In a feed engineering post published in March 2026, LinkedIn announced it had rebuilt its feed: instead of a scattered set of independently running ranking systems, a single large language model now represents both the post and the member in a shared semantic space. The academic foundation for that model is detailed in the 360Brew research paper.
The Feed Now Runs Semantically
In the previous system, a post was represented by predefined numeric features — word count, format, past engagement rate, and the like. In the new system, the post text, the member’s past engagement, and their profile data are all processed as natural language; the model ranks by grasping what the post is actually about.
Keyword repetition provides no advantage
Semantic evaluation makes it easier to spot text that repeats a topic unnaturally. The keyword-density tactic that can work in SEO has the opposite effect on social media — it lowers readability.
Dwell Time: Time Spent Gets Measured
The feed system now uses not just whether a post gets clicked, but how long a user stays on it, as a signal. Lingering on a post for more than a few seconds signals a different level of interest than scrolling past it quickly.
Format choice affects dwell time
A short text post can be read and scrolled past within seconds, while a document (PDF carousel) or video format can naturally hold a user’s attention longer. That’s a structural reason document and video formats show up more often in feed tests.
Active Engagement Carries Different Weight Than Passive Engagement
Active engagement — comments and shares — carries more weight in the algorithm than passive engagement like reactions. The reasoning is that active engagement shows the user is genuinely engaged with the content, while passive engagement can be little more than a momentary reaction.
Comments in the first hour shape distribution
After a post goes live, it gets shown to a small audience first; the early engagement from that audience determines whether the post gets distributed further. Responding quickly to comments in the first hour after publishing can reinforce that early signal-gathering process.
A Common Mistake
The most common mistake is reducing the algorithm to a single tactical question like “what time should I post.” Timing is a factor, but the feed now semantically evaluates the post’s content and the quality of the actual engagement it gets — picking the right hour doesn’t turn a weak post into a strong one.
Summary
The feed system LinkedIn rolled out in 2026 evaluates posts by semantic content rather than keyword matching, weighting dwell time and active engagement. Format choice and post-publish engagement management in the first hour are the two controllable variables that directly affect distribution under this system.