The LinkedIn algorithm, by LinkedIn’s own description, works in two steps. A retrieval system picks candidate posts by matching what each post is about to what a member engages with, using embeddings from a large language model. A ranking model then orders those candidates by reading the member’s history as a sequence of more than a thousand past interactions. LinkedIn described both in an engineering blog post that went out with its 12 March 2026 feed announcement. Everything on this page was checked on 11 November 2026.
Most guides that rank for this term mix two kinds of claims. Some come from LinkedIn: its engineers, its pressroom and its research papers. Others come from creators and tool vendors who study post data, and some come from nowhere at all. The “golden hour”, the link penalty and the “360Brew rollout” are often written as facts LinkedIn confirmed. This page keeps the two apart. The first half covers only what LinkedIn has said, with dates. The second half covers what outside studies found, with their method and sample size. The last part is a routine for B2B founders that rests on both.
What has LinkedIn said about how the feed ranks posts?
These are the primary sources, oldest first.
| Date | Source | What it says |
|---|---|---|
| 27 Jan 2025 (withdrawn) | 360Brew paper on arXiv | A 150B-parameter “research pre-production model” trained on LinkedIn data matched or beat production systems on offline metrics |
| 11 Aug 2025 | Tim Jurka, “How Does the LinkedIn Feed Work?” | LLMs identify the context of a post; ranking factors in the author’s professional identity, not age, race or gender |
| 12 Feb 2026 | Feed SR paper on arXiv | A transformer-based sequential ranker replaced the old DCNv2 ranker and served most feed traffic |
| 12 Mar 2026 | LinkedIn Pressroom, Tim Jurka and the engineering blog | New LLM retrieval and Generative Recommender ranking; action on pods, comment automation and engagement bait |
| 4 Jun 2026 | LinkedIn Pressroom, “Keeping conversations real” | Generic AI content is less likely to spread beyond a person’s immediate network |
| 8 Oct 2026 | Tim Jurka, “The Technology Behind LinkedIn’s Fight Against AI Slop” | AI-slop classifiers now cover all posts shown beyond the immediate network |
Tim Jurka is a LinkedIn vice president of engineering, as Inc. reported in October 2026. In his August 2025 post he says he has worked on the feed for over a decade.
Step one: retrieval picks the candidate posts
Retrieval is the step that decides which posts are even considered for your feed. LinkedIn’s engineering post says the old feed pulled candidates from separate sources: a chronological index of your network’s activity, trending posts in your region, collaborative filtering and several embedding systems. The new feed replaces them with one system that turns members and posts into embeddings from a fine-tuned LLM and matches them by meaning.
The same post lists what goes into each description:
- For a post: format, author information (name, headline, company, industry), engagement counts, article metadata and the post text.
- For a member: profile information, skills, work history, education, and a time-ordered list of posts the member engaged with.
LinkedIn says it trained the model only on posts members engaged with, and used posts that were shown but got no engagement as “hard negatives”. Its example is a member who follows electrical engineering and reads about small modular reactors: the LLM knows the topics are related, where keyword matching would miss it.
Step two: ranking orders what you see
Ranking decides the order. The engineering post describes a Generative Recommender that reads more than a thousand of a member’s past interactions in time order, with the actions taken on each one. It names those actions as long dwells, likes, comments and shares, and it separates passive signals (click, skip, long dwell) from active ones (like, comment, share).
The research paper behind it, Feed SR, says the new ranker served most of LinkedIn’s feed traffic for over three months and, in online A/B tests, raised time spent by 2.10% and likes, comments or reshares by 3.52% against the old model. The paper also says LinkedIn tried LLM-based ranking architectures and chose Feed SR for the best mix of online results and production efficiency.
Two points matter for anyone who posts:
- Dwell is a named signal. LinkedIn lists “long dwells” among the actions the ranker reads. It has not published how much each action weighs, so any “a save is worth five likes” figure you read is not from LinkedIn.
- The feed mixes your network with strangers. LinkedIn says the system balances content from your network, accounts you follow and suggested posts from people you have never connected with.
What about 360Brew?
360Brew is real, but LinkedIn has not said it runs the feed. The arXiv paper, first submitted on 27 January 2025 by Hamed Firooz and 22 co-authors, calls 360Brew V1.0 a research pre-production model with 150 billion parameters that handled over 30 predictive tasks on offline metrics. All four versions of the paper are now marked withdrawn; the arXiv note says the submitter did not have the right to agree to the licence at the time of submission. None of LinkedIn’s 12 March 2026 posts use the name. When a guide says “LinkedIn replaced its algorithm with 360Brew”, treat it as unconfirmed.
Your profile is part of the input
Jurka wrote in August 2025 that the ranking system factors in the author’s professional identity, such as experience, content shared on the profile and skills. The engineering post lists the author’s name, headline, company and industry inside each post’s description. LinkedIn has not said how much weight these carry. The practical reading is simple: a headline that names the topic you post about gives the system a consistent signal.
What LinkedIn says it shows less of
The 12 March 2026 pressroom post and Jurka’s post the same day name four kinds of content the feed will show less of:
- Engagement bait such as “Comment ‘Yes’ if you agree”.
- Videos that have nothing to do with the text of the post.
- “Recycled” thought-leadership posts with little substance.
- Comments from engagement pods, comment automation and third-party software or browser extensions, which Jurka says are not allowed on LinkedIn.
The June 2026 pressroom post, first published by LinkedIn’s executive editor Laura Lorenzetti on 20 May 2026, adds AI slop: low-effort AI content that “may sound polished on the surface but lacks any real unique perspective”. LinkedIn says such posts are less likely to be distributed beyond a person’s immediate network, that its first tests identified generic content correctly 94% of the time, and that it also acts on comments posted at scale with automation tools and replies that only restate the post. It says using AI to help you write is fine as long as the post represents your voice.
On 8 October 2026, Jurka wrote that the classifiers now cover all posts distributed beyond the immediate network with 94% precision, and that, as of August 2026, members saw 40% fewer views on what LinkedIn classifies as AI slop than a few weeks earlier. He also describes AI agents that each check one policy, such as whether a post is promotional, celebrates an achievement, or is timely.
You can turn ranking off for yourself
LinkedIn’s Help Center says the feed shows “Top” posts by default, personalized by your activity and interests. You can switch to “Recent” for posts from people you follow in reverse chronological order. Outside the European Economic Area, LinkedIn says Recent still shows the updates most relevant to you, in time order.
What do third-party studies say about LinkedIn reach?
Outside studies measure outcomes, not the code. They show patterns across many accounts, they cannot show cause, and the authors often sell LinkedIn tools or training. Read each one with its sample.
| Study | Sample and method | Main findings | Caveat |
|---|---|---|---|
| Richard van der Blom, 2026 Algorithm Insights report (as he described it on Creator Science, 2 June 2026) | 1.3 million posts from about 50,000 creators | Reach down about 60% for active creators over two years; engagement down 20 to 25%; optimal posting fell from 5 to 6 a week to about 2 to 4 | Full report is gated; figures here are from his interview |
| Richard van der Blom, 2025 Algorithm Insights report (as summarised by Expandi) | 1.8 million posts, 58,000 personal profiles | Organic reach down about 50% and engagement down 25% year over year | Second-hand summary by a vendor of LinkedIn automation |
| Buffer | Over 2 million LinkedIn posts | Posting 2 to 5 times a week improved impressions and engagement | Buffer’s page does not describe how the sample was drawn |
| Buffer | Over 1 million posts | Carousels earned nearly 3x the engagement of videos and nearly 6x that of text-only posts | Engagement, not reach or pipeline |
| Buffer | 72,000 posts | Replying to comments was linked to about 30% more engagement | Correlation; active authors may differ in other ways |
The common claims you will meet elsewhere, such as “70% of reach is decided in the first 90 minutes”, “external links cut reach by 60%” or “first-comment links are penalized”, do not appear in any LinkedIn source listed above. Some may be true on average. None is a rule LinkedIn has stated.
The 60% drop for active creators fits what LinkedIn says it changed. If the feed now matches posts to readers by topic and fills more slots with posts from outside your network, the creators who used to reach most of their followers will see the largest fall. That is our reading, not a LinkedIn statement.
How should B2B founders post for the 2026 feed?
Each step below ties back to a LinkedIn source or a named study. None of it needs a hack.
- Pick two or three topics your buyer pays to solve. LinkedIn’s retrieval matches posts to readers by meaning, and the author’s headline is part of each post’s description. A founder who posts about agency cash flow every week gives the system a clear match. Put the main topic in your headline. Our guide to LinkedIn personal branding covers the rest of the profile.
- Write posts people stay to read. “Long dwells” are a ranking signal by LinkedIn’s own account. Specific numbers from your work, a worked example, or a short framework hold attention longer than a one-line opinion. Our LinkedIn post examples show formats that do this.
- Format for the phone. Short paragraphs and a first line that states the point help a reader decide to tap “see more”. Our free LinkedIn text formatter adds bold and italic text, and the LinkedIn character counter shows where the cut falls.
- Use AI as an editor, not the author. LinkedIn says it limits generic AI content beyond your network with classifiers it reports at 94% precision. Start from a call, a deal or a number only you have, then let a tool tighten the wording.
- Drop the bait. “Comment YES”, unrelated videos and recycled advice are on LinkedIn’s own list of things it shows less of.
- Keep comments human. Pods and comment automation are named targets. Reply to comments yourself; Buffer’s 72,000-post sample links replies to more engagement.
- Post at a pace you can hold. Two to five posts a week is where both Buffer’s data and van der Blom’s 2026 view land. A LinkedIn newsletter gives you a second channel that notifies subscribers directly.
- Measure by topic, not by post. Jurka suggests looking at posting patterns by topic, format and timing. Once a month, tag your posts by topic in LinkedIn’s analytics export and compare the median impressions and the job titles that saw them.
For a SaaS company, the SaaS LinkedIn marketing guide covers how founder posts, the company page and ads work together. If the hard part is the hours, our LinkedIn ghostwriting service writes posts from short calls with you, and you approve each one.
What has not changed
Some things hold across all of LinkedIn’s statements. The Help Center describes the default feed as personalized by your interests and activity. Jurka says ranking does not use demographics such as age, race and gender. And the March 2026 posts say LinkedIn wants the feed to be “relevant to your interests, not a popularity contest”. Our reading: reach now depends more on what a post is about and who it suits than on how many people follow you.
We will re-check this page when LinkedIn publishes its next feed update. Last checked 11 November 2026.
Frequently asked
How does the LinkedIn algorithm work in 2026?
In two steps, according to LinkedIn's engineering blog. A retrieval system built on LLM embeddings picks candidate posts by matching what a post is about to what a member engages with, and a sequential ranking model then orders them using more than a thousand of the member's past interactions. LinkedIn says the inputs include profile details such as industry, skills and experience, plus what the member read, liked, commented on or scrolled past.
Is 360Brew the LinkedIn algorithm?
LinkedIn has not said so. 360Brew is a 150-billion-parameter research model described in a January 2025 arXiv paper that calls it a research pre-production model, and every version of that paper is now marked withdrawn on arXiv. LinkedIn's March 2026 announcements about the new feed describe LLM-based retrieval and Generative Recommender ranking without using the 360Brew name.
Does LinkedIn penalize AI-generated posts?
LinkedIn says it reduces the spread of generic AI content it calls AI slop. Its June 2026 pressroom post says such content is less likely to be distributed beyond a person's immediate network, and Tim Jurka wrote on 8 October 2026 that LinkedIn's classifiers now cover all posts distributed beyond the immediate network with 94% precision. LinkedIn also says using AI to help you write is fine if the post represents your own voice and perspective.
Does LinkedIn penalize engagement pods and automation?
Yes, by its own statements. LinkedIn's 12 March 2026 pressroom post says it is working to make engagement pods ineffective and to curb comment automation and third-party tools that create fake conversations. Its Help Center says third-party software and browser extensions that automate activity violate the User Agreement and can lead to account restriction.
How often should you post on LinkedIn?
LinkedIn does not publish a number. Buffer's analysis of more than 2 million LinkedIn posts found that posting 2 to 5 times a week improved impressions and engagement, and Richard van der Blom's 2026 report says the posting frequency he considers optimal fell from 5 to 6 a week to about 2 to 4. Pick a pace you can hold for months.
Why did my LinkedIn reach drop?
LinkedIn does not report reach changes for individual accounts, so no one can say for sure. Third-party studies point to a broad drop for frequent posters: van der Blom's 2026 report, based on 1.3 million posts from about 50,000 creators, says reach fell about 60% for active creators over two years. LinkedIn has also said it now shows more posts from outside your network and less engagement bait, which changes who sees what.



















