaways of this paper. So it sounds like this so far, if I were a marketing manager trying to understand okay, how do I, you know, make my Twitter account better? It sounds like I need to post more and get more engagement? Did I really need to read a paper? Because that sounds like a general best practice. Anyway. So what is the paper telling me about the specifics versus what my gut is telling me to do?
Christopher Penn 21:54
In this case? It’s, I think it lends people don’t think about the timeliness as much right of the importance of frequency, and recency. Yes, tweet, tweet valuable content is a pretty abrupt, well known trope these days. But we did not know the time was explicitly a not just part of the equation, but a fundamental underpinning of the model. And that’s that, to me is what’s different about this is saying, Okay, time is as important as you tweet quality is not just in the mixer like it is as it is a peer of quality. So it is it is theoretically possible that as long as you are talking about the relevant topics, because that’s discussed in another paper, that you could be seen more by, and really, people were interested in that topic just on the timeliness alone. But the engagements obviously helped boost that substantially more.
Katie Robbert 22:55
Got it. So, and I have to say this URL, so I’ll just keep dwelling on in my head. So for example, so I am one degree from Kevin Bacon. However, my connection point to Kevin Bacon is a relationship that stems from high school, which I haven’t been in for a while, a while. And so therefore, just because I can claim that I have one degree to Kevin Bacon, it doesn’t really matter, all that much in this context of this algorithm, because my connection point is over 20 years old, therefore the time has degraded. And there’s going to be other people who have the same, you know, degrees to that single source that are newer, that are more recent, that are more active.
Christopher Penn 23:51
Exactly. Now, the same person you had lunch with last week, your ability to say, Hey, could you pass a message to Kevin? It’d be much higher
Katie Robbert 23:59
at it, which, okay, so that makes sense. So I’m trying to I’m trying to put it into a context that my sometimes non technical brain can try to like, wrap around it. John, I don’t want to monopolize, what kind of questions or reactions do you have to all of this?
John Wall 24:16
Well, the thing is that this just really highlights the stuff that we see spread, right, all of this, oh, hey, tell us about a movie with a great soundtrack, or what’s your last favorite three things on your photo roll, like all that kind of stuff. It’s not that compelling. But people are, you know, do interact with it, and they do it immediately. And that’s the kind of stuff that spreads like wildfire. It’s not that it’s, you know, any good but it’s the kind of stuff that yeah, oh, yeah, I’d love to tell you about the last TV show I saw
Christopher Penn 24:46
or bingo. This is why misinformation spreads so easily because it’s immediately believable it gets you engage and it’s fast.
Katie Robbert 24:59
Well, that’s depressing. What else you got?
Christopher Penn 25:01
So let’s move on to LinkedIn. But no, I would encourage people again to dig into these papers because that, again, the so what is I understand now why some content does better than others? And why asking people on Twitter, pineapple on pizza, right? immediate engagement, fast engagement, frequent engagement, all those things are things that a temporal graph network would naturally favor.
Katie Robbert 25:33
I was talking with someone yesterday who mentioned that pineapple can go on pizza if it’s paired with a meatball.
Christopher Penn 25:41
And a little sweet and savory.
Katie Robbert 25:43
And I think that’s what it was. And I understand that that’s the whole idea. And I don’t want to go down the rabbit hole of pineapple on pizza, but I can understand why it’s something that’s more universal versus niche. And people can respond to it quickly and engage with it quickly, and then move on.
Christopher Penn 26:01
Exactly. And the more universal it is, the more engagement you’ll get. Which means the nodes stripping an edge strength between two nodes grows. So you have pineapple on pizza, which way does your toilet paper go over under DC or Marvel? You know, do you wear shoes in the house, should you wear socks with sandals, Apple or Android, all of these things are there, they’re chewing gum, right? They’re not nutritious, but they get people doing something. And and those that, by the way, actually also is something that we’ve talked about for private social networks to generate community engagement as well. So same process applies to LinkedIn, LinkedIn has an engineering blog. And they have. What’s interesting about LinkedIn is that LinkedIn has so many different components and pieces. And it’s difficult to understand that all these pieces are related to each other, and that the underlying technology talks to each other in ways that should inform our behavior on LinkedIn. So let’s look at a couple of different things. I’m going to review each of these pieces a relatively quickly, and then we can talk about how they integrate together. The first is the skills graph. So every time LinkedIn processes your data, based on what you’ve put in, it’s going to be it’s building a natural language model of who you are. And one of the things that they’ve done is they’ve built in what’s called an inferred skills graph, they have a structured list of data, so that they understand if you’ve post about artificial neural networks, you probably know about machine learning, you probably know about deep learning, you probably know about statistics. So even if those skills are not on your LinkedIn profile, it understands from a topic perspective, you probably know about those things. And those things are probably important to you. So LinkedIn looks at at a few different things, they look at you as the person and what’s on your profile, they look at you and your content, the language using your content, and they look at your network, and the profiles and the content, you know, talks about, and all of that together gets sort of mashed up into his understanding of what’s likely to appeal to you. When they when you look at the papers that talk about how LinkedIn decides what to show you in his newsfeed. They show you all the different interaction features and embeddings that go into this model to say, Okay, this because you probably know about machine learning. And because you’re connected to these people, you will probably like a post about x. And so it will show you that and then obviously, you takes into engagement measures like if I show you a post about artificial neural networks caving, you don’t engage with it like okay, clearly, even though your network says you might like that you yourself don’t. And so that will that will change those those weights will change. Here’s where it gets interesting. Because of the way LinkedIn is set up, its architecture governs a lot of what gets shown. So for example, you’ll see all this random advice on LinkedIn, like, oh, you should be posting videos, and you should be posting this and that and this gets higher engagement stuff. When you look at the underlying architecture for how they design things like product search, there’s only two categories of content. There’s articles and posts. That’s it. So from a, a content structure, those the only two major chunks of of data. So a post is anything you put in that isn’t an article and an article. Of course, it’s the long form content that you put on LinkedIn, your newsletter, for example, if you take a look at the Trust Insights, free LinkedIn course we have, it’s a TrustInsights.ai AI slash LinkedIn course. One of the things we talk about in the course is if you’re building your personal brand, you must publish a newsletter. Just you got to do it because it’s true because of you. If you look at the way the search Federation works about how their their internal databases work, articles, including newsletters are treated differently than posts so your article will get typically we get more engagement and more spread than a simple Post, however, so go ahead, good, no good.
Katie Robbert 30:04
Well, I was gonna like my like, this is one of those moments where like, my brain is working so hard to try to keep up and have intelligent conversation about this. So I guess there’s two things that I want to sort of comments, slash ask. So the first one being, it sounds like, in some ways, like if I’m not saying it works identically, but it sounds like in some ways, it’s similar to the way that we think about semantic SEO. So related key word. So if you know about artificial intelligence, you likely also know about machine learning. And so again, sort of trying to wrap my brain a boat around it that way. So it’s obvious it doesn’t work the exact same, but it’s a similar idea. The second observations, last question, and this is true for Twitter as well is, you know, we’re talking about so let’s say you share with me an article on artificial intelligence, and I don’t engage with it. It sounds like the algorithm decides, Okay, that’s not the right thing to be showing to Katie, she didn’t engage with it. But it doesn’t take into account the fact that maybe I was just really busy that day and couldn’t get to it. But I want to engage with it. And I just like I just don’t have time to and so like, I feel like that piece that actual human behavior piece is missing from both Twitter and LinkedIn. And I don’t have a solution for that. But it’s more of an observation of like, it’s making assumptions, but it knows nothing about me. Maybe I’m on vacation. And there’s 10 articles that would have shown up in my feed. That would have been amazing for me, but I’m never going to see them because I wasn’t there to actually see them.
Christopher Penn 31:46
It depends it so age, particularly Twitter is a factor less so on LinkedIn, because it’s looking at probabilities, right? So your if you don’t read that article today, but in the past, you’ve read articles, no, six days later, there’s a probability that that will reoccur. I mean, we’ve all been on LinkedIn where a post shows up as like a week old, like where they come from, I have seen posts resurface on LinkedIn that are two months old, like, why did this come now? Oh, because it’s about this thing related to this person that I’m connected to, and so on, so forth. So the, the neural network is experienced enough with my member history, to resurface things, when there’s there’s evidence that I might engage with it. Because LinkedIn has three different measures of engagement under the under the hood. And this goes, this is in one of the multitask papers, there is passive consumption, there’s active engagement. And then there’s upstream engagement, there’s two or three different papers to talk about this passive consumption is, what is the likelihood you’ve got to read an article or post right? And there’s a probability engine to guess that there is active consumption? What is the probability you are going to like, comment or share a piece of content? And then there’s upstream engagement, where, what is the probability of this creator, being engaged by these metrics to create more, one of the things that makes LinkedIn very different from every other social network, and one of the reasons we like it, is that there are incentives built into its algorithm for creators to reward them by getting higher levels of engagement to keep them creating to encourage them to create more. So if you go on LinkedIn, you post once and you know, not much happens, like, okay, whatever. But if you, if that post definitely does well, for the wall, what does that incentivize you to do? Like, oh, I liked that little validation on me, I’ll do that again. And so part of the algorithm is balancing the things that make creators happy so that he can take the new to contribute to the platform. Got it. Let’s dig into the relevance of content because this is this is the meat and potatoes of how LinkedIn decides how LinkedIn understands the content is going to show that it goes through three big passes, there’s a first pass that sort of establishes some basic metrics and says, Okay, let you know this, this, this should be considered not considered the second pass is okay, let’s, let’s rank that stuff. And then the third passes. Okay, let’s, let’s try displaying that stuff. So in that first, if you think about this is this is very similar to Google, right crawl index rank, and then display. So very, very similar. In the in that crawl slash index FeS, there’s four different models that work together as the P 13. And model that’s a translation model basically says, if I post in Spanish, probably show my posts to other people who speak Spanish, right? Probably show it to two people who engage with Spanish language content don’t show my posts to people who engage with Russian language content. The second is the relevance quality models. The relevance quality model, is what we in SEO, you would just call term document relevance. So If I search for or I talk about machine learning, is the piece of content that’s being evaluated about machine learning, right? Yes or No. If it’s not topically relevant, then probably give it a give it a thumbs down. So if I post about influencer marketing and fashion, machine learning posts that it was being evaluated to be shown to me probably is going to get a thumbs down like that I don’t think this member is going to be interested in this piece of content. The third one is the content quality model. And this is very, very specific. LinkedIn says this is query agnostic document quality, ie static rank. Those are magic words, if you understand
Katie Robbert 35:43
that, again, query ignore very
Christopher Penn 35:45
agnostic document quality, ie static rank.
Katie Robbert 35:50
My whenever I talk,
Christopher Penn 35:53
when everyone talks about oh, he high quality content on LinkedIn, everybody else goes, what does that mean? What is quality content? Right? Does it mean pictures, oil Dancing animals? What does that mean? The key word phrase here is static rank. This is essentially a measure of centrality. What does that mean? It’s PageRank. So it means inbound links is what that means. When you have all the content when you have content, how many people link to that content, it doesn’t matter if it’s a dog, a cat or five ways to you know, your, for your founders to, to get lines of credit, whatever it is, if the member and their content was getting inbound links, ie mentions were reshares, from other from other people in the network that goes into the content quality. And all of that is what content quality means it is inbound links. It’s it’s SEO for LinkedIn. And the fourth one, hey, there’s our old friend recency. Right. So time is in this model as well. That’s the first pass. The second stage is the member context. Right? So this is a question of, for this piece of content, what is the context of them of the member their first degree network? If I’ll give you a real world example here, three weeks ago, literally, nobody in my network had any idea of including myself about this, this TV show Warrior Nun, right. And I got involved this whole bunch of people. Now this is topic that I post about. And suddenly the models like, should I even show this because this does not match the context of me in my first degree connections, right. So some of that stuff will do less well. Because Lincoln’s like, I don’t see what this has to do with all your connections. All connections are about marketing analytics, and you’re talking about a TV show, like why. So there’s a potential for that to be downright there, which is why it’s very important for you to be somewhat thoughtful about who you let into your LinkedIn network as first degree connections. If you just let in every Tom Dick and Harry you know who’s got their their multilevel marketing Ponzi scheme that if they’re Hawking, or crypto or whatever, scheme talking this week, that changes your context, which means it’s going to change what you see on LinkedIn. So we want to be we want to be a little bit thoughtful. And then we have affinity and other Nearline features. So again, these are the things like Who do I interact with the most? If I like something and I interact with say our friend Justin lovey a lot when a post comes up for consideration and Justin speed is going to look at how often I’ve engaged with with that post of that creator and say, Hmm, Chris has engaged in this lot Chris’s friends Justin Neil show this posted to Justin and see if he likes it too, because of that, that affinity and then it goes into you know, how you do the actual display of the content itself. So this is linked ins, the entirety of the of the algorithm of the composite go into this system to decide what to show. Once you understand those different components. You go, Okay, I should write content and intelligent language I should have my content should be relevant to various terms and things I want to be known for. My content should be linked to is shared by other people. So when even when you read a LinkedIn post, like, hey, please share this with your network or please what everyone’s like, leave a comment like, yeah, leave a comment, but get people to share your stuff. This says beyond LinkedIn right now, don’t just stop in like once a day for like two seconds, actually spend some time do some activities, member contexts, engage with other people’s content. Now, to your point, Katie, from previously, this is not new information, right? He’ll be on LinkedIn and be relevant. It’s not new information. But the new information for us is this PageRank stuff, which is not something that’s commonly talked about on LinkedIn. And the fact that these features occur in two different passes. And so LinkedIn algorithm looks at the quality of the content first and then your network. So that’s a really important thing. So it’s not just making stuff for the people that you like, but it’s making stuff that is high quality first that people will link to, and then sharing it with a well chosen network.
Katie Robbert 40:17
What strikes me in both of these examples, both LinkedIn and Twitter is that there’s not there’s talk about your network, but there’s not a lot of talk about building your network relationship building. Because I feel like this the way that I mean, and this is just my might be my interpretation, but the way it’s being described is that it’s, you know, for just like growing, you know, your awareness, but, you know, it doesn’t factor in that you need to take time to build that network to foster those relationships to get that trust, even if it’s just over social network to say, you know, what, I trust that when John posts something, it’s going to be good quality, so I feel comfortable sharing it, you know, and so it’s, it doesn’t, again, sort of the human behavior side of it doesn’t take into account what I consider to be good and trustworthy. I feel like that’s very subjective. And that’s going to be harder for these algorithms to pin down. You know, I would imagine that like, you know, John would feel comfortable posting anything that comes from Trust Insights, because he knows us, he knows the kind of content that we put out, he’s part of the process of that content creation. And so that inherent trust is not something you can build into an algorithm.
Christopher Penn 41:40
It’s not something you can build into an algorithm. Explicitly, but it finds its way in there, implicitly, I was trying to find a piece that they did not too long ago about the people you may know, such and because that algorithm itself has a lot of the human side here is optimizing people, you know, for equity and content creation. So they talk about the model behind the people you may know and how it recommends, you know, who you’re going to see and who they are one of the we take, I think we talked about this and want to say, last year, sometime, one of the things that is in the people, you may know algorithm is a propensity model for dealing with people who basically says some people don’t like getting lots of random connections. So the more of these invites that are sort of stacking up in your inbox, the less your you are going to show up. And people you may know, because Lincoln’s like, clearly, this person’s not engaging with it. So let’s show you less in the recommendations for people to grow your network. If you keep this clean, even just hit ignore, like over and over again, you’re keeping it clean, and therefore you’re going to be shown more. So just a very simple tactical thing from this part of the LinkedIn algorithm. And again, this all goes back to those language models to say, Okay, who are we going to show you, people who match your skills, people who match your content, people who have affinities like schools you went to or companies you worked at in the past, that’s going to be to be shown and ranked higher in these systems. LinkedIn in particular is very good about trying to identify people will be good candidates for your network, Twitter less so but because Twitter focuses so heavily on its graph and temporal graph network, that it it doesn’t pick as people I think, as well, in terms of the people that you should follow as LinkedIn does. LinkedIn has more data to draw from, it has richer data to draw from, it has much more personal information to draw from just off the bat because of your profile and things. And so I think in the because of that, it does a better job. So is it explicit in the product? Not necessarily. But Are there parts of the product that are designed to help you broaden your network and grow? Yes, and if you know how those pieces work, they work they work pretty well.
John Wall 43:57
It’s interesting that there’s kind of this push pull then of, you want to have as many followers as you possibly can to increase your odds of stuff being shared or forwarded engaged with the rise up. But at the same time, you don’t want to be following all the garbage accounts, because that would hit you in that second stage where it would presume that none of these other people want to see this stuff. So it actually stopped showing your stuff. So you kind of you have to that has to be actively managed to be optimized.
Christopher Penn 44:25
Exactly. And if you think about it makes total sense, right? If you’re just being followed by a bunch of bots, you know, they’re not going to reshare, your content, things like that. And LinkedIn wants the highest quality content shown to real people. And so if your entire LinkedIn network is junk, then yeah, your stuffs not going to go anywhere. It’s not going to get accelerated, it’s not going to get seen. If, on the other hand, you’re adhering to the the eight or nine different pieces of the algorithm that we know is part of the architecture, your chances of being seen go up dramatically.
Katie Robbert 45:01
Yeah, I’m I’m one of those people who notoriously has a bunch of outstanding not looked at invites connections in my LinkedIn network because I have this rational slash irrational fear that the second I connect with someone, they’re immediately going to pitch me something. So I mostly just ignored unless I know someone, and I’m just looking I had 24 outstanding, which I feel like isn’t that bad, but I still had 24 outstanding that I needed to do something with so I can go through and clean those up. But yeah, it’s, yeah.
Christopher Penn 45:36
Here’s the thing about LinkedIn. That is, again, something I love about LinkedIn that I wish other social networks are do. LinkedIn has. And there’s a whole paper on this. It’s called multi objective optimization. And they talk about the different factors they use for optimizing, you know, the member experience, one of those factors that no other network does, is complaints, right. So if you complain, if you do have a complaint style action of some kind, that waits very heavily in the network. So when I accept an invitation from somebody, they immediately slapped me with a sales pitch, I hit, I hit block, and I hit report, right, and I report them for spamming me. What does that do to their reputation, they now have a big ol honkin red flag on them for a while that says this person has been complained about. And this is a problem. And so they get down ranked pretty dramatically, pretty quickly. So it’s a way for LinkedIn to say we want we because what LinkedIn desperately needs is they need to hold on to their users, because that is the product that they sell in their HR software is user so they cannot afford to lose people, you know, because someone’s spamming this political cause or this thing or handing out he’s doing inappropriate behavior. And direct messages, the moment you hit complaint, you’re basically sticking a big ol red flag on that person’s back to say, like, yeah, this person’s can be problematic. And if enough people do that, that person goes into the end of the algorithm, you go to the end of the line, you never get seen again. Whereas compare that to Twitter, particularly after the management change, or all the big see, the trust and safety team has gone complaining does nothing, you don’t even find out this accident taking, it becomes a much more of a cesspool, very, very quickly. So with LinkedIn, that’s, that’s a major factor in this favor. That means that you can accept those connections more readily. And it’s also a warning for anyone who’s listening that if you are that person who immediately sends a sales pitch, you’re gonna get the crap smack data,
Katie Robbert 47:27
right? Oh, well, we have been. Or Chris, you have been explaining to myself and John for well over 45 minutes, how Twitter and LinkedIn algorithms work. I can say that My head hurts. But it has been incredibly informative. Because it feels I feel like it reinforces the things that we make assumptions about. But it gives us that more granular point of view of like how to approach and the big thing that I’m taking away from this is that you cannot just have one blanket, social strategy across platforms, you have to treat each individual platform as its own thing. And so for Twitter, its frequency and recency. LinkedIn, same is frequency and Ricans recency. But also the things that your profile says you’re interested in the way in which you engage the amount of people that you have engaged with, you know, in your network and brought into your network. So there’s there’s nuance it’s similar, but different.
Christopher Penn 48:36
That’s exactly right. And the other thing I would say is that we talked a lot about how to think about deconstructing these algorithms. And one thing I want to go back to the John setup at the very beginning of the show is, if you have a question about a feature of some kind, how many hashtags? Should you put a tweet or whatever? The data from these services is something you can export, and you can run that statistical analysis yourself. So you can say, yes, it turns out for our audience, our little corner of Twitter, maybe the number of hashtags does matter. I was answering this question to some PR folks. And they’re like, Well, you know, how many hashtags is matter. I will show that here’s a statistical map of the factors that get Twitter engagement. Hashtags is a zero has no relevance, none whatsoever. This entire little ecosystem and like, so stop obsessing about it, it doesn’t matter. It is never mattered, but it definitely does not matter now. And I would encourage that, and that’s another way to pick apart these algorithms is to do your own statistical testing so that you can say, Yeah, this, these factors do not matter. Let’s not worry about them. Let’s focus on the tried and true which is create good content and build a strong network of people to share it.
Katie Robbert 49:50
And I think most importantly, we slash Chris can be the one to decipher all of these academic papers and help you put To gather that strategy for this specific algorithm. So if you hit us up TrustInsights.ai slash contact, you will first get to talk to John. John will talk to you all about his Twitter and LinkedIn experiences. And then Chris can absolutely help you decipher what works specifically for you. Because I think that’s the other piece of it is these papers are talking about the general how they work, as you have your own specific network, your own specific audience. The way in which you know, Chris, you approach posting on Twitter is going to be different from how I approach it, it’s going to be different from how John approaches it, because we’re three different people with three different agendas, three different styles, content, etc, all those things. So it’s still going to have to be personalized.
Christopher Penn 50:49
Exactly right. All right, folks, that’s gonna do it for this week. So we will see you all next time. Take care. Thanks for watching today. Be sure to subscribe to our show wherever you’re watching it. For more resources. And to learn more, check out the Trust Insights podcast at trust insights.ai/t AI podcast, and a weekly email newsletter at trust insights.ai/newsletter Got questions about what you saw on today’s episode? Join our free analytics for marketers slack group at trust insights.ai/analytics for marketers, see you next time.
Transcribed by https://otter.ai
Need help with your marketing data and analytics?
You might also enjoy:
Get unique data, analysis, and perspectives on analytics, insights, machine learning, marketing, and AI in the weekly Trust Insights newsletter, INBOX INSIGHTS. Subscribe now for free; new issues every Wednesday!
Want to learn more about data, analytics, and insights? Subscribe to In-Ear Insights, the Trust Insights podcast, with new 10-minute or less episodes every week.