How the LinkedIn Feed Actually Ranks Your Posts

LinkedIn is unusual among the big platforms in one useful way: its engineering team publishes. Not marketing copy about "authentic content," but architecture write-ups with model names and tradeoffs in them. In March 2026 they published a long post describing a rebuilt feed, and it changes what good advice looks like.

What LinkedIn changed in 2026

The post, written by Hristo Danchev, describes a feed that "serves more than 1.3 billion professionals" and a new ranking stack "powered by LLMs and GPUs, that better understands what a post is actually about and how it relates to a member's evolving interests and career goals."

Two structural changes matter to you.

Retrieval was unified and made semantic. LinkedIn used to pull candidate posts from several separate sources: a chronological index of your network's activity, trending posts in your geography, collaborative filtering, topic indexes. They replaced that with one retrieval system built on LLM-generated embeddings. Their own example: a member interested in "electrical engineering" who engages heavily with posts about "small modular reactors" is a connection "traditional keyword-based systems might miss."

Ranking became sequential. Instead of scoring each post independently, LinkedIn built what it calls a Generative Recommender that "processes more than a thousand of your historical interactions to understand temporal patterns and long-term interests." The model treats a member's feed history as an ordered story, not a bag of preferences.

Practical translation: writing to a topic consistently now compounds harder than it used to, because the system builds a semantic representation of what you post about and matches it to people whose behavior sits near that meaning. Keyword stuffing was never great and is now beside the point.

The signals LinkedIn admits to

LinkedIn's help documentation, last updated in 2026, groups feed signals into three families.

Identity: "your profile details such as location, workplace, and skills" and your professional background. Your profile is a ranking input, not a business card.

Content: "how often a post is viewed or engaged with," "what the post is about and whether it provides knowledge or advice," how recent it is, whether it comes from someone you follow, the language, "how constructive or professional the conversation is," and tagged people or companies.

Activity: what you reacted to or commented on, "who you interact with frequently or recently," "content you spend the most time viewing," and "engagement others have with your posts and how recent it is."

The engineering post adds detail the help page does not. Feed signals include "what you've read, liked, commented on, returned back to, or simply scrolled past." Post popularity is fed to the model as a percentile bucket rather than a raw count, because raw numbers tokenize badly, and LinkedIn describes popularity as "such a strong relevance signal." And in the ranking head, tasks are split into "passive tasks (click, skip, long-dwell) and active tasks (like, comment, share)."

A separate LinkedIn help page on feed ranking adds two useful lines: "Our AI systems and algorithms consider hundreds of signals to determine what content appears in each member's Feed," and "our algorithm and systems do not use demographic information, such as age, race, or gender, as a signal to determine the visibility of content, profile, or posts in the Feed."

Skips are modelled explicitly. That is not new. Back in 2020 LinkedIn Engineering published a post on dwell time explaining why: clicks and likes "can be rare, especially for passive consumers of the feed," and "clicks are noisy indicators of engagement," because members click an article and bounce within seconds. So they built a P(skip) model that predicts whether your dwell time on a post will fall below a threshold, and they reduce a post's score in proportion to that probability.

Two things follow. The first line of your post is doing more work than the rest of it combined. And a post that people stop on, even without touching it, is not wasted.

One more line from the help page is worth knowing, because business owners ask it constantly: "Feed distribution isn't influenced by payments from third parties to LinkedIn, except for promoted (paid) content, which is clearly labeled." Buying ads does not lift your organic reach. Not buying them does not suppress it.

What LinkedIn does not publish

Be suspicious of anyone who tells you these with a straight face, because LinkedIn has published none of them:

• A penalty for outbound links. LinkedIn has never documented one. Third-party consultants report one, using their own scraped samples. It may be real. It is not confirmed by the platform.
• An optimal posting frequency. LinkedIn does not publish one.
• A "golden hour" threshold. No published number exists.
• Reach differences between personal profiles and company Pages. LinkedIn documents that Pages exist and that Page search ranking considers followers and "quality of updates, comments, and reactions." It does not publish comparative organic reach.

When a number in a LinkedIn tip has no platform source, it came from someone's dataset. That can still be useful, but treat it as one agency's sample, not physics.

The format and limit realities

• A post is capped at 3,000 characters, per LinkedIn's help page on posting, last updated in mid 2026.
• You can post as yourself or, if you are a Page admin, as the Page, from the same share box.
• You can restrict who can comment (anyone, connections only, or off) and schedule posts natively.
• Documents, polls, images, video and events are all native post types in the composer.

The composer detail that matters most is the least exciting one: scheduling exists, so "I was busy that day" is a solvable problem.

What a founder should actually do differently here

Fill out the profile like it is a ranking input, because it is. Industry, skills, geography and headline are named identity signals. A vague headline makes you harder to match to the right readers.

Pick a lane and stay in it for a while. Semantic retrieval rewards a coherent body of work. If your last twenty posts are about restaurant ops, the system has something to represent. If they are about ops, politics, a marathon, and a webinar, it has noise.

Write the first two lines like a headline, not a windup. The model that decides your fate is predicting whether the reader will skip. "I've been thinking a lot lately about..." is a skip.

Treat comments as distribution, not manners. Activity signals explicitly include who you interact with frequently and recently, and engagement recency on your own posts. Commenting is how you stay in the recent-interaction set of the people you want reading you.

Do not confuse impressions with the point. LinkedIn ranks for the reader's interest, and a post read closely by 60 people in your industry is a better business outcome than 6,000 impressions from a viral take about hustle.

None of that is a mechanics problem. The hard part is having something worth saying this week, and a reader who will take it seriously. That is the part Ripple Room was built around: a standing hour, business owners who know what you do, and comments written in a person's own words.

When this was last confirmed

The engineering description here reflects LinkedIn's March 2026 post and its help documentation as of August 2026. The dwell time work is from 2020 and describes a mechanism, not a current setting. LinkedIn does not publish model weights, thresholds or reach benchmarks, and it changes ranking continuously. If something here stops matching what you see in your own analytics, believe your analytics.