What is LinkedIn's 360Brew algorithm?
360Brew is a research model described in a paper, and LinkedIn's own feed announcements don't say it runs your feed. The paper, by Hamed Firooz and 22 co-authors, was first submitted to arXiv on 27 January 2025. It calls 360Brew V1.0 a "research pre-production model": a decoder-only model with 150 billion parameters, trained and fine-tuned on LinkedIn's data and tasks. The authors report that it handled more than 30 predictive tasks across LinkedIn and matched or beat the production systems on offline metrics.
Dre Olgiati, one of the co-authors, announced the paper on LinkedIn on 29 January 2025. Two details on the arXiv page change how much weight the name deserves:
- It is withdrawn. All four versions (27 January, 1 February, 7 February and 23 August 2025) are marked withdrawn. The arXiv admin note says the submitter did not have the right to agree to the licence at the time of submission. The PDF can no longer be downloaded there.
- The results are offline. The abstract compares 360Brew with production systems on offline metrics. It does not report an A/B test or say the model was deployed.
So when a vendor post says LinkedIn replaced its feed algorithm with 360Brew, it is using a research name for a feed LinkedIn describes in other terms. Those terms are more useful, because they say what the system actually reads.
What did LinkedIn actually announce in March 2026?
On 12 March 2026 LinkedIn published two posts about a new feed. One was an engineering post by Hristo Danchev. The other was a member update from Tim Jurka, who works on LinkedIn's feed. Neither uses the name 360Brew. They describe two changes.
- Retrieval, the step that picks candidate posts. LinkedIn used to pull candidates from separate sources: a chronological index of your network's activity, trending posts in your geography, collaborative filtering, industry trending content and several embedding systems. It replaced them with one system built on LLM-generated embeddings of members and posts.
- Ranking, the step that orders them. A Generative Recommender treats your feed history as a sequence and processes more than a thousand of your past interactions, instead of scoring each post on its own.
The ranking model has its own paper. Feed SR, submitted to arXiv on 12 February 2026 by LinkedIn authors, replaced a DCNv2-based ranker and had served the majority of feed traffic for over three months. In online A/B tests it raised time spent by 2.10% and likes, comments or reshares by 3.52%. The same paper says LinkedIn first tried an LLM-Ranker that read every post feature as text, starting from a pretrained model such as Llama 3. It "never achieved superior online performance" over the existing production model, partly because the strength of network relationships was hard to encode in a text prompt. Two of the 360Brew co-authors, Adrian Englhardt and Souvik Ghosh, are also authors of Feed SR, which does not mention 360Brew.
| Claim you will read | What LinkedIn published (checked 24 Oct 2026) | Our verdict |
|---|---|---|
| 360Brew is the LinkedIn algorithm | Not named in either 12 March 2026 post; the paper is withdrawn | Unconfirmed |
| AI reads what your post is about | LLM embeddings now drive retrieval; Jurka says LLMs better understand what a post is actually about | Confirmed |
| Your profile feeds into ranking | Retrieval inputs include the author's name, headline, company and industry; Jurka (Aug 2025) says ranking factors in the author's professional identity | Confirmed |
| Changing topic gets you suppressed | No such rule in these posts or papers | Unconfirmed |
| Engagement pods are targeted | Jurka says LinkedIn is working to make pods ineffective and curb comment automation | Confirmed |
| Engagement bait and recycled posts lose reach | Jurka says the feed will show less of both over the next few months | Stated plan |
Why does sticking to one topic matter now?
Because the new feed describes each post partly by who wrote it, then matches it to readers by meaning. The engineering post lists what goes into a post's description for retrieval: format, author information (name, headline, company, industry), engagement counts, article metadata and the post text. A member's description includes profile, skills, work history, education and a chronological list of posts they engaged with. LinkedIn found the model learned better when that list held only posts the member engaged with, and it trains on posts that were shown but got no engagement as hard negatives.
Tim Jurka, who has worked on the feed for over a decade, made the author side explicit in an August 2025 post on how the feed works. The ranking system factors in the author's professional identity, such as experience, content shared on their profile and skills, and does not use age, race or gender. Jurka also wrote that each post finds its own path to an audience, and suggested looking at your posting patterns by topic, format and timing.
Here is our reading, which LinkedIn has not stated in these words. When your headline says "pricing for B2B SaaS" and your post is about pricing, the two inputs agree, and the readers whose history is full of pricing posts are the obvious match. When the same author posts about crypto on Monday and parenting on Thursday, each post is described with a headline that fits neither. LinkedIn has not published how much author identity weighs, and no source here describes a penalty for changing topic. The gain from focus is a clearer match, and that is reason enough.
What does the 360Brew hype get wrong?
It states as fact things LinkedIn has not published. Valley's blog says that LinkedIn replaced its feed algorithm with 360Brew, that the model takes roughly 90 days to map your expertise, and that a growth marketer posting about crypto gets suppressed. Expandi's LinkedIn algorithm guide says that 360Brew now handles feed ranking, job recommendations and connection suggestions, and reached 40 to 100% of platform surfaces by autumn 2025. We could not find the 90-day figure, the suppression rule or the rollout share in the LinkedIn posts and papers linked in this article.
To be fair to both, their practical advice is close to ours: narrow your themes, write with substance, skip engagement bait. The trouble is the certainty. If a 90-day clock were real, you would plan a topic change very differently from how you would plan one under a system that matches each post on its own. upGrowth ran a similar check against the primary sources and also found that the March 2026 posts never use the name.
The changes LinkedIn did announce matter more for most founders: less engagement bait such as asking readers to comment Yes if they agree, fewer videos unrelated to their text, fewer recycled thought-leadership posts, and work against pods and comment automation, including third-party software and browser extensions, which Jurka says are not allowed on LinkedIn. Our guide to what LinkedIn allows and what gets accounts restricted covers that risk. Our post on what still gets reach after LinkedIn's 2026 changes to AI-generated posts covers that side.
How do you pick your two or three topics?
Pick topics your buyer pays to solve and you can write about from your own work for months. Run each candidate through four tests:
| Test | Question | Fails when |
|---|---|---|
| Buyer | Would a person who signs your contracts read this? | Only peers in your own job care |
| Material | Can you list 20 posts from calls, numbers and decisions you have? | You run dry after five |
| Headline | Can your headline name it in four words? | The headline needs an "and also" |
| Proof | Have you done the work you're writing about? | Every post is opinion with no example |
An example for a founder selling invoice automation to agencies: (1) getting paid on time as an agency, (2) finance operations for 10 to 50 person firms, (3) building the company, with honest numbers. The third is the one to keep smallest, because it is closest to your own interests and furthest from the buyer's. Rotating formats inside a topic keeps it fresh without changing subject: a lesson learned on Monday and a customer story on Thursday can both be about late invoices.
If you get stuck for ideas inside a narrow topic, our guide to what a founder can post when there seems to be nothing to say mines sales calls and your inbox for material.
A 30-minute topic audit using LinkedIn's own analytics
Check whether your topics already work before you change anything. LinkedIn's analytics give you enough to do it in half an hour.
Export your recent posts
In creator analytics, open Content analytics, scroll to Most recent posts performance, click Show all posts, then Load more, and export the table as a CSV file. LinkedIn says this table is desktop-only and still rolling out, so you may not have it yet. If not, copy impressions and engagements from each post by hand.
Tag each post with one topic
Use one column and at most four labels, plus "other". If a post fits two, pick the one in its first line.
Compare medians, not totals
Take the median impressions per topic. One viral post skews an average. Illustrative numbers, not real data: 9 pricing posts with a median of 2,100 impressions, 6 hiring posts at 1,400 and 9 "other" posts at 650. That account should post more on pricing and stop most of the "other".
Check who saw them
Open post viewer demographics on your top posts per topic. Per LinkedIn Help, demographic breakdowns stay available for 180 days and don't exist for posts shared only with connections or in Groups. LinkedIn also adds a small amount of noise and shows each category as a percentage of members reached, so read the shares as rough.
Keep a topic only if it reaches people and the right people. A topic with big numbers and the wrong audience is a hobby.
Make your profile and your posts say the same thing
Your headline is one of the author fields LinkedIn's engineering post lists for describing each post, so make it name your main topic. Then line up the rest:
- Headline: who you help and with what, in plain words. Our LinkedIn headline examples show the pattern.
- About: two or three sentences on the problem, then proof from your own work.
- Experience and skills: list the skills behind your topics. Jurka names experience, profile content and skills as part of the author's professional identity.
- Pace: pick a pace you can hold for months. Our guide to how often a busy founder should post on LinkedIn covers what the data says.
Siela's Social Media agent can keep the pace for you. Its LinkedIn Presence agent posts in your voice at 2, 3 or 5 posts a week, in your active hours, in formats such as lessons learned, customer stories, industry takes, behind the scenes and how-to lists. A second AI reviewer checks every post. If you choose "show me every post first", you can turn down any draft that drifts from your two or three topics before it goes out.
Frequently asked questions
Is 360Brew live on LinkedIn?
Not in any LinkedIn post or paper we checked. 360Brew is a 150-billion-parameter research model from a January 2025 paper, now withdrawn from arXiv. LinkedIn's 12 March 2026 posts about its new feed describe LLM-based retrieval and a Generative Recommender ranking model without using the name. Treat claims that 360Brew decides your reach as unconfirmed.
Why was the 360Brew paper withdrawn from arXiv?
The arXiv page says the version was removed by arXiv administrators because the submitter did not have the right to agree to the licence at the time of submission. All four versions, from 27 January to 23 August 2025, are marked withdrawn. The page gives no other reason, and the abstract is still readable.
Does LinkedIn read my headline when it ranks my posts?
At the retrieval step, yes. LinkedIn's March 2026 engineering post says each post's description for retrieval, the step that picks candidate posts, includes the author's name, headline, company and industry, along with the post text. Tim Jurka wrote in August 2025 that ranking factors in the author's professional identity, such as experience, profile content and skills. LinkedIn has not said how much weight the headline carries.
Will posting about a new topic hurt my LinkedIn reach?
None of the LinkedIn posts and papers cited in this article says so. Claims of a 90-day expertise clock or suppression for off-profile topics come from vendor guides, and we found neither in those LinkedIn sources. LinkedIn does match posts to readers by meaning, so a new topic may reach a different audience at first. Update your headline if the new topic becomes your main one.
How do I see which of my LinkedIn topics work?
Export your posts from creator analytics, tag each with one topic and compare median impressions per topic. Then check post viewer demographics on the best posts per topic to see whether the right job titles saw them. LinkedIn keeps demographic breakdowns for 180 days and adds a little noise to them, so treat the percentages as rough.
Sources
- arXiv: 360Brew: A Decoder-only Foundation Model for Personalized Ranking and Recommendation (withdrawn) (first submitted 27 Jan 2025)
- Dre Olgiati on LinkedIn: 360Brew paper announcement (29 Jan 2025)
- LinkedIn Engineering: Engineering the next generation of LinkedIn's Feed (12 Mar 2026)
- Tim Jurka: Updates to the LinkedIn Feed focusing on authentic, relevant conversations (12 Mar 2026)
- arXiv: An Industrial-Scale Sequential Recommender for LinkedIn Feed Ranking (Feed SR) (submitted 12 Feb 2026)
- Tim Jurka: How does the LinkedIn Feed work? (11 Aug 2025)
- Valley: LinkedIn's new 360Brew AI algorithm (checked 24 Oct 2026)
- Expandi: LinkedIn algorithm guide (checked 24 Oct 2026)
- upGrowth: LinkedIn algorithm 2026, March update and 360Brew fact check (checked 24 Oct 2026)
- LinkedIn Help: View your creator analytics (checked 24 Oct 2026)
- LinkedIn Help: View post analytics for your content (checked 24 Oct 2026)
- LinkedIn Help: Recent changes to viewers demographics (checked 24 Oct 2026)