So What? Understanding the LinkedIn Algorithm for Engagement

All right, welcome back, folks. It is Thursday. Uh, this is so what the marketing analytics insights live show. Uh Katie is on vacation this week, so it's just John and I. Um this week, we're gonna dig into uh LinkedIn.

Uh specifically one of the things that's really interesting about LinkedIn is it's you know it's it's definitely the marker's uh choice of of social network is a lot of business folks, particularly for B2B. Um, but one of the things I think is really interesting about it is just how much LinkedIn tells us about how the network works. So before we dig in, John, uh in terms of like your usage of LinkedIn for marketing and things, what have you found uh what do you found that works in the especially recently? Yeah, well, and disclaimer first, too, I have to say they're a sponsor of the Marketing Over Coffee podcast and have been for many, many years. So, you know, I am biased on that front.

But I I mean, and we've talked about this on the show all the time. The fact that it's very different from the other social networks for a number of reasons. I mean, one is that it is business focused, right? None of the other ones are. Because that's what the you know the insane traffic comes from.

But LinkedIn is all about business. And it's amazing how it's completely replaced the resume, really. You know, I in fact, I I don't think I've seen a paper resume floating about in like 10 years. I know I haven't bothered with one. And um, so there's that.

Um, and then there's a whole nother layer of them being acquired by Microsoft and the Linda acquisition, them to be able to do training. There's some crazy stuff on the horizon you can think about because that they know how many companies have openings, how many people are looking, where they all are on the globe. I mean, there's there's just an insane amount of data there. So yeah, I've you know, I'm a huge fan. I love posting stuff over there, and you've been doing some cool stuff too about cleaning up your feed.

I imagine we'll talk about that, which is great. But yeah, I'm uh you know, all in as a tool. Now, I guess and not to make it sound like it's all rosy and perfect. The one thing is you do pay to be there. You know, it's um ever I can't think of an exception, but every time we've done ad campaigns, people have said, yeah, the the leads are better and more qualified, but it's more expensive to do business over there.

Yep, definitely. Well, today we're going to take a look at um some things that we've found digging around in the academic papers and engineering uh content that LinkedIn publishes, because that to me is where you'll find really cool stuff about how the network works um to a greater degree than I think any other social network, period. I mean, Facebook often has its like experts talking about like uh operations engineering, whereas LinkedIn has its folks out there saying, Oh, yeah, here's how this exact feature works. If and if anybody cares, and it's always the machine learning folks. So all the marketers are sitting here going, uh what's the other page?

Math, math. Um so let's dig in because we have four things that we want to talk about today. The first is uh people you may know. So we'll go ahead and bring up this uh blog post. This is again from the engineering team.

Engineering has been uh really big on saying, like, hey, we want to to make sure that everyone's getting a fair shake. And they talk a lot about this concentration problem where you know the the sort of the rich get richer, right? So you know the the popular people get more and more popular. How do you deal with that? How do you uh how do they balance for it?

And they talk about some of the algorithms, but one thing that stood out really interestingly here is they have a decay metric for how often you're shown in LinkedIn. And I didn't know this. The decay metric is how many invitations are sitting in your inbox that you haven't answered to. So you need to get that list of like, you know, respond to these people. And I don't know about you.

I'm kind of lazy. I'm like, I'll go in like once a week and just like clean it out or whatever, like oh yet another business development manager who's like, I want to connect with you. Um and yeah, the decay metric says the more you've got piling up, the less we show ya. That is so you know that's one of those things that that's a discrete learning. It's just like once that's explained to you, like, oh my god, of course it does that, because they wouldn't want to be showing you this stuff for some guy who has a hundred and fifty invites buried in his inbox that he doesn't look at.

You would send it to the people. Okay, so yeah, to do clean inbox. That's uh as soon as we get off this call. Exactly. So uh they do talk about other things uh uh in the what's called the LinkedIn fairness toolkit or lift that looks at making sure that uh for example, you're shown uh diversity of people, like so people obviously you know and things like that, but then out of network people, there's some some randomization and stuff shown in there to make sure that no one gender or ethnicity or even you know a geography is being overshone uh to a person.

So I thought that was that was really cool. But yeah, the big takeaway was yeah, there's a decay metric at work, and it's under your control. That's well, and it's very interesting to see that too, because now I remember years ago, you know, you and I were both been in the same boat. We worked at some of these places where there's a lot of young 20-something women, and I've seen this more than once where you go to a company and you're like looking at the exec, and it's like people who looked at this exec looked at this exact, and it's like a stack of supermodels. Like there's no business relation between them and somebody just some Cretin trolling around.

So it's good to see that that's being the biases being spun out of that. Exactly. So that's number one. Uh so your homework, uh, if you do nothing else from this show, uh, make sure that you go and clean out that you know, outstanding invitations in your LinkedIn network that will that will help uh things out. The second thing that I thought was really interesting is how LinkedIn uses language models.

So one of the things that they do is they try to analyze uh the text that is used in thing uh to just make decisions about what you get shown. Um and so they have this really, really in-depth paper on a piece of software called D text, which is using um the one of the most advanced language models out there, BERT, the bidirectional encoders for representation bidirectional encoding representation for transformers. And without getting into uh the bloody guts of this thing, one thing that's interesting is that they don't use just one version of their model, they have split it out and trained it so that it there's four separate language models, one for searches, one for member profiles, one for job posts, and one for the help center. And each language model behaves differently. So something that you and they point this out in their paper, something that you type in um, say for uh a search query will be treated differently language wise than if you're you're in the job hunting section because they don't, you know, if you're looking for you know uh job openings at uh Microsoft in regular search, that's different than you have a different intent there in terms of what you're looking for than you're in the job section.

And one of the things they talk about in the section is that they pre-train this model on these different domain areas with what are uh it's something different than most implementations do. They train in multiple source fields. What that means is what's in your profile, what comments you leave, what uh interaction, what you post on LinkedIn, but also what happens in your first degree network. So what the comments that people leave there, all of this is used and essentially boiled down in each of these four domain areas to try and give uh members a better experience and sort of rank and rate what gets shown. Well, what's interesting to me about this is if you are on LinkedIn all day and you're talking, I don't know, let's take something you're uh you're talking about some politics some politician, you're just ranting on and on about this person.

This model will start to interpret what you're doing and say, okay, you should be associated with this topic, and we're gonna show you more for you know, in in certain types of search and stuff for that thing. So if you're on LinkedIn talking about stuff that is not related to what you want to be found for, you're kind of you're hurting yourself. Right. So you're saying that the way you behave as far as your posts will actually ultimately determine the audience you've got access to? Yeah, well, determine what your uh to some degree what your audience sees, right?

So the kinds of things that they see. Um there and we'll talk about some of the other performance metrics uh about the LinkedIn feed too, but uh the language that it's interpreting is coming from multiple different places that you are providing, and you're it you essentially have a role in training LinkedIn uh for what you are about. So if there's something that you've been talking a lot about on LinkedIn that you don't want to be associated with, you might want to stop talking about that. Right, save that for Facebook. Exactly.

Oh gosh, yeah. So that was one of the the big takeaways from this. That and the fact that they had there isn't a single LinkedIn language model. There's there's these four different domains. And this goes back to um an episode of This Weekend Machine Learning and AI, the podcast, where one of LinkedIn's engineers was on saying, uh we look at a number of different success metrics.

Like Facebook is very monolithic. Like we just want you paying attention to the website no matter what, right? Uh we'll we'll show you fake news, we'll show you uh aliens, we'll show you, you know, whatever it takes to keep you on the site. And Facebook's algorithm uh in many ways has has learned this and will show us stuff that makes us angry and upset all the time because we keep tomb scrolling. Um one of LinkedIn's uh balancing features is complaints.

They look at to see how often a member complains and what and it rebalances your fee feed. So we've talked about this in the past. If you want to tune up your LinkedIn experience, not only hit the like button on stuff you want to see more of, but then choose that little three dots in the corner of a post and say, I don't want to see this, and say I'm not interested in this topic or this person or whatever, do that you know, three times a day for uh you know 60 seconds uh a time, and within a week, LinkedIn is gonna be totally different for you. That's really interesting. And now, as far as running these separate models, do you see this just as the enterprise plan going forward that there's not going to be a monolithic single train model, but you're going to want to train every that's kind of crazy.

It's almost like employees, like you're gonna want specific ones for specific things. Exactly right, exactly right. Because yeah, you want a model that is well tuned for specific domain tasks, you know, just like you you're you wouldn't, I hope put your HR person in charge of finance, right? At least not in a big company. That that might not go so well.

And now that you say that now I see how that would work too. What you would have to have is you would have yet another model on top of that. That model's whole purpose would be hey, which one does this go to? Exactly. Uh there's actually you know, term terminology for like hypervisor networks that can essentially hand things out in an assignment.

So we're in in some ways we're we're taking like jobs, single-task jobs, and having these models be very specialized in that because you don't want the language that people use when they're searching for a job to be determining what they see in their LinkedIn feed necessarily, right? They you want those to be discrete models. And so what they've they've shown here is that they've actually done that. And there's even some more detail uh further on down about how they fine-tune each of these language models based on this the performance metrics that you know uh they're getting from users. So it's learning from what w other experiences uh people have had on LinkedIn, like positive stuff.

Like if there's this post about uh you know Wiley Coyote on LinkedIn that that nobody likes, uh over time, if that happens enough, it's actually gonna it's actually get discounted further and further in language while until you can put up a post about Wiley Kodio and your LinkedIn feed in and no one will ever see it. Right. You get shut down. Exactly. So that's number two is be really careful about what you post on LinkedIn and what you comment and things like that.

Not that you should censor yourself, but make sure that you're talking about the things you want to be known for. Okay. Number three uh is looking at content publication. This I thought was fascinating. LinkedIn has incorporated the idea of both upstream and downstream metrics.

So when you post something on LinkedIn, uh there are downstream metrics, like who uh who shares it, you know, how far does this content get shared in your network downstream? And they're and obviously they're looking to see the more that other people engage with it and reshare it, the more it engages the rest of the network. And that's a positive thing. LinkedIn's like, cool, we like that. But a couple years ago, they also added something uh creator side metrics, they're calling them upstream metrics.

Where if Richard, the the example they say is if Richard Branson posts something on LinkedIn and you hit like on it, the impact to him is pretty much zero, right? He's like, oh good, I went from 8,600 likes to 8,601 likes. You know, there's no impact. But if you or I, as average people post on LinkedIn and you know, so you know, one person likes it, like, oh cool, like we actually see it, we like it, we engage it. And one of the things they realized in their machine learning training is if you keep the creator engaged, they stick around longer, they are more engaged on the network, and they create more stuff.

So they now have these these bi-directional downstream metrics and upstream metrics. So things like likes and comments and stuff are now factored to upstream metrics. The likelihood that a post to the creator publishes that is likely to give them feedback to keep them sticking around is part of the way that LinkedIn analyzes content and decides what's out what you know what it should and should not show. I thought it was really interesting because what it does is it it tries to essentially solve for that. Um, you know, you know, when Bill Gates publishes, everybody reads it, or when Richard Branson or Elon Musk and try and help out getting smaller creators to get their stuff seen.

Um they went through and said, Yeah, this thing works really well, and we're getting creators engaged. So one of the um takeaways from this is that if you if you publish something on LinkedIn and somebody engages it, likes it or comment, you go back in and keep the conversation going, right? Because it's going to basically say that that post is valuable enough that it got you, the creator, engaged. And because it's a balance of upstream and downstream metrics, you want to make sure that you're you're doing your part on the upstream side of like, oh yeah. So if you see a post you like on LinkedIn, you know, anytime somebody comments, you know, like or whatever, you know, engage with it, reply back if it's appropriate to you, even just to say thank you, so that you keep boosting that post's um upstream metrics and saying, Yeah, this is keeping me around as a creator algorithm.

Please keep showing it to more people. Yeah, that's really interesting too to think about how like you know, it just creates that feedback loop and going over and over again. Um, and how those big names are actually getting filtered by the general public. You know, it'd be people are just gonna follow them because they're a huge name, but if it gets reshared at the lower levels, that makes a lot of sense um to actually test the content and see what it's really like. Um so really like you said, the the big takeaway from that is you've got to stay engaged as stuff as comments keep coming on, you've got to go on and keep those posts alive.

It's also interesting too to you think about how you accounts that are just spamming that never go back and you know, reply to comments or likes or whatever, that makes it really easy to see who those bad actors are. Exactly, yeah, because they get no engagement. And one of the things that they said in here, and this is interesting because it it mirrors something I was reading in a in a different paper on a different topic about Google attribution, is that one of their metrics they look at is time to first engagement. So if you post this thing at, you know, how long does it take before somebody else engages with it to to boost that creator metric? So um one of the tactics that we've seen a lot of people do, and it because it works, is they'll put up a LinkedIn post to like tag 10 friends, right?

You know, or whatever. Why? Because you get that time to first like that much faster than a post where you have to wait for the algorithm to show it to people. If you can proactively hit hit up people that you know, it'll work better. So I don't know if we want to encourage like pods, uh, you know, sort of the idea of of uh groups of people doing stuff, but certainly tagging people in that you know are just gonna hit the like button might have some value.

Well, and yeah, you and I both know tons of people who have done those kinds of pod things that have just been able to generate insane amounts of traffic, but we won't talk about that on the live stream. I mean, we can. Um, but at the very least, I mean it's an easy, obvious win is it, particularly if you're at a larger organization where you've got a marketing team, company puts out something. Every employee go go hit the like button. Even if it's just the marketing team, like everybody, hey, I'm putting the post up at nine o'clock, you know, 9 01.

Go hit the like button. Yeah, do your likes to drive the action. And it's like you said, that calling out to pushes notifications. So that's just that's a huge benefit because it's it's getting forced in front of their face. Exactly.

Um, and it also again, if we go back to what we were looking at um when we were looking at the the NLP model, your first degree network is what's being hit. So if your employees are also connected to people um within the network, or you have uh uh a couple of influencers who are you know sure bets, that time to first like is really going to matter because then that influencer's audience is more likely to see it because they're one degree away from that post. Um, so I mean, for influencer marketing, you might even want to specify contract-wise, like hey, you know, part of the part of the service level agreement for how much we're paying you is that when we publish something, you have to hit like on it and within 60 minutes. Yeah, or even share to push it out. Yep.

Yeah, do the upstream and downstream metrics. So you like to boost that creator score, and then you share it downstream to make sure it propagates. It'd be very interesting. We should we should see if uh anyway, anyone wants to test that idea. Okay.

So that is number three. Um, when you're looking at content, the upstream metrics and the downstream metrics matter. You've got to get that time to first like down. Um, and as a creator, make sure that you are engaging with every single person that responds to your stuff, even if it's just saying thank you, you know, over and over again, just to keep that engagement chain going. The last one that we want to take a look at today is this relatively new post.

This is from uh last year on decay rate and uh and dwell time. So I thought this was really interesting. They don't look just at clicks because not that many people click on something, but they can absolutely measure how long you spend looking at an update. And they've got this thing called dwell time, where they look at the probability of you taking action, you know, upstream and and downstream, those metrics, and they're saying that clicks certain or viral actions are not always measurable, they're a very binary indicator of engagement, they're noisy and positive signals are sparse. Dwell time, aka how fast you scroll by an update or engage with it, is real valued.

It's always measurable, it can be more reliable, and there's they've got tons of data. So every time you open up your LinkedIn app or you're on the website and you're scrolling you know through your feed, they're looking at how fast you go by. Is something thumb stopping or not? And they built these models to look at just how fast something gets skipped. One of the things I thought was very interesting in their model is they looked at you know, things like uh is this a post about a job, is it a job anniversary post, is it a job change, is a video, is an article, and the skip time is almost identical across them.

So it doesn't matter like the content type. What it does matter is whether it's of interest to the person. So as you're creating content on LinkedIn, you've got to do something to stop that person's thumb. You've got because the moment that they they just cruise on by it, like you know, and you see that that sort of first uh thing there, uh, that it's negatively impacting how easily your content is seen. Yeah, that's interesting.

And they are all very tight, it's very close, but it's also interesting to see that images are just sell you know, visual stuff grabs, and the job change thing is interesting too. If you're uh people are interested in, like, oh, who's moving where, what's going on, the soap opera is attractive. Um this is why you see a lot of you know, when when you read how people structure LinkedIn posts, they start off with that provocative statement as that first sentence. Um, and it's not just because they're they're aspiring journalists or something, it's because they know, you know, the folks who get good engagement know whether or not they've been told that by uh reading these models or not, that if they can get you to stop and get you past that skip threshold, um, then your your your post, your content is not going to get downranked in in how often show it to other people. On the other hand, if everybody skips past it, um it's gonna show up less and less and less.

So you've got to do something to stop the thumb. Tastefully. Everything goes to the sewer eventually, but yeah, make it uh prove its quality, not uh just PT Barnum. Exactly. Now the one thing they don't say is they don't tell you what the skip time is, they've normalized it, obviously, because you know, if they if they know it's three and a half seconds, then you know whatever, but it's definitely short.

It's a it's a short period of time. Um so what are some of the things you've seen people doing on LinkedIn to to stop the the thumb? Yeah, that's a great well, and uh it was interesting what you just mentioned about normalizing the time. Because yeah, I could totally see somebody f hiring a bunch of Turks just saying, okay, spend four seconds here on this, you know, you could easily map that out and see how it goes. Um LinkedIn content, I it's a lot like other social stuff.

I think you get in this bubble of users. You know, there's certain types of users. Like there's a local incubator that I follow, and I read everything they put out because they don't put out too much, and it everything tends to be relevant and on target. So that's a big deal. I think there is a lot to be said.

You can, you know, some of the classic stuff of social is have a good image. You know, a good image is key. If you don't have an image, you're just kind of choking yourself off. Um, yeah, and then tagging people is the other. I pretty much get looped into stuff that I've been tagged on or that other people have found interesting.

That goes right to the top of the pile. And then after that, it just becomes kind of morning or afternoon news check. You know, I just kind of scroll through the pile to see. And um, yeah, in fact, I think this is good for you to talk about tuning and you know how you've tuned the algorithm a little bit because I think that's huge. There's a couple people who show up all the time because obviously the algorithm knows I like their stuff and I'm gonna engage with it.

And so I get fed that stuff on a regular basis. Yeah, exactly. Um it really does come down to reinforcing that network, right? Uh so as the consumer telling the algorithm this is what I like, this is what I don't like. Show me more of this, show me less of this through through your actions.

Um making sure that you start outright blocking people that you just don't want to see. Uh, a block is essentially a complaint. We know that uh in discussions about a multi-objective optimization, LinkedIn does take that into account because 40% of their revenue comes from their their talent management software, their HRS, uh, the ability for them to uh plug into an enterprise company's HR system and say, Yeah, here's 500 candidates that that you know all match the the job description you're looking for. And if they get people leaving because they're pissed off at you know seeing you know so and so politician in in their feed that that's not a good thing. So to the extent that they will deprioritize things like that to reduce complaints, that's a good thing.

But yeah, from a consumer perspective you want to make sure that you're do you're doing that from a a a marketer's perspective that first headline really matters. The image does matter. It was interesting when looking at this that um video performed relatively poorly uh you know people uh didn't stay engaged with it as much as like you said an image or a job change uh as things in fact articles seem to be you know outperforming video so if you're spending a lot of time on LinkedIn video you may want to do some A B testing of your own you know to publish an article publish an image publish a video and see if what kinds of performance you get because right now it's looking like you know images sort of winning the winning the battle there. Yeah that is interesting. Well and how about too what more have you heard because we had seen some modeling and some stuff talking about how like for company pages you only want to post a post a day.

It's not like other networks where you want to be spamming seven, eight times a day. Is that still the case or is there anything you know what do you suggest on that phone? I I think based on what we've seen, um if you can keep the upstream and downstream metrics solid on every post, then yeah, post as you as often as you can. If on the other hand you're posting seven times a day and it's getting skipped every single time in the feed um you're just shooting yourself in the foot. I mean, it's it's stuff that's no good.

Um, so post only as often as you have the bandwidth to to do what you need to do to keep that post alive. Um, getting that time to first like that we were talking about, um, having somebody with your community management software uh interacting with people all the time, um tagging in people, uh, and making sure that nobody scrolls past it, you know, to the to the best you can. If if that that might mean you might not post maybe more than twice a week, right? If if that's all you can afford to um to marshal your resources to get that engagement, you might be resource constrained. Yeah, but that's a great point.

So then it comes down to make it engaging and keep it uh keep the engagement going so that it keeps running. You probably don't want to be pumping up your press releases four times a day on that. That would be a really bad idea. Exactly. In fact, you know, after this, one of the things I'm thinking about doing is I've been curating content and putting it up on LinkedIn, and it gets occasional engagements here and there.

I think I'll just go pull a plug on that and just you know do an original post every couple of days, you know, something that where I can spend the time to feed it. Um, but that way when I do have an announcement, I've still got essentially that social credit, you know, built up, that social currency stored up that the algorithm is likely to help show it. Um, as opposed to just the curated stuff that doesn't get a ton of engagement. That's fine for Twitter, but clearly, based on everything we've talked about today with on LinkedIn, it's getting scrolled by. Yeah, it's not the kind of stuff you want up there.

Yep. So recapping, people you may know, clear out your inbox. Uh just sooner rather than later. Um, as a creator, um, as a participant on the network, make sure that you are writing and commenting and discussing and interacting with the stuff that you want to be known for. Uh that is your area of expertise.

Uh, and if you're changing areas of expertise or you want to change areas of expertise, make sure that you're participating in things, you know, following hashtags, etc. that are in the career or the subject that you want to be known for rather than sticking solely to the stuff that you are known for. If you are a creator, you need to make sure your upstream metrics are just as important to you as downstream metrics. So create, tag in, get that time to first like it down, and engage like crazy with the with the content that you publish. And then from a content perspective, make sure that you you increase dwell time as much as possible.

Stop that thumb from continuing to scroll by. You need to break that moment. Otherwise, your post is going to get downranked if it's below whatever the arbitrary threshold is. And these are machine learning models, so I would imagine that number is a moving target day to day as it is constantly retraining. Any other tips, John, for uh helping folks get more out of their LinkedIn feeds?

No, that's solid stuff. Again, the big one is go clean up all your stuff. I've got a lot of I've got a lot of delinquent invites I gotta go deal with. Exactly. So uh if you got comments or questions about stuff that we talked about today, you know, pop on over to the Slack group over at analytics for uh trustinsights.ai slash analytics for marketers.

Yeah, happy to chat about this stuff there. Uh in fact, I may put together uh the links to the different papers and posts and push that into the group so that folks can uh can read in greater detail. I will warn you it's very some of it's very, very technical. So uh just be aware of that. So with that, uh we will see you next week.

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 TrustInsights.ai slash TI podcast at our weekly email newsletter at Trustinsights.ai slash newsletter. Got questions about what you saw in today's episode? Join our free analytics for marketers slack group at trustinsights.ai slash analytics for marketers.

See you next time.