So What? Analyzing Your LinkedIn Strategy with Agentic AI

Christopher Penn – 00:01

We’re not even bothering with the intro. We’re going straight in.

Katie Robbert – 00:04

Yeah, thank you for anyone who is still here. We were having some sort of strange technology issue with our streaming software. None of us could get in. We were all scrambling around. For what it’s worth, we all try to log in about 15 minutes before the show for things like that. If you’re in any sort of speaking role or you do webinars, there’s always a tech check, and we do that every week.

However, despite doing that, none of us could get in or troubleshoot. So we finally figured it out. We’re a few minutes late, but we’re gonna roll with it. On today’s live stream, “So What?”, we are analyzing your LinkedIn strategy with agentic AI. I’m Katie, that’s Chris, that’s John. Let’s just roll right into it. Happy Thursday.

Christopher Penn – 00:59

Happy Thursday. All right. For those who don’t know, we republish our unofficial LinkedIn algorithm guide approximately whenever LinkedIn creates a big update. LinkedIn dropped a massive update about a week ago. You can find it at the TrustInsights.ai LinkedIn guide URL.

If you’ve not gone in and looked at this, it is a 138-page guide that walks through how LinkedIn works. As much as people like to say these things are opaque, if you read engineering blogs and the technical papers that underpin them, LinkedIn actually tells you exactly how the system works. There are some things they don’t give away, but there’s enough from an architecture perspective that we can understand the system.

Today we want to talk about how we would understand and analyze our LinkedIn performance so that we can do better on LinkedIn. Part of that has to start with how the system works at all. Before we dig into this, I want to ask you two: how do you think LinkedIn works in terms of showing up in the news feed, which is what marketers really care about? I want to be seen either individually or as a brand.

Katie Robbert – 02:36

I think it is fueled on misogyny and bias. Final answer. The reason I say this is because we have actually found the data. Maybe this is part of the update this time with the algorithm, but we’ve seen the data where the LinkedIn algorithm favors male-presenting profiles over female-presenting profiles, which is a big problem.

Why is it doing that? All of this stuff is programmed by humans, and humans have inherent bias in their brains. They may not consciously realize they’re doing this, or they are, which makes them terrible. They may be subconsciously including just their own information. Basically, that’s my final answer.

What we know is that anytime we think we have the LinkedIn algorithm cracked — take out your pronouns, change your name, comment more, post more, like more, share more — they make a change. It feels like a moving target. But the consistencies are that you’re never going to see what you want to see. That’s how it works. You’re welcome.

John Wall – 03:52

I have read the Marvel Comics version. It’s like 2,000 posts. They enter the Hunger Games and all fight to the death. If it makes it through the first round, it goes into a bunch of safety scanners. This is where all the bias and weird stuff is applied. At the end of that, it spits out a bunch of stuff for me which is real estate, sketchy deals, and a lot of crummy stuff. That’s the comic book version of what I read in the latest LinkedIn paper.

Christopher Penn – 04:26

Here’s at a very high level how it works. This is in the NotebookLM version of our paper, which you get if you download the paper. There are three different AI models under the hood. There is the Causal LLM; there is Qwen, which is a profile embedding LLM; and then there is a Transformers-based sequencer called Feed SR, or Generative Recommender, which is the third stage.

If we think about this, it is classic old-school AI in a lot of ways: retrieve and rank. If you’ve been in SEO, you know exactly what this is. Get the data and then arrange the data in some sort of fashion. First, from things like profiles, posts, and comments in any language, LinkedIn gathers up about 2,000 candidates that it could show you. It essentially does this weeding with a model called Llama 3, which was created by Meta.

They use the three billion parameter model, which is a very small model that will run on your phone. They use it because it’s extremely fast. Back in December, Trust Insights published an evaluation of bias in this model, which is what Katie was talking about. As far as we know, we never got word back from LinkedIn saying they fixed this, so those biases probably are still there. But that first model says, “Out of all the posts, here’s 2,000 that we think semantically match what a target member would be looking for.”

It then takes that 2,000 set of posts and calls a second model called Qwen. The 600 million parameter models are very small models that look at your profile. It pulls your profile data, turns that into embeddings, and then grabs about 1,000 of your last interactions — likes, comments, reposts, dwell time, and so on. With a formula that they did not publish, it forms a chronological sequence of your last 1,000 interactions.

We then bring in your profile, the profile embeddings of the people nearest you, and the semantic match candidates. We say, “Out of these 2,000 candidates, based on all this sequential data, what’s this user going to interact with?” Their goal is interaction. They want you to engage, like, comment, and share. In combination, these systems say, “Here is all your past history. Katie, I think you’re gonna like this post the most. Here it is.”

What’s different about this than previous versions of LinkedIn is that because they’re using language models, it’s now very heavily conditioned on two things: language consistency and your interaction history. If Katie is constantly posting about strategy, digital transformation, change management, and organizational behavior, it will see this long history and the language associated with it and say this is what Katie is about.

If there’s somebody else — let’s pretend John is the CEO of Marketing Over Coffee — and John is scrolling around on LinkedIn and is interested in change management, John will see Katie’s post before mine because maybe I’m posting about super tech nerd stuff. If Katie posts about something totally random like football, and that’s never been a part of John’s interaction history in the last 1,000 interactions or 30 days, John won’t see that post. Even if John’s interacted with Katie a lot in the past, because there’s no semantic language overlap, he won’t see it.

Katie Robbert – 08:52

That seems like a huge missed opportunity. One of the functions of LinkedIn is you can follow someone without connecting with them. If I really like the writings of John Wall, I can click follow on his profile, and theoretically in my human brain, I’m going to see the stuff that he posts.

But if he’s posting about hockey and I’m looking for football, then even though I’m following him, I’m not going to see his post. I have to seek him out and search him to see what John posted about recently. I can interact with it, but it’s still not necessarily going to show me the next time I open LinkedIn.

Christopher Penn – 09:28

We’ve had that experience. You’ve gone to a friend’s profile and noticed three posts you never saw. That’s why, because the system did not see a semantic overlap of interest between the two. That is the big frustration for a lot of LinkedIn users right now: the machine is trying to guess at what you want to see despite you saying you follow these people or like these pages.

Katie Robbert – 10:01

In some way, it wants us humans to be singularly focused on one topic and very uninteresting. God forbid we post outside of what we normally post to expand our audience; then everybody loses that information.

Christopher Penn – 10:22

Yes, that’s true, because it all revolves around semantic distance. It’s not exact match keywords. It’s words that are related from a statistical relationship. If I’m talking about XGBoost, which is a classical machine learning algorithm, that is semantically within distance of generative AI. Someone reading about ChatGPT has a higher chance of seeing my XGBoost post than a post about football, because football and generative AI are much further apart in the training data.

Katie Robbert – 10:55

It sounds like what you’re describing is like playing the SEO keyword game. If I have five core themes I always write about, I should probably do keyword research to figure out the related keywords and then make sure I’m including them. The LinkedIn algorithm now kind of operates the way a search engine does in terms of those semantic keywords.

Christopher Penn – 11:43

You’re on the right track, but keywords won’t help. When we’re talking about semantics and embeddings, we’re talking about statistical probabilities of words, and that may mean there are words that are not keywords that are topically related. Think about topics, not keywords. Should I be writing about change management? Well, what’s involved in change management? The 5P Framework by Trust Insights, the ADKAR model — all these different frameworks.

If you were to do a keyword search for ADKAR, it’ll take a couple of jumps to get to change management. But in a semantic distribution or a probability matrix, ADKAR and the 5P Framework should be right next to change management because they’re all conceptually the same thing. People use them near each other in the massive amounts of data these tools have been trained on.

If you want to improve your LinkedIn performance, especially in terms of showing up, you want to get a semantic understanding of the space of the people you’re trying to interact with. Then take your content and see how closely it matches that semantic space. You want to do this at a person level and a topical level.

At two different checkpoints in the LinkedIn system now, language models are ingesting your profile data. That’s not just, “I’m an AI expert and I do AI all the time.” I will have things in my profile like the causes I volunteer for, like the Bay Path Humane Society, certifications, and education. There will be semantic overlaps that matter. LinkedIn knows this because it often suggests you might know people from your college or organization. Those semantic overlaps are how it picks up those aspects. All the things in a person’s profile matter, even if they’re not what you talk about 24/7.

Katie Robbert – 14:05

So what do we do with that? For those who aren’t part of our community, over in our free Slack community, Analytics for Marketers, the question of the day was for people to drop their LinkedIn profiles so we could all connect. Since we get new members every week, it’s nice to make sure people feel like they can make those connections. Not pitches, just connections. You can join for free at TrustInsights.ai/analytics-for-marketers.

You just went through a lot of technical information, which I feel like I have a decent high-level grasp on. What do I do today? How do I make sure I’m showing up? It’s great to understand this, but how do I make sure I’m showing up? It’s not only your profile but also your LinkedIn strategy. That includes what you’re posting, when you’re posting, and how you’re interacting. What do we do?

Christopher Penn – 15:30

When really doesn’t matter. Anyone telling you the best time to post on LinkedIn is Tuesdays at 1:00 PM is wrong. The fact that generative AI is now in charge means that recency doesn’t really matter. This has been an issue for a year. People ask why they are seeing posts that are three weeks old. It’s because they’re relevant, not recent. Relevancy is king on LinkedIn at the moment.

We’re going to look at this from two different ways of what you can do next. We’ll start with the hard way. I took the paper we wrote and all the source papers and said, “Let’s build an application that can measure two profiles and see what you should do to improve your profile to be semantically relevant.” There are four different AI models running behind the scenes here that attempt to replicate as closely as possible the current LinkedIn algorithm.

There is Llama 3, Qwen, a BERT model for the mettoids — which are semantic concepts in posts — and then there is Nemotron. We don’t have access to the Feed SR model. That is LinkedIn’s custom baby, and no one has access to anything close to it, so we’re using a generic model as a standard. This is an approximation, but what I did was take my profile, Katie’s profile, and our last 10 posts.

I put them in here and said I want to improve my profile. Llama 3 — that first gate model — says there’s about a 96.8% overlap. Katie and I are very closely aligned in terms of the content we share. We’re highly aligned at 98.4%. We talk about very similar things in terms of the language we use, and our composite score is 96.95%. This says there’s a high probability I will show up in Katie’s feed because I’m so well aligned.

It’s going to try to rewrite parts of my profile to get an even closer match. If I wanted to specifically target a certain kind of buyer on LinkedIn, I would use the profile of the CEO of Apple, Tim Cook, instead of Katie’s. I would have Tim Cook’s posts and profile, and it would say, “If you want to close that gap semantically, here’s how you should change your LinkedIn profile to match.” Then I could put a post in and see what I could do to modify it to be more semantically aligned with who Tim is and what he cares about.

Katie Robbert – 19:05

That’s interesting, but that’s going one person at a time. Could you do something like run your ICP — your ideal customer profile — and get the technographics, firmographics, and demographics in a cumulative way? Could you put that information here as the target profile so you can show up for a larger potential audience? It’s hard to do it one by one because you’d be changing your profile every other day.

Christopher Penn – 19:52

The problem with an ICP is that it’s an averaging out of a bunch of different characteristics. University names, degree names, and charity names are good examples. Tim Cook is going to look different from Larry Ellison, even though they’re both high-tech CEOs. Those two individuals are going to be very different. What you would do is grab five LinkedIn profiles — like Larry Ellison and Marc Benioff — and their last 10 posts.

You would use a tool like Claude Co-work to programmatically take control of your browser, gather their posts, and store them. Instead of having one person’s profile and conversation history, you have the group’s history and profiles as the target. Then you could do your calibration against the group as a whole. It’s not quite the ICP, but it’s a similar idea with all the nuances you want access to.

This is the hard way because it uses the exact models. It’s going to give you the most faithful replication of what LinkedIn does for people who are not LinkedIn employees. If you wanted to do something more practical, you would fire up a tool like Claude Co-work and say, “I want to build a semantic understanding of these two profiles.”

Let’s start with a basic prompt: “Today we’re going to do a semantic comparison of two users on LinkedIn, Chris and Katie. I’m going to provide you with their profiles in text format and the last 15 posts from each person in JSON format.” Then you tell it to write a piece of Python code that will compare them using Cosine similarity and BM25 similarity. Then provide suggestions for Chris to adapt his language to more closely resemble Katie’s.

Your final output should be an HTML file styled with CSS that’s easy to read and prescriptive, telling Chris exactly what he needs to do to be found more semantically relevant by Katie.

Katie Robbert – 23:08

I like how this is the simple version, but you still had to understand things like looking for BM25 and Cosine similarity. You still have to understand that’s part of it.

Christopher Penn – 23:23

This is why there’s so much snake oil on LinkedIn about “the ultimate algorithm hack.” There is no “just do this.” When you’re dealing with language models, you’re dealing with very complex systems. But if we understand how those systems work based on what LinkedIn has told us in their technical papers, we can use today’s generative AI models to get the closest approximation possible. We know the outcome is semantic similarity, which is the heartbeat of all those systems.

Katie Robbert – 24:20

Chris, while that’s running, we have a question about ranking. If you had to rank semantic understanding, member profiles, content posts, and the last 1,000 actions of the author/reader, what has the lowest to highest leverage?

Christopher Penn – 24:45

The answer is to use the algorithms LinkedIn uses. They stated in the Causal LLM paper last fall that they use Llama 3B, and they specifically use Matryoshka embeddings with 3,072 dimensions. You just need the words to put into Claude to say, “I want you to help me install Llama 3, three billion parameter version, eight-bit quantization. I want to use it for embeddings and generate 3,072 dimension Matryoshka embeddings to compare content.”

Effectively, you’re using the vanilla version of the same model LinkedIn is using in production, but you’re doing it with your desktop computer instead. In terms of leverage, what gets you leverage is higher similarity. If you have one pool of embeddings about change management and another about college football, those are far apart. It’s going to be difficult for there to be enough overlap for LinkedIn to show the post.

Katie Robbert – 26:07

John, I hope you were taking notes; there’s going to be a quiz later. The second part of that question is: do we know how those 1,000 to 2,000 candidate posts are shortlisted from a bigger pool and which one we will see first?

Christopher Penn – 26:18

Those are shortlisted by using Llama 3 and the Matryoshka embeddings to process everything in the LinkedIn system. That’s why it’s such a small model. A three billion parameter model can run on your phone, and LinkedIn uses it because it has to process everything. When you use regular generative AI like ChatGPT, Gemini, or Claude, you’re using models that are 10 to 15 trillion parameter models, which is why they’re so slow.

When you use a three billion parameter model for embeddings, because it’s such a finely focused task, it can process all of that. As for how they know which you see first, that’s phase two: retrieve and rank. Phase two is that ranking to sort through those 2,000 posts. Qwen pulls your profile embeddings, the Feed SR pulls in your last 1,000 interactions, and it does a statistical prediction of what post is going to have the highest score across those three measures. That is what shows up first in your feed.

Katie Robbert – 27:48

LinkedIn has tools like Sales Navigator that they want you to purchase for advanced searching and lead generation. Does this version of what you’re saying we’re running in Claude Co-work act as a way to do this if I don’t have the money for Sales Navigator? John, as head of business development, is this something you feel would accurately represent where you are today versus five years ago?

John Wall – 28:52

Sales Navigator is a whole separate thing. I look at that more as data enrichment and being able to put together lists of people and getting additional access to message them. To me, this is a whole different thing; this is more of a content play. This is about how we post stuff that gets in front of the people we want to reach and how we make a convincing case.

I’m interested to hear more because so far this has been a black box. We don’t have our own last 1,000 interactions easily unless we’re doing some kind of extra exporting, which is probably against the terms of service. You can run models and get a good idea of what should be done, but you can’t promise what works. I’m also hoping that Clippy is going to help guide us along and show the way; I have to give him a shout-out as a special guest star.

Christopher Penn – 30:06

Claude Co-work finished an assessment. They used a different similarity method called TF-IDF (term frequency-inverse document frequency). This is an older version of natural language processing and is word-based. There’s a much bigger gap here than when you use the models LinkedIn actually uses. This is a great example of the difference between going keyword-based, which says things suck, versus embedding-based, where there is a strong overlap.

But there are still useful things in here. Katie talks a lot about leadership, change, people, process, framework, and method. Chris, not so much. We both talk about AI strategy, data, analytics, comms, and content. If I wanted to close this semantic gap, these are the blatant terms Katie uses that I don’t: leadership, product, strategy, team, people, architecture, and decisions.

It also gives some prescriptive recommendations. To attract Katie’s attention, instead of saying, “Here are three ingredients for success with any significant project,” I would start by saying, “The leaders who actually see ROI from AI are not starting with code; they’re starting with a decision document.” It’s the same thing but couched in the language Katie uses. I’m like, “Wait, did I write that?”

Katie Robbert – 33:02

What’s interesting is you still have to find your own voice and not just replicate someone else, or you get into IP infringement. But if you’re trying to get the attention of a prospective hiring company — let’s say I want to work at McKinsey — I could take the McKinsey business profile and see how I can use some of their language so I’m a more appealing candidate. It demonstrates that I understand their culture and the work they’re doing.

Christopher Penn – 34:15

There’s a couple of important use cases there. If you are looking for work and you know who the hiring managers are, you follow the prompt we used in Claude Co-work to understand the language you want to be using. If I wanted to attract Katie’s attention, I should trim the old crap in my profile that isn’t helping my similarity scores and write content about leadership failure framing.

Case number two: if John had a billion-dollar deal in the pipeline and needed to be seen in the LinkedIn feeds of the C-suite for the next 30 days while they are deliberating, John would take those profiles, run this analysis, and start cranking out content with the highest probability of being seen by that group. You wouldn’t do this for every profile, but if you need to throw the kitchen sink to win a huge deal, this is what you do.

Katie Robbert – 36:00

I think that’s a smart way to use this. We can’t change everything about our profile all the time, but we can create the right kind of content. That goes back to understanding your ideal customer. By the way, if you want help with that, go to TrustInsights.ai and contact John. We just redid the services page. It’s lovely. I really like the idea of appealing to the right people. We talk about how to show up, but once you show up, are they going to stick around?

Christopher Penn – 36:51

What is interesting is that you and Katie are saying the same things; you’re just saying them to different people in a different dialect. For Trust Insights, that is fantastic because we want to appeal to different networks. Your dialect is precise, technical, and practitioner-facing. Her dialect is operational, organizational, and leader-facing.

Katie Robbert – 37:21

What’s nice about using us as examples is we can look at this and see it’s correct. I’m not worried about you showing up in my feed because I know where to look for your stuff. But when John posts — who is a less frequent poster because of the nature of his work — I want to make sure I don’t miss his content. If we have a big deal, you want to see what those people are talking about, their pain points, and if they’ve made a new hire. Those are things you probably want to know.

John Wall – 38:38

It’s definitely the prime place to stock these contacts and see where they’re at. Another side door is to see who else they’re engaging with, whether it’s competitors or other things. If someone was enterprising, they would find all the prospects for competitors and pitch more appropriate content to get a better fit than everyone else.

Christopher Penn – 39:08

At the highest level, you would have cohorts of pre-gathered curated data for the major influencer segments. If you brought on a client that wanted to do influencer marketing, you would use a curated influencer segment for the reach they want and give them an audit of their editorial calendar. You would use the same technology LinkedIn’s using to make sure it aligns.

You cannot game this system. There is no secret hack because the weights of the models recompute hourly and flush daily. You may discover a hack that works for 12 minutes, and then it’s gone. What you should be focusing on is profile coherency, content targeted toward the audience you care about, and behavioral consistency over a 30-day window. You can get that paper for free at TrustInsights.ai/linkedin-guide.

Katie Robbert – 41:11

I want to acknowledge our friend who asked the question. We are all in the same boat of information overload. A lot of these major technology companies put the information out there, but they don’t write it in a way the average user can understand. We rely on thought leaders like Chris to find that information and distill it for the rest of us. Chris does an excellent job on this paper every time LinkedIn makes a major update, so go get it.

Give the paper to a large language model and say, “Summarize this for me in terms of ordering pizza” or whatever makes sense to you. If you love horseback riding, describe it in terms of a horse event. That is one of the beautiful things about a large language model: you can take these highly technical concepts and break them down. But first, you’ve got to get the paper.

Christopher Penn – 43:36

Go get the paper. There are four pages of citations so you can find the underlying sources. We used Claude Co-work today, but you can use any system — Google Gemini, OpenAI, and so on. Any system that allows you to put in data and programmatically generate code. You will save yourself so much time. The hardest part is getting the vocabulary down to understand what to ask for. That’s where people run into trouble, especially with something this highly technical.

Katie Robbert – 44:39

The big takeaway is that algorithms are a moving target. There is no gaming it. The best you can do is try to understand it and find the three things you can do to make your experience on these platforms more meaningful and connected. John, check the book.

John Wall – 45:14

And the book is free.

Christopher Penn – 45:17

Go get it.

John Wall – 45:30

We get real-time notifications, and copies of this are flying out the door continually. It’s definitely road-proven. There’s a reason we’re releasing this every time they do the updates.

Katie Robbert – 45:45

We’re not in the business of gatekeeping. This kind of information benefits a wide audience. The name of the company is Trust Insights, and we need people to trust the insights we’re giving them. If we start to charge for this, it loses that credibility. I’m grateful we give it away for free because so many people benefit from it.

Christopher Penn – 46:30

And we write our own software for it; maybe we’ll charge for the software. That’s going to do it for this episode. Be sure to subscribe to our show wherever you’re watching it. For more resources, check out the Trust Insights podcast at TrustInsights.ai/tipodcast and our weekly email newsletter at TrustInsights.ai/newsletter.

Got questions about what you saw in today’s episode? Join our free Analytics for Marketers Slack group at TrustInsights.ai/analytics-for-marketers. See you next time.