Reverse-engineering LinkedIn's feed algorithm from their published papers

LinkedIn published two research papers and an engineering blog post explaining their new feed algorithm.

Their goal is:

to connect every member to insights, ideas, and inspiration that move them forward. The most valuable content is timely, relevant to their professional goals, and grounded in trust.

As a professional exercise, I reverse-engineered these articles into the new posting rules below.

Admittedly, these rules leave out the most important part. Human behaviour.

I’m not an outreach specialist. Just an ML engineer interested in recommender systems. Below is how it works from a technical perspective.

Tier 1: The high-impact mechanics

Your profile is read alongside every post you publish. The retrieval system pulls in your name, headline, company, industry, and title and processes them together with the post content. The model is built on LLaMA 3, so it understands these fields semantically. If your profile says one thing and you post about another, the system has weaker context for placing your content with the right readers.
The model understands meaning. A post about "recommendation systems" can reach someone whose history is "content discovery" without any shared keywords, because the model already knows what topics relate to what. Keyword stuffing doesn't help anymore. The model gets it from natural language.
Active engagement counts more than passive engagement. The system tracks "professional interactions": long dwell, react, comment, repost. Active actions (like, comment, share) flow through a separate gating layer from passive ones (click, skip). A post that sparks comments and reposts trains the model far more strongly than a post with the same number of fleeting clicks.

Tier 2: Smaller effects that compound over time

Posting consistently about the same topic compounds. The ranker reads each reader's last 1,000 interactions in chronological order, with recent activity weighted more heavily. It's detecting where someone's interests are heading, not just what they've engaged with overall. If a reader is on a learning journey in your field, every post you publish in that field reinforces the trajectory. They see more of you. Switching topics dilutes the signal because no single trajectory builds.
Long dwell is one of the actions the system optimises for directly. Not just one of the things it tracks. Thresholds vary by post type. Posts scannable in two seconds may earn a click but rarely a long dwell. Substantive posts that reward reading time generate stronger signals than headline-bait.
The same encoder handles you as a reader and you as a creator. Your profile and behaviour shape both. A consistent professional identity across headline, posts, and engagement makes your content easier to place semantically. Mixed signals make that placement fuzzy.

Tier 3: System-level dynamics

Cold-start works differently now. The model can place a new account based on profile alone, before any engagement history exists. The blog highlights this is "especially powerful for cold-start scenarios." A new account can start being discovered almost immediately, provided the profile gives the model something to work with.
There's no fixed lifespan for a post. The ranking paper notes posts can persist for weeks, even though most engagement arrives in the first 24 hours. Engagement signals feed back into the retrieval system within minutes, keeping the system's view of the post current. Older posts that keep earning interactions stay in circulation. Ones that stop earning fade naturally.
Your reach isn't capped by your network. Posts can be shown to people outside your connections if the system finds them relevant. The retrieval paper is explicit that its LLM system serves "suggested content from outside of the member's network based on the member's topical interests". Follower count doesn't define reach.

Sources:

LinkedIn Engineering Blog (March 2026): "Engineering the next generation of LinkedIn's Feed"
arXiv 2510.14223 (Ramanujam et al., October 2025): "Large Scale Retrieval for the LinkedIn Feed using Causal Language Models"
arXiv 2602.12354 (Hertel et al., February 2026): "An Industrial-Scale Sequential Recommender for LinkedIn Feed Ranking"

LinkedIn claims the new algorithm brings +2.10% time spent on Feed in A/B testing.

But I think in the long term, it's user behaviour that will define its success or failure. Do people want to see strangers' opinions in their feed? Or strangers showing up with critical comments under their posts?

Currently most of the ideas, insights and opinions have very strong self-promotional bias either in the posts or comments. That would need to change for the new model to take off in my view.

Comments

AutoModerator•2 months ago•1

If this post doesn't follow the rules, please report it to the mods.

I am a bot, and this action was performed automatically. Please contact the moderators of this subreddit if you have any questions or concerns.

Dramatic_Object_8508•2 months ago•1

this is actually one of the more interesting breakdowns i’ve seen because it lines up with what a lot of creators already suspected, linkedin is moving way more toward semantic understanding instead of simple keyword matching

the “consistency of signal” part is probably the biggest takeaway. people think the algorithm rewards random daily posting, but it seems more like it rewards clear identity + repeatable topics over time

also makes sense why engagement quality matters more now. a few real comments from the right audience probably teach the system more than thousands of empty impressions

honestly feels like most platforms are converging toward the same thing now, less gaming hashtags, more building recognizable expertise and trust

UnoriginalSandwich•2 months ago•1

Spent about eight months as a content strategist at a B2B SaaS company trying to figure out why our LinkedIn posts kept getting buried. This was late 2023, right when everyone started noticing the algorithm felt different. We were doing everything "right" on paper: consistent posting, decent engagement, relevant topics. Still flatlined.The profile coherence thing in this post is real and massively underrated. Our head of sales had a profile that said enterprise CRM but kept posting about startup culture and founder mindsets. Total mismatch. The content just didnt reach anyone useful.We ran Shield for analytics to figure out what was actually performing. Helped us see the engagement patterns but didnt explain the distribution problem. Eventually the team switched to Taplio partly because the scheduling and engagement builder let us actually control who we were commenting on, which fed back into who the algorithm thought we were. That targeting piece mattered more than the AI writing stuff.The part about active vs passive engagement being processed through separate layers tracks with what we observed. Posts that got three genuine comments from relevant people in the first hour outperformed posts with thirty likes by a wide margin. Not three times better. More like ten times in terms of reach.Keyword stuffing was already dying back then. The semantic matching described here is why. Stopped trying to cram "B2B SaaS marketing" into every post and just wrote clearly about the actual problem. Reach went up. Topic clustering over time is slower to build but it compounds in ways that feel almost unfair once it kicks in.

Adventurous_Ebb7614•2 months ago•1

The active vs passive engagement distinction is the part most people are sleeping on. We spent about four months last year trying to grow two consultant accounts in the legal tech space and kept hitting a ceiling around 800 impressions per post. My colleague was obsessing over posting frequency, which turned out to be almost irrelevant compared to what was actually happening in the comments section.The semantic matching piece checks out too. We had one client whose profile said "enterprise software" but who kept posting about change management. Reach was patchy for months. Once we tightened the profile language to match the content topics, distribution got noticeably more consistent within about three weeks. Not a dramatic overnight shift, just steadier.On the tooling side, we had been running Shield Analytics to track what was working at the post level, which gave us decent visibility into engagement patterns. We also had one account on Taplio for scheduling and the engagement builder feature, which basically lets you prioritize commenting on specific people's posts rather than just scrolling and hoping. The comment-first approach ended up being more useful than anything we did with posting cadence. The model clearly rewards accounts that are active participants in conversations, not just broadcasters.The profile-as-context point from the paper is probably the most actionable thing in this whole breakdown. Most people treat the headline as SEO for recruiters and ignore how it interacts with post distribution. Treating it as a semantic anchor for your content topics seems to be where the real use is, at least based on what we saw.

More like this

EVERYTHING ABOUT THE INSTAGRAM ALGORITHM IN 2026

When you post, Instagram doesn't evaluate your content all at once. Distribution is continuous and adaptive. The system is constantly re-ranking your post based on signals it collects over time. Early engagement matters a lot, but posts can pick up hours or even days later, especially Reels. It's not a single batch test. It's an ongoing one. What you need to understand is that the algorithm is always watching the same core signals, and most people are optimizing for the wrong ones. **What actually moves the needle** Instagram's CEO confirmed this year that three signals are driving distribution more than anything else right now. Watch time is number one by a significant margin. Viewers decide within about 1.7 seconds whether to keep watching. If people are dropping off in the first 3 seconds, your post dies. If they make it past 50%, that's a strong signal. If they rewatch, that's explosive. Your retention curve is more important than your like count, full stop. Second is likes per reach, meaning the percentage of people who actually liked your post out of everyone who saw it. This matters more for reaching your existing followers than for growing to new audiences. Third, and this is the one most people are underestimating, is DM shares. When someone sends your post to a friend, Instagram treats it as a stronger endorsement than a like or even a comment. It signals that your content is worth recommending to strangers. Every post should have a built-in "send this to someone who needs it" moment, intentionally. If you're still optimizing primarily for likes in 2026, you're behind. **The format breakdown** Reels are for reaching new people. Carousels and photos are for your existing followers. Stories are for keeping those followers from leaving. They're not interchangeable. They serve completely different purposes in the algorithm. Carousels are underrated right now. Instagram does re-rank posts over time, which means a carousel that didn't land on the first impression can get another shot. The takeaway: make every slide worth stopping on, not just the first. Stories aren't optional if retention matters to you. Accounts that post consistently to Stories see meaningfully fewer unfollows. Stories keep your existing audience warm while your Reels pull new people in. **What you should actually be doing** Forget posting volume targets. Quality is the prerequisite. High frequency with low quality lowers your retention metrics and actively hurts your distribution. Run this instead: Every day: one high-quality Reel with a hook in the first 2 seconds and a clear share trigger built in, plus 3 to 5 Story frames to stay visible and keep your audience connected. Three to four times a week: a carousel optimized for saves and shares, something educational, useful, or worth returning to. Every single post should pass three checks before it goes out. Does the hook land in under 2 seconds? Is there one clear idea? Is there a reason someone would send this to a friend? **On niche consistency** Your last 9 to 12 posts define how Instagram categorizes your account. The algorithm rewards tight topic focus and punishes accounts that drift between unrelated content. Whatever angle you've built your account around, stay in it consistently. It's not about being in a broad niche. It's about having a distinct point of view within one. A hundred fitness creators exist. Only a few have a perspective that's immediately recognizable. That's the real differentiator. **Where I've seen this work** I grew from 100 followers to 360k using these principles. Grew 10+ accounts from 0 to 10k and sold most of them. The process was the same every time: understand what the algorithm is currently rewarding, make content that earns retention and shares, stay consistent for months not weeks, and adjust based on what the data tells you. It's not exciting. It's a system. Systems win. **TLDR** This post blew up last time so I'm bringing it back with answers to the most common questions I got. Before anything else, few things I wish someone told me earlier: 1. **Consistency** is the only thing that actually matters. I know everyone says this and everyone ignores it. That's literally why most people fail. The people winning are not smarter than you, they just didn't quit. 2. Video quality matters more than most people admit. Drop CapCut, get Adobe Premiere or hire an editor. Skip Fiverr, find editors in **Discord communities** instead, way cheaper and actually good. 3. Stop wasting hours on scripts, hooks, and hunting for content ideas manually. I use **SocialHunt** for all of that. You can train it on viral content in your niche and it handles the research and scripting side so you can just focus on filming. 4. Use **Superflow** to handle distribution, workflows, and repetitive ops. If you’re doing things manually, you’re capping growth.

3 months ago

255

Any people that runs 100 to 10k followers pages (any social media)?

Hey all, I own a business that involves many clients that promote their product, art or website with TikTok (also Instagram, Youtube and Facebook). I want to offer them a way to post their content, or to get content created and posted for them on existing TikTok pages. Basically you got pages in a specific theme (sport, anime, cinema, culture, memes...) and you rent it for a specific period of time. You get paid every month, and all you have to do is basically connecting the account once to our system. You can opt-out anytime you want, you keep the full ownership of the account. That's basically a way to monetize your pages passively. For now I'm doing it myself with my own pages (3 tiktok pages with 500 to 3k followers, and 1 6k subscribers Youtube channel), I made around $800 renting these for less than a month. If someone is interested, please comment or DM!

3 months ago

222

Everything dying at 300 views for so long before I finally caught the problem

I've been absolutely obsessed with short form content for the last two years. Like people have staged actual interventions about my health level of obsessed. I'm talking 11-14 hour days breaking down what separates successful videos from failures, experimenting with different hook variations, rewriting scripts until my brain hurts, testing every editing approach I could possibly get my hands on. Why this level of obsession? Because I'm absolutely certain short form video is the backbone of everything right now. Growing followers, selling anything, generating opportunities, creating brands from nothing. Every part of it depends on whether you can hold someone's attention for 30 seconds. But here's what nearly made me quit entirely: despite the constant daily grind, nothing was hitting. I'd pour 7-8 hours into crafting one video only to watch it crash at 300 views. Tried every tactic from every person claiming to have figured it out. Bought their courses. Applied their "proven" methods. Still going nowhere. I seriously started thinking maybe I'm just not the type of person who can make this work. Like maybe there's some fundamental ability I'm completely lacking. Then something clicked. I'm grinding constantly, but I'm operating completely blind. I don't actually know what's broken. I'm essentially just trying random things hoping something eventually works. So I stopped hunting for some mythical viral code and started analyzing actual data. Analyzed my last 50 videos second by second, documented every retention drop, and discovered 5 consistent patterns that were systematically killing my performance: 1. **Vague mysterious hooks are totally invisible** "This will transform you..." gets scrolled past every time. But "I used resistance bands for 55 days and my shoulder mobility actually decreased" stops people mid scroll. Specific concrete details destroy vague teasing without exception. 2. **Seconds 5-7 are where everything gets decided** Most viewers leave between 4-7 seconds if you haven't proven it's worth watching. I was creating slow buildups like a complete amateur. Now my strongest visual or most compelling number arrives exactly at second 5. That's where the hook that genuinely holds people. 3. **Any gap beyond 1 second absolutely kills your retention** Tracked this obsessively, anything past 1.2 seconds makes people think the video stopped. What feels like natural comfortable pacing to you reads as complete dead time to someone scrolling. Cut significantly tighter than feels normal. 4. **Visual variety is absolutely critical** If nothing changes on screen for more than 3 seconds, attention vanishes without warning. I started constantly rotating camera angles, cutting to b-roll, moving text placement, literally anything to maintain constant visual movement. Went from losing 50% at the halfway mark to keeping 70%. 5. **Rewatch rate is dramatically more important than most people realize** Videos people watch more than once get pushed exponentially harder by the algorithm. Started planting subtle details that aren't obvious first viewing, editing faster, adding elements worth discovering on rewatch. Rewatch percentage jumped from 8% to 31% and reach went completely through the roof. Honestly the biggest shift was abandoning all guesswork and actually measuring what was happening at every second. Came across this one app that goes way beyond showing where people drop off, it literally tells you why and exactly how to correct it. That's when everything transformed. Went from averaging 300 views to hitting 18k in about 4 weeks. Regular analytics show you people are leaving. This one shows the exact second, the actual reason, and what to adjust before your next post. If you're uploading consistently but stuck below 1k views, your content isn't the problem. You just don't know what's genuinely working versus what you assume is working. Listen, I'm sharing this because breaking through was honestly one of the hardest things I've tackled. I really wish someone had just explained exactly what needed fixing when I was stuck there. Would have saved months of confusion and doubt. So that's what I'm doing now for anyone who needs it.

4 months ago

221