I made a mistake in my algorithm video

I made a mistake in my first algorithm video.

And someone from LinkedIn called me out.

Two months ago, I published a YouTube video breaking down LinkedIn's algorithm based on a research paper their engineers had quietly published on arXiv. No announcement, no blog post, no press. Just a paper uploaded by their own engineers. So everything in that video was me interpreting what they'd put out there.

We still didn't have official confirmation from LinkedIn.

Then two things happened.

First, a Senior AI/ML Leader at LinkedIn named Adam Bird left a comment correcting something I said. I told you the algorithm reads your entire profile before distributing your content. Skills, job history, education, certifications, all of it.

He said that's wrong. And he was right.

Second, LinkedIn went fully public about their algorithm. On March 12th, 2026, three things dropped at once:

For the first time, we have confirmation from the people who built the system. So I went into research mode, cross-referenced everything with the original arXiv paper, third-party data, and what we're seeing across client accounts at Distinctiva.

And I made a second video covering all of it.

But I know some of you prefer reading. So here's the full breakdown:

What I got wrong about your profile

I said the algorithm uses your entire profile to decide who sees your content. Adam Bird's correction: only 5 fields travel with your content into someone else's feed. LinkedIn calls this your "author data":

That's it.

Your full profile (skills, job history, education, certifications, languages) is used on the viewer side to understand what each person wants to see. But it doesn't travel with your posts.

This matters a lot. Your headline and the industry/title you select carry way more weight in distribution than most people realize. The "full profile optimization" advice? It helps your profile page convert visitors. But it doesn't affect how the algorithm distributes your content.

What to do: Open your profile right now. Read your headline. If your ideal client wouldn't immediately think "this person can help me" from those words alone, rewrite it. That headline is one of only five fields the AI uses to decide who should see your posts.

I built a free audit kit around exactly this - including how to fix your headline and the other four author data fields. Grab it here:

→ [DOWNLOAD THE LINKEDIN ALGORITHM AUDIT KIT]

The old system has been replaced

For years, LinkedIn ran multiple separate systems at the same time. One tracked your network's activity chronologically. Another handled trending posts by geography. Another ran collaborative filtering based on similar members' interests. Several more ran embedding-based retrieval.

Each system had its own logic, its own biases. That's why the old advice worked. Post at 8 AM. Get your pod to comment immediately. Stuff your post with hashtags. You were gaming individual machines. And it worked because those machines were simple enough to trick.

LinkedIn replaced all of those with two things:

1. A unified retrieval model powered by a fine-tuned LLM (Meta's LLaMA 3) that matches content to viewers based on semantic meaning. It doesn't care about your hashtags. It reads your post and understands what it's about conceptually.

2. A Generative Recommender built on transformer architecture that processes 1,000+ of your historical interactions as a sequence. It understands your professional interests over time, not as random data points but as a curiosity arc. If you engage with machine learning on Monday and distributed systems on Tuesday, it maps that pattern.

Your first 50 words are your algorithm audition

This changed how we approach every client post at Distinctiva.

LinkedIn's retrieval system truncates your post text to the first 60 tokens (roughly 45-50 words) when deciding whether to include you in the candidate pool. Everything after that matters for ranking. But the initial filter, the part that decides whether your post even gets considered, focuses on the beginning.

Your hook now has two jobs: stop a human from scrolling AND give the AI enough semantic signal to match you to the right audience.

What to do: Pull up your last 5 posts. Count the first 50 words. Do those words contain your topic and your audience? If they could belong to anyone in your industry, they're too generic.

Saves are worth 5x a like

The system is optimized for what LinkedIn calls "Professional Interactors" - people who take meaningful actions like long dwells, saves, comments, and reshares. AuthoredUp's analysis found one save equals roughly 5x the reach impact of a single like.

The system also converts raw engagement into percentile buckets. It sees "71st percentile of view counts," not "12,345 views." This change improved the model's correlation with relevance by 30x.

Posts with frameworks, checklists, specific numbers, and step-by-step breakdowns get saved. Opinions and motivational takes get liked.

What to do: Before you publish your next post, ask yourself: would my ideal client bookmark this? Would they send it to a colleague? If the answer is no, add something specific. A framework. A set of numbers. A step-by-step breakdown.

If you depend on a pod, maybe it’s time to rethink your strategy

LinkedIn's March 12th announcement said it directly: they're working to make engagement pods ineffective and curbing comment automation and third-party tools that create fake conversations. They're reducing repetitive, low-substance posts and engagement bait.

The system tracks dwell time. It distinguishes between someone who read your post and left a thoughtful comment vs. someone who typed "great insight" in two seconds without reading.

What to do: If you're in a pod, get out. If you're using automation tools for comments, stop. Write for humans. The AI is watching whether the humans actually care.

Smaller accounts have a structural advantage now

LinkedIn's A/B test data showed the new system produced a 3.29% revenue increase for members with fewer connections and a 1.17% increase in professional interactions.

The old system was biased toward large networks. If you had 50,000 connections, you had built-in distribution just through network activity. The new system matches content to interest regardless of network size.

Shield Analytics data from 50,000 posts backs this up. A top-performing post from someone with 5-10K followers averages about 5,500 impressions. A median post from someone with 25-50K followers averages about 2,400. Execution is outpacing audience size.

What to do: If you've been waiting to "build a bigger audience" before going serious on LinkedIn, stop waiting. The system just got rebuilt in your favor.

Your engagement history trains your own feed

The system maintains a time-ordered list of every post you've positively engaged with. And Adam Bird confirmed: your engagement with other posts doesn't affect whether your content reaches someone else. This is about your own feed.

But your feed shapes what inspires your content. And what you create is what the algorithm distributes.

LinkedIn also tested including negative signals (posts you saw but didn't engage with) and it made the model worse. Removing negative signals reduced memory usage by 37%, processed 40% more training data per batch, and made training 2.6x faster.

What to do: Be intentional about what you engage with. Like, comment on, and save posts from creators covering your topic area. You're training the algorithm to show you the best thinking in your space, which sharpens what you create.

Go deep or go invisible

Tim Jurka confirmed it in his post: exceptional content can be distributed broadly across LinkedIn to members who are interested in the type of content you post, even if they don't follow you. Every piece of content has its own path based on topic, format, and timing.

If you try to be about everything, the AI can't build a clear embedding for you. You become unmatchable. But if you go deep on a specific professional topic, the system finds exactly the right audience.

What to do: Look at your last 20 posts. Could the AI identify a clear topical pattern? Or would it see a random mix of productivity tips, leadership quotes, and industry commentary? Pick your lane and commit to it for 90 days.

The algorithm audit kit

I put all of this into a free downloadable kit. 8 sections based on the 8 rules from the video, audit checklists for each one, plus AI prompts you can paste directly into ChatGPT or Claude to audit your own profile and content.

→ [DOWNLOAD THE LINKEDIN ALGORITHM AUDIT KIT]

If you'd rather watch

The full deep dive is on YouTube. About 18 minutes. Covers everything above in visual format, including Adam Bird's comment and how I traced the correction back to the official sources.

→ [WATCH THE VIDEO]

Talk soon,

D.