Are you tired of posting consistently on LinkedIn only to watch your content disappear into the void? You’re following all the “expert” advice, like posting at optimal times, using hashtags, and engaging with comments—yet your reach keeps declining and your posts feel like they’re shouting into an empty room. You’re not alone in this frustration, and you’re probably wondering if there’s something fundamentally broken with your approach or if the platform itself has changed the rules without telling anyone.
That’s exactly what our guest today discovered, and instead of accepting mediocre results, he did what any good data scientist would do: reverse-engineered LinkedIn’s algorithm from the ground up. Christopher Penn is an authority on analytics, digital marketing, and artificial intelligence, co-founder and Chief Data Scientist at Trust Insights, best-selling author of 9 books with over 750,000 copies sold worldwide, and an IBM Champion who has shaped four key fields in the marketing industry, including AI adoption in marketing.
Social Pulse Podcast host Mike Allton asked Christopher Penn about:
- Algorithm Reverse Engineering – How Christopher used data science techniques to uncover the real factors driving LinkedIn’s algorithm, not the surface-level advice most marketers follow.
- Viral Content Formula – The specific technical elements and mathematical patterns that actually determine whether content gets distributed widely on LinkedIn versus getting buried.
- Data-Driven Strategy Implementation – How to apply Christopher’s research findings to create a systematic approach for improving your LinkedIn reach and engagement using measurable tactics.
Learn more about Christopher Penn
Resources & Brands mentioned in this episode
Full Transcript
(lightly edited)
Mike Allton: I wonder if we could start just talking about how most marketers are frustrated with their declining organic reach on LinkedIn and probably every other social network.
What made you decide to take this data science approach to solving this rather than just accepting it?
Chris Penn: Oh, I still get crappy reach.
No, this was actually inspired, as many things are with LinkedIn, with people on LinkedIn, real people who come across as snake oil salesmen, saying, “If only you follow my secret guide, comment guide in the comments. I’ll send you the secret to the LinkedIn algorithm,” and that general approach irritates me on a variety of levels, mostly because I know it to be factually incorrect.
And this is something that I think we can start with: There is no LinkedIn algorithm.
An algorithm implies that there is a methodology, a system that has a repeatable input and it has a repeatable output. LinkedIn does not have an algorithm. LinkedIn has 14 AI systems networks together that process the data and turn it into a result, because when we look at the system that makes up LinkedIn, LinkedIn is one of the world’s largest knowledge graphs, according, at least, to LinkedIn Engineering. It is a knowledge graph kept entirely in memory and six petabytes of RAM on 14 servers around the planet. And the underlying mechanics of the system make it. They have to keep this thing operating. They have to keep it updated quickly. They have to scale it, and it has to be able to make split-second decisions as it renders things.
So the idea, just from a data science perspective and an AI perspective, of one algorithm doing all that is absolutely ludicrous. But more importantly, once you realize the system itself is all these subsystems, then you realize how difficult it is to quote, hack the algorithm. You can’t; you might briefly find an anomaly in one of the models, weights somewhere along the line.
But, oh, by the way, according to LinkedIn Engineering, the weights recompile every hour. So what worked for you before lunch for your super secret LinkedIn algorithm hack probably stopped working after lunch.
Mike Allton: So walk us through your methodology.
How was your approach different?
Chris Penn: I reverse engineer nothing. Absolutely nothing.
What I did was—and LinkedIn is probably the best of all the social networks for this—I gathered up everything that they said was in the system.
So one of the things that LinkedIn does that they do really well, probably better than any social networking company, is they tell you how the system works. So there’s a LinkedIn Engineering blog that is stuffed full of blog posts about how different components of LinkedIn work, from spam traps and safety issues to news feeds to all that. So that’s one aspect.
So, you go back and just gather up a few hundred of their blog posts because the engineering team is prolific. And some of it is very, I would say, business and user-focused, like saying, here’s how our trust and safety systems work. And other posts are very technical. Here’s how we use Apache Kafka in our message queuing system.
Then the other side of the coin is that LinkedIn sends its engineers, its distinguished engineers, out to submit papers to conferences like KDD, which is part of American computing, something like that. The CM is the name of the conference. ACM, KDD, and publishing papers for things like NeurIPS (Neural Information Processing Systems) and other AI, like hardcore AI. We’re not talking ChatGPT, we’re talking about the low-level AI stuff sending papers. They submitted two papers, one for ACM last year, one for ACM this year, that walked through step-by-step the architecture of the LinkedIn system, and in extremely technical detail.
And that this is one of the reasons why your average marketer, your average business professional, your average social media marketer doesn’t know what’s going on. It’s not because they’re ill-informed. It’s because these papers, in some cases, are so technical that you, if you’re not swimming in the world of AI, don’t know what they mean.
Like, looking at the graph neural network that they built, they have a custom neural network just for managing the size of the graph. So what what I do is I gather up all the data that I can find, all the information I can find, and I use generative AI systems like Google Gemini and stuff to distill it down, sequence it, and then figure out, okay, how, and based on their own stuff, how does it actually work?
Google’s Notebook LM is a fantastic resource for this because it can handle lots of documents, and it also generally, unless you convince it otherwise, will not give you an answer if you did not provide the data. So’s likelihood of it hallucinating is much lower. So you put all the blog posts and all the papers and all this stuff into it from the official source only.
You ask it questions. Okay. How does the first pass ranker work? How does the second pass ranker work? How do liger and liran, and liquid all work? And the tools will tell you the system will say, based on this paper and or this blog post, this is how this part of the system works. We did that.
I did that over and over again for all the different pieces until I had a working understanding of the system. All 14 components. And then what we did was, after that, take the findings, put that in Google’s Gemini, and say, okay, now let’s make this accessible to non-technical audiences. Because it’s great to know something like, liger or Venice or Kafka, but that doesn’t help a marketer going, I just want people, I want more than two people to comment on my post.
Mike Allton: Yeah, and I was going to say, that’s why frankly we’re so happy to have people like you who understand and have the time and want to take the time to review all this information to even know that it exists out there. Many folks did not know all that. You sent this to me when you first created this document. Months ago. And I was completely unaware of a lot of the content and the sources that you were referring to. And I appreciate you pushing back on how we’re framing a lot of these things; that’s important.
And I also really enjoy that you talked about Google Notebook LM. For those listening, if you haven’t used it, that is a fantastic tool. Chris has given you one example of the many ways that you could use this tool to put a bunch of different pieces of information, documents, and reports, and get really succinct summaries.
It’s frankly, it’s really nice if you just. I don’t have a lot of time. You’ve got a link to report, drop it into Google Notebook LM, and let the podcast hosts give you a really entertaining two or three-minute summary of that particular document. A great tool, definitely bookmark that. But I’m wondering, Chris.
What were the biggest surprises about how these LinkedIn systems actually work?
Chris Penn: Okay, so there are a lot of answers to this question.
There is a paper called The Unofficial LinkedIn Algorithm Guide for Marketers. It’s published by our company, Trust Insights. It’s totally free. You don’t have to pay any money for it. No, you’re more than welcome to grab a copy of it. What changed in this version? We did the first version in mid-2024, and then this new version, which came out about a month and a half ago. It has been newly revised based on LinkedIn’s changes.
So LinkedIn itself has made some pretty large changes on the backend, and the major changes that they made were in the rich ranking section of their system. They have new inference models that effectively apply a type of language model to what’s going on LinkedIn for deciding how to show you content. In the past, they had more classical AI systems, things that did regression analysis and stuff, and that’s still in there. In the new one, though, that is a language model.
There are several of them that are part of the process. And what that allows ’em to do is take advantage of more of the language. So that was a big architectural change. A, it’s a lot more compute-intensive now. So LinkedIn has had to make the system more efficient, but B, it changes how we think about language.
One of the things that’s most common about LinkedIn advice is Hey, do this at this time, respond to this number of comments, and comment on this. A number of people say that’s a lot of the general what to do, but very few people are talking about the language that they use. When you look at LinkedIn’s architecture and you look at the fact that does tokenization and embedding, which is essentially computer nerd speak for, reads, the words that you use and figures out how they’re relevant, all the words that you use on LinkedIn, count all of the words condition the model for how you get your recommendations, but also how other people get their recommendations about you.
So let’s say, Mike, in your case, you are a person who’s talking about social media, right? We’re on the Agorapulse Social Pulse podcast, and you’re talking all about social media, and then suddenly you start talking about sailboats, right? And you’re, and you talk about this. It’s looking at your profile, it’s looking at your first degree connections, which also matter.
It’s looking at the language you use in your post. It’s looking at the language you use in your comments. It’s looking at the language people use on your posts and so on and so forth in these concentric circles of language. And then it tries to figure out, based on the language model built in.
What should I show Mike when he logs in? Hey Mike, you’ve been talking a lot about sailboats. Maybe your thing is sailboats now. So here’s some more content about sailboats, and it’s showing your content to other people who might have an interest in sailboats. But if you’re all in on sailboats, guess what?
You’re not being shown for social media, right? Yeah. So if you are trying to boost the Agorapulse Social Pulse podcast, and you’re talking about sailboats. You are going to resonate less with the social media audience that you want because you’re not talking about social media. So one of the things that this investigation digs into, in terms of how it goes against marketing devices, is that people talk about the language that you use every day on LinkedIn, who and who you’re connected to, and who connects with you, and the language that you use in your interactions. If you’re all in on generative AI and ChatGPT, great.
But if you’re trying to promote social media. And you’re not talking about that anymore, it’s going to, you’re going to have challenges reaching the audience you want to reach. So a big thing that you have to be thinking about carefully is what kind of audience you want to curate. Who’s your ideal customer profile if you’re on LinkedIn, and then are you talking about the things that they talk about so that you have a greater affinity on a topic level with what they’re there for?
Mike Allton: Okay. First of all, I’m disappointed you didn’t use a food example like pizza. You went with sailboats. I’m used to you making me hungry on this show and talking about breaking things down by food. But it sounds to me like it’s almost more similar to Google search signals, and if I were trying to rank a piece of content or a website in Google, I’d be paying attention not to the the H1 could, but a lot of people tell you to focus on that kind of stuff and instead focus on what topics am I talking about? What other articles am I writing about? How are they interlinked? Am I getting links from other sites that are also talking about those topics? Am I understanding that correctly?
Chris Penn: To a degree, yes. Yeah. So I mean that there’s a bunch of other things, ’cause again, there’s 14 systems and six major stages that are all trying to decide and referee what’s going to happen with your content. But that first. The input gathering in the candidate generation is looking at who you are, what you talk about. Who are you connected to? What do they talk about? What do you talk about in other people’s places, and what do you talk about in your own places? So, yes, there is an element of that language, and that to me is the big thing is it’s when you look at the system from a language perspective, you start realizing, oh. It’s great if I go and comment on stuff, but if I’m not being substantive in my comments, like if you’re one of those dumb donkeys who’s got a comment bot, it’s like, Hey, great insights, Mike Rocketship. That’s worthless. That is completely worthless. That does nothing for you. That does nothing for the other person, and critically, it does not create language that the system can understand.
This is what you are about.
Mike Allton: That’s exactly where my mind was going at that moment. I was thinking about having more thoughtful comments. I imagine while we wouldn’t necessarily see a direct connection or correlation between, we’re spending more time on comments and then we’re getting more reach and visibility, I imagine if we were mapping that out over time, we should expect to see improvements as a result.
Chris Penn: Yeah, in general, if you’re, this is the thing. These systems are designed to foster engagement, right? Because ultimately, the goal of a social network, at least today’s modern ad-driven social networks, is to sell you more ads, right? That is the goal. In order to sell you more ads, you have to be there longer.
So for you to be there longer, you have to stay engaged. For you to stay engaged, you have to see and be seen for the things that you supposedly care about. And if you are. If you are on LinkedIn and you are doing things that other people don’t care about, just the mathematics of that alone are going to push you towards irrelevance, right?
One of the things I think about, even though we both use and absolutely love Agorapulse as software, right? You have to be very thoughtful about how you do your automation. Posting something that does not get any engagement, including you, harms you in two different ways. It harms you on the downstream metrics, like the propensity for likes, comments, and shares, and harms you on the upstream metrics, the creator metrics.
Are you likely to even engage with your own post? That’s something that came out three-ish years ago in a paper. What they implemented was some creator metrics upstream to say, What is the likelihood that the creator will engage with their own content once it’s published? And will you respond to comments if you’re posting like, Hey, here’s yet another blog post we wrote on a blog, and you don’t care, hence why you’ve automated it, then you are going to cause compounding harm to your ability to get reach on LinkedIn. So one of the things I would strongly advise is to use social media monitoring and social media and management tools, but use ’em thoughtfully and use them in a way that will encourage you as a human to engage more rather than less. If you’re like, I’m just going to set it, forget it, then yeah, you can enjoy your five impressions per post.
Mike Allton: Yeah, and I just finished a little experiment of my own. This is obviously anecdotal. But I was using Jetpack for WordPress and allowing its social aspect to automatically publish new pieces of content to social channels. And you can see right on your posts what your reach is like for a LinkedIn post. And you could go back and you can see that just going down. For me, even though I’m continuing to post other kinds of content natively, like I normally would long-form posts and videos and that sort of thing, repeatedly sharing short link posts is killing my reach. So I took that bullet for all of you listening. You don’t have to do that.
Don’t do that. Don’t automate your systems, your LinkedIn posts in that respect. But I want to stick with this commenting and engagement theme for just another second, because LinkedIn still has these collaborative articles.
Are you seeing any benefit in commenting or engaging with collaborative articles on the rest of your profile?
Chris Penn: I don’t even see them like, I don’t engage with, ’cause I don’t even see them, they’re not what LinkedIn chooses to show to me. Mostly because I am very heavily engaged in the AI community and so that is all I see all the time. And people are complaining about M dashes.
Mike Allton: Yeah, I did recall. I spent some time in 2024 looking at and potentially engaging with AI-focused collaborative articles. And not surprisingly, they were filled and riddled with AI-generated responses, so they seemed fairly mediocre at best and a waste of time at the least. So folks, we’re talking with Christopher Penn about what factors go into social media systems and processes like LinkedIn and how social media managers should be navigating them.
How do you think these LinkedIn systems may differ from some of the other platforms?
Chris Penn: None of the systems has an algorithm. Meta has published system cards for all of their services. Facebook and Threads, Instagram, and the system cards show you the 22 different components that make up Facebook. They don’t go to Facebook’s website itself.
They don’t go into as much technical depth as LinkedIn does. You have to go hunting in other places for those academic papers for more of that. But every system, every modern social network uses some combination of multiple systems because they have to. There is no way you could make one algorithm to do everything.
Now, LinkedIn is a bit more complicated than that because LinkedIn since its early days, LinkedIn has had multi-objective optimization built in, which is a challenge that other platforms don’t have. So LinkedIn has three businesses, right? Three fundamental business lines: LinkedIn Marketing Solutions, LinkedIn Talent Solutions, and LinkedIn Sales Solutions. LinkedIn Sales Solutions is to sell tools to salespeople to sell more stuff. LinkedIn Marketing Solutions is also known as ads, and LinkedIn Talent Solutions is selling people like us to employers and so their incentives have to be balanced among those three things.
It’s not just ads, whereas a platform like Facebook is just selling ads. All they care about is selling ads. They don’t really care all that much whether you stick around or whether you’re happy or things like that. In fact, it’s pretty clear from Meta’s testimony in various congressional hearings that they would actually prefer you to be unhappy as much as possible because unhappy people stick around longer and engage more as they doom scroll their way to poor mental health.
LinkedIn, for you to stick around and be of value to like talent solutions, which is for candidates, you have to not be pissed off enough that you quit, right? Yeah. And so the LinkedIn algorithms for multi-objective optimization, which happen in the re-ranking stage in their system, are different from, say, what a Meta would do or what, whatever the hell is going on at X, formerly known as Twitter.
Mike Allton: Yeah, that makes sense, ’cause like those listening, if you’re ever in Sales Navigator and you’re looking at contacts, you might see something like a little tag that says Intent to Buy, which is not information that you’d ever see on another platform. But they’re using other signals that the individual, I don’t know, maybe responded to DMs, maybe, but they did something that would’ve let the platform know that they’re not buying e-commerce stuff. They’re willing to talk to salesmen from another organization.
For someone who’s not a data scientist, how do they apply your research findings to their LinkedIn strategy starting today?
Chris Penn: We made that easy. So in, in the paper itself, there is a series of, there’s explanations of what’s going on. I’m sharing my screen here. There’s an explanation of what’s going on. And then there’s checklists. Okay, here’s what to do for your LinkedIn profile. These are the things that you should do to improve your profile.
There is a launch checklist before you post, what you should do as you post, what you should do, and what to do after you post. And then there’s the engagement, quick things to do every day, five, 15-minute things to do, a couple times a week, and then stuff to do that will take you longer. And those sections of the paper are exactly what to do. That’s, that was the whole point because it’s fun to dig into the tech, but ultimately, very few people actually care about the tech other than nerds like me, everybody cares about how I make stuff work better so that I get more than five impressions on my posts, that’s what the paper’s designed to do.
Mike Allton: Awesome. We will obviously have a link to that in the show notes, folks, so don’t worry, you don’t have to go searching for it. The link will be below. Obviously, you’ve talked a lot and you’ve written a lot about using AI and machine learning in marketing.
How do you personally use these technologies to optimize your LinkedIn content strategy?
Chris Penn: I don’t, ’cause I don’t care. Good answer. This is a strange thing, right? But for me, on my profile, I don’t care all that much about a content strategy. For our company page and our company content, we absolutely use tools. So we, for example, have built a machine inferring ideal customer profiles.
And our team at Trust Insights takes the ideal customer profile, takes the content we want to post, and tries to figure out, okay, what wording should we use that will appeal to this? And then we have our writing style guides, which are just technical files that we have. We, so we give that to a generated AI tool. Say, here’s the writing style. So if I want to impress my CEO, Katie Robert, this is how she writes on LinkedIn. They have my ideal customer profile of what the customer cares about, and then those get prompted together to create content that we put onto our company posts. And then we measure that, we schedule it in Agorapulse, and we measure it within the Agorapulse reports.
My personal content strategy is whatever I feel like ranting about that day. And I just see what happens.
Mike Allton: And it works for you.
Chris Penn: It works for me because, again, because I don’t care. LinkedIn, social media in general. And this is weird, saying on the Social Pulse podcast, social media in general is not something that drives a lot of business for me.
It drives a little bit of awareness. It drives a little bit of community, but public social media. Not it. Private social media. Very much so. So we’re talking Slack, Discord, group chats, things that happen behind the velvet rope, as we call it. That is highly impactful. Slack is one of our biggest drivers of prospect and lead generation because we run a Slack community: Analytics for Marketers, 4,200 members in there. Because we run it, we get to set what the rules are like; here’s what you are and are not allowed to do. But the biggest driver for me personally and also for the company is email. Email marketing is orders of magnitude more effective for me because I’ve also spent 15 years building my email list.
My email has just like 300,000 readers weekly. I spent 15 years building it. So even though I’ve got like 45,000 odd followers on LinkedIn, from a, Hey, I want to talk to you about doing business with you, email is where I spend my time. So from my strategy on LinkedIn, personally, it’s whatever I want to talk about that day and roll the dice.
Mike Allton: And I know many of the social media managers listening will relate to that because they’re managing social media on behalf of a brand. Their own profiles, if they post. If they don’t. Many of the folks that I see doing social media for big companies may never post a LinkedIn unless they’re in between jobs.
They feel like they need to go back to that job function. But the last question I have for you, Chris, is I’m wondering. For folks who, obviously, this is a concern for everybody, for if you’re working professionally for a brand, then you need to get rich on your social posts. You need to get engagement, you need to be able to foster that community, and then have them take a next step for your brand, whatever that might be.
How do you recommend that they keep up with what’s truly happening with these platforms? Is it just to follow you, or are there other sources, resources you think they should be turning to?
Chris Penn: My narcissistic egomaniac would say, Follow me. My useful, helpful answer is to follow the technology companies themselves, follow the engineers, read the engineering blogs, ’cause that’s where all the good stuff happens. That’s where all the announcements happen. That’s where all the details happen. And today, in the era of generative AI, there’s no longer an excuse for saying, I don’t understand. You get it on the LinkedIn blog post, which talks about Apache Kafka. You’re like, I don’t understand this great copy. Paste it into Gemini or ChatGPT or whatever, and say, explain this to me in marketing terms. What does this mean for my marketing? The answer might be that it has nothing to do with marketing, right? The ability to scale Hadoop clusters means nothing for your marketing, and you can say, Great, I didn’t, I don’t need to read this post.
On the other hand, if you’re talking about messages queuing in Kafka, you may say, What does this mean for my marketing? And it will say, if LinkedIn is using Apache Kafka for message queuing, it may indicate they’ve got a bottleneck that they need to solve for rapid speed. And then you can give you some ideas about how to interpret that as it applies to what you’re doing on LinkedIn.
But that’s where there’s an expression that I love in the AI community. The truth is in the code, right? Yeah. You can write position papers all you want, you can prognosticate, you can pretend you’ve hacked the algorithm, whatever the truth is in the code. And the closer you are to the code, the more involved in reality you are.
So I would say for any social media manager, get close to the engineering, get close to the systems to understand what they do because they are the vehicle that you’re driving your marketing around in. And if you don’t understand. You’re what? What bus are you on? You may be going to the wrong place.
Mike Allton: And folks, you can set up a Feedly, even a free Feedly account, to monitor RSS feeds for Meta’s newsroom and that sort of thing.
Do you feel like X is still the place to follow engineers and the kind of people you’re talking about?
Chris Penn: As much as I despise it. As much as I despise the owner and the management, yes. A lot of the AI community, the hardcore AI community, is there and on Discord.
Discord is actually one of the best places to go for hardcore AI stuff because all the big tech companies have their own Discord servers. Because Discord is super cheap to operate, it’s 500 a year for a tier three server. That is pennies for a tech company. Like they can find that in the change in the couch, in the lobby.
And so there’s a lot of stuff that happens there. And what is also interesting is because it’s behind semi-closed doors, there’s no clout, right? When you’re on LinkedIn, you’re trying to figure, how to craft the perfect post to get the attention of your audience. And you go onto Discord, and people try to do the same thing.
And boy, you get your ass handed to you super fast by people like, dude, this is not LinkedIn. Get out. Or another person saying, like this, these are the 18 things that you just said that are all wrong. And so it’s a much less polite, but more truthful place than public social media.
So those are the places that I would suggest people hang out because you will see stuff. But the other thing is to make use of your data. So if you go into Agorapulse and you go into the reports tab, right? And you, there’s an export button, you can say export CSV of your reporting, and you’re like, great, I still don’t know what this all means.
Go to Google Colab, write a prompt that says, Hey, here’s my social media data. I don’t know what I’m doing, but I’d really like to know what’s working in my social media. Here’s my CSV file. Google Colab will then take your prompt, take your CSV file and start writing Python code to analyze it because you never let generative AI do math, but you always let it write code, and it will say, Hey, I did a regression analysis on your LinkedIn post that you exported from Agorapulse, here’s what’s working for you. The tools are there that will allow you to make use of your data. The closer you can get to the code, the more the closer you are to the truth.
Mike Allton: Fantastic. Chris, you’ve been amazing as always. I know everyone’s minds are burning with ideas and questions that they want to reach out to you and ask you questions.
Where should they go?
Chris Penn: The easiest place to go is TrustInsights.ai, cleverly placed below my name.
Mike Allton: Fantastic. That’s all the time we have for today, friends. Of course, we’ll put all of Chris’s links in the show notes below, as always. But don’t forget to find the Social Pulse podcast on Apple. Drop me a review. Let me know what you thought of this episode, what questions you wish I had asked Chris, and I’ll bring him on here another time.
And please join our exclusive community on Facebook Social Pulse community, where you can meet great experts like Chris, as well as network with literally thousands of other social media professionals just like you. Until next time,
