What is 360Brew? What defines your LinkedIn feed
Partner · Strategy, advertising and measurement
360Brew is a 150-billion-parameter artificial intelligence model that LinkedIn's research teams described in January 2025 for ranking and recommending content. LinkedIn tested it with a small group of members, then shut the test down: since April 14, 2026, the company has said plainly that 360Brew is not part of its ranking system. What ranks your feed since March 2026 is a different system, documented by LinkedIn, which also reads your text and your history. For your company the consequence is the same: what used to be won through engagement is now won through the precision of what you say.
- 360Brew is a research model that LinkedIn tested in 2025 and then abandoned. It never replaced the feed's algorithm.
- The system in service since March 2026 finds posts with a language model, then ranks them by reading more than 1,000 past interactions of each member.
- The median reach of a post fell 47% between June 2024 and May 2025, for 95% of members.
- LinkedIn has said it wants to make engagement pods "entirely ineffective" and to reduce posts that fish for reactions.
- Content that speaks precisely to a precise audience is served better than broad content, because the system reads the text, not just the reactions.
- The budget and time that went to reach tactics move toward what your company genuinely has that is specific to say.
On this page
Foundation model
A foundation model is a large model trained on highly varied data, able to perform many tasks without being retrained for each one. It stands in contrast to specialized models, trained separately for one specific task. Applied to recommendation, it lets a single system handle the ranking of several surfaces of a platform, reasoning about context rather than applying rules learned surface by surface.
What is 360Brew?
360Brew is an artificial intelligence model that LinkedIn's research teams described in January 2025. Built on the open model Mixtral 8x22B, it has 150 billion parameters and reads, as text, a member's profile, their history and the content to evaluate, in order to predict their interest. LinkedIn tested it with a small group of members, then abandoned it.
The research paper, signed by Hamed Firooz and some twenty colleagues, presents 360Brew as a "research pre-production model." It covers more than 30 prediction tasks on at least 8 surfaces of the platform (posts, job postings and connection suggestions are named), with a single method: everything is written as text. Instead of receiving hundreds of numeric signals, the model receives an instruction, the member's profile spelled out, the list of what they viewed, liked or ignored, then the item to judge. It answers the way it would answer a question.
That is what struck the industry: one model that reads, in place of dozens of models that count. The authors report results "comparable to or exceeding" production systems, measured offline, on recorded data. The clearest gap appears on members with a short history: with five known interactions instead of a hundred, the model that reads keeps the edge.
Two facts got lost along the way. The first is that "measured offline" means the model was never compared live, on real members. The second is that the paper was withdrawn from arXiv, every one of its four versions, because the author did not have the right to agree to the licence at the time of submission. The texts still in circulation cite it as if it described today's LinkedIn.
LinkedIn does not use 360Brew to rank your feed.
On April 14, 2026, Tim Jurka, LinkedIn's Vice President of Engineering, asked the question himself in a post: "Is 360Brew used as part of your ranking system? Short answer: No!" He explains that LinkedIn runs hundreds of tests at any given time, that 360Brew was tried in early 2025 with a small group of members and that the team concluded it was not the right fit for the platform, then shut the test down.
A month earlier, on March 12, 2026, LinkedIn had published two texts on the overhaul of its feed: an announcement to members signed by the same Tim Jurka and a technical article on the engineering blog signed by Hristo Danchev. Neither one says the name 360Brew.
Why do so many articles say the opposite? Because the research paper came out in January 2025, reach collapsed the same year, and the two facts fused in the texts of trainers and agencies. When reach drops, the name of an algorithm serves as the explanation. LinkedIn took fifteen months to deny it. The denial was picked up far less than the rumour.
What matters for your company is not the name of the model. It is that the real system, the one LinkedIn documents, rests on the same idea as 360Brew: read a member's text and history rather than count their reactions. The practical consequences attributed to 360Brew remain true, for a different reason.
What ranks your feed since March 2026
LinkedIn's feed works in two stages, like most recommendation systems. A first system finds, among hundreds of millions of posts, a few thousand candidates for a member. A second one ranks them. LinkedIn rebuilt both.
Retrieval reads the text. Since a technical paper of October 2025, confirmed by the March 2026 article, retrieval rests on a language model (Meta's Llama 3, fine-tuned by LinkedIn) that turns the text of a post and the text describing a member into an "embedding," then brings the two together. This system replaces several separate sources (trending posts, the network, collaborative filters) with a single one. LinkedIn draws a clear promise to members from it: discovering content from authors you don't follow, on topics you care about, even when the words differ.
Ranking reads the history. The ranker, which LinkedIn calls the "Generative Recommender," treats a member's history as an ordered sequence of more than 1,000 past interactions (what they read at length, liked, commented on, shared) and predicts what they will do with each candidate post: read it at length, skip it, like it, comment on it, share it. The announcement to members adds that the profile (industry, skills) enters this calculation and that the feed adapts within a session, as interests shift.
A LinkedIn research paper from February 2026 gives the measure of the change: this sequential ranker, in service for the majority of feed traffic for more than three months across 1.2 billion members, gained 2.10% in time spent and 3.52% in likes, comments and shares compared with the previous model. The same paper notes that rankers built directly on a language model, in the line of 360Brew, did not give the best result live: the sequential model won on results and on computing cost alike.
| Element | Before 2025 | Since March 2026 |
|---|---|---|
| Finding posts | Several separate sources (trends, network, collaborative filters) | A single language model that reads the text of posts and members |
| Ranking | A model that weighs hundreds of numeric signals | A model that reads the sequence of the member's last 1,000 interactions |
| What is predicted | Likely reactions | Long reads, reactions, comments, shares |
| A niche topic | Few signals, so little distributed | Matched to the members the topic interests, even without volume |
| The member's profile | Approximated through segments | Read as text, with industry and skills |
| Time to appear | Hours | Minutes, according to LinkedIn |
Why your reach dropped
The drop is real and measured. AuthoredUp, a writing tool for LinkedIn, analyzed 3 million posts from more than 30,000 personal profiles: the median reach of a post went from 1,211 people in June 2024 to 636 at the end of May 2025, a drop of more than 47% in one year, for 95% of members. Every format is down, video the most (72%), then images (45%), documents (43%) and text (34%).
This drop does not come from 360Brew, which never ranked the feed. It comes from a choice by LinkedIn, stated in the March 2026 announcement: serve each post to the members for whom it is relevant, rather than to the greatest number. A broad post, addressed to everyone, used to reach a lot of people through inertia. A system that reads the text ties it to few people, because it speaks precisely to no one.
Company pages start from lower down. Metricool's 2026 study, on 673,658 posts from nearly 65,000 accounts, measures an engagement rate of 2.60% on personal profiles against 1.74% on pages. The reach gaps reported by tool vendors go well beyond that, but with no published method: we do not repeat them. What we see in our engagements points the same way. It is a distribution mechanic, independent of the ranking model: people follow people.
What to know before the call
A post lives one to two days on this platform, from what we observe in our engagements. Any strategy that rests on the reach of a single message is therefore paid for again and again, accumulating nothing. What changes with the system in service is that artificial reach tactics cost the same as before and return markedly less. The budget moves toward substance, which cannot be bought.
- Who posts for us, a personal account or the company page? Why was that choice made?
- How many real leads did this channel produce last quarter?
- Which precise audience do our posts address? Could we name it in one sentence?
- Are we still paying for reach, engagement or follower services?
- What do we post that no one else in our sector could post?
The useful answer names a precise audience and an identifiable personal account. A weak answer talks about reach, impressions, engagement rate, or proposes raising the posting cadence.
Deciding whether this channel deserves the time of one of your executives (and in what form) is an annual trade-off. A 90-minute consultation settles it from your actual market, with a written summary that circulates through your organization.
The tactics that lose their effect
They all rested on the same idea: produce engagement signals for a system that read nothing but signals. LinkedIn now goes after them in two ways, through ranking and through the rules.
On November 6, 2025, Gyanda Sachdeva, LinkedIn's Vice President of Product Management, told Social Media Today: "Our goal is to make engagement pods entirely ineffective." LinkedIn is increasing the ways it detects them, limiting the reach of their posts and reminding members that these practices breach its terms of service. The March 2026 announcement goes further: third-party software and browser extensions that automate comments are not allowed on LinkedIn. LinkedIn is also reducing repetitive, click-driven posts, with specific examples: the "Comment 'Yes' if you agree" posts, videos unrelated to the text that accompanies them and recycled thought-leadership posts.
| Tactic | Why it worked | Why it weakens |
|---|---|---|
| Mutual engagement pods | Generated fast interactions at low cost | LinkedIn detects them, limits their reach and bans them in its terms |
| A closed question at the end of a post | Prompted short comments | LinkedIn has been reducing these posts, by name, since March 2026 |
| A video attached to an unrelated text | Rode on the format's reach | No longer gains additional reach, according to LinkedIn |
| A sensational hook | Raised the stop rate | The gap between the hook and the content is legible to a system that reads both |
| Posting at the optimal time | A real but small effect | Stays marginal: posts now appear within minutes |
| Recycling the same message | Caught those who missed it | Still useful if the text says something, targeted if it is generic |
Artificial engagement tactics do not become useless overnight. They become less profitable, and above all riskier, because LinkedIn has said publicly that it is looking for them, and because a system able to read the message spots the gap between what is promised and what is delivered more easily. The cost of these tactics has not changed; their return has collapsed.
What stays in-house is the account itself and the voice behind it. A company page does not replace an identifiable executive or specialist. No provider can borrow that voice credibly over time. What gets delegated is the preparation: spotting the topics moving through the sector, formatting, the calendar, watching what competitors publish and reading the distribution data. The split to aim for is thirty minutes a week from the person who speaks, everything else outside.
What to do concretely
The recommendations that follow stem from what LinkedIn published about its feed in 2025 and 2026, not from an official visibility guideline: LinkedIn publishes none. None requires a budget; all require time from a person who knows the trade.
Treat the profile as a text the system reads
Retrieval turns the text describing a member into an embedding. Ranking takes industry and skills into account. A profile headline that states a trade and a clientele ("Accountant for dental clinics in Quebec") gives the system what it needs to match your posts to the right people. A headline for effect ("I grow ambitions") gives it nothing.
Post from profiles, not from the page
The gap between a personal account and a company page stays sharper on LinkedIn than on any other platform. It is a distribution mechanic, independent of the ranking model. A budget spent keeping a company page alive therefore buys less visibility than the time of an executive.
Write for a precise audience
Content that speaks specifically to plant managers at companies under two hundred employees is easier to tie to a relevance than content addressed to managers in general. It is also the only kind that produces leads rather than congratulations.
Say something verifiable
A figure from your own engagements, a client case, an observation drawn from your practice. That is what LinkedIn calls actionable insights and a thoughtful perspective in its announcement. It is also what no competitor can copy.
Keep the promise of the hook
The gap between the first line and the actual content has become the most costly of shortcuts, because that gap is legible to a system that understands both, and because LinkedIn is reducing click-driven posts by name.
Reply to comments within the hour
The ranker predicts long reads, comments and shares: a real exchange under the post feeds all three. Posting and leaving remains the simplest way to waste the only real cost of this channel, the time of the person who writes.
A system that reads the message rewards those who have one. That is bad news for tactics and good news for those who know what they are talking about.
Falia analysis gridThe seven points that follow cost nothing to fix and are decided in one meeting.
What you can adjust right now
The full LinkedIn strategy, organic and paid, is covered in LinkedIn, when the high cost is justified. The role of organic presence is detailed in organic social media. The logic of recommendation systems applied to advertising is explained in the Andromeda update.
The overall view of AI visibility is in who measures what, and who measures nothing.
Adapting to AI recommendation systems is at the heart of the Strengthen your visibility in AI answers goal.
Already running a marketing team? See how we plug in as reinforcement on LinkedIn.
Frequently asked questions about 360Brew
What is 360Brew?
It is a 150-billion-parameter artificial intelligence model described by LinkedIn's research teams in January 2025, built on Mixtral 8x22B. It reads, as text, a member's profile, their history and the content to evaluate, in order to predict their interest. LinkedIn tested it with a small group of members, then shut the test down.
Does LinkedIn use 360Brew to rank my feed?
No. On April 14, 2026, Tim Jurka, LinkedIn's Vice President of Engineering, answered publicly: "Short answer: No!" The model was tried in early 2025 and judged not the right fit for the platform. The texts that still present it as the algorithm in service are out of date.
Then why does everyone talk about it?
Because the research paper came out in January 2025 and the reach of posts collapsed the same year: the two facts fused in the texts of trainers and agencies. LinkedIn's denial, fifteen months later, was picked up far less than the rumour.
What ranks my feed now?
Two systems documented by LinkedIn in March 2026: a retrieval stage built on a language model, which matches the text of posts to the text describing each member, then a ranker that reads the sequence of more than 1,000 past interactions of the member and predicts whether they will read, like, comment on or share each post.
Why did my reach drop?
According to AuthoredUp, the median reach of a post went from 1,211 people in June 2024 to 636 in May 2025, a 47% drop for 95% of members. LinkedIn serves each post to the members for whom it is relevant rather than to the greatest number: a broad post reaches fewer people than before, a precise post reaches the right ones.
Do engagement pods still work?
Less and less, with added risk. LinkedIn said in November 2025 that it wants to make them "entirely ineffective": it detects them, limits the reach of their posts and reminds members that they breach its terms of service. Tools that automate comments are not allowed.
Should you post from your company page or your profile?
From profiles, relayed by the page. Metricool's 2026 study measures an engagement rate of 2.60% on personal profiles against 1.74% on pages. The reach gap is sharper still. It is a distribution mechanic: people follow people.
Does LinkedIn publish visibility guidelines?
No. LinkedIn documents the architecture of its feed and says what it is reducing (posts that fish for reactions, videos unrelated to their text, automation), but does not publish the weighting of its ranking. The recommendations in this article stem from those publications, not from an official guideline.
How long does a LinkedIn post live?
One to two days in practice, from what we observe in our engagements. That is longer than on chronological feeds, where the useful span is counted in hours, but markedly less than on discovery platforms, where a piece of content can circulate for weeks.
- Hamed Firooz et al. (LinkedIn), 360Brew: A Decoder-only Foundation Model for Personalized Ranking and Recommendation, arXiv, submitted January 27, 2025, since withdrawn over a licensing issue. Accessed November 2026.
- Tim Jurka, Vice President of Engineering at LinkedIn, LinkedIn post of April 14, 2026: "Is 360Brew used as part of your ranking system? Short answer: No!" Accessed November 2026.
- Tim Jurka, Updates to The LinkedIn Feed Focusing on Authentic, Relevant Conversations, LinkedIn, March 12, 2026. Accessed November 2026.
- Hristo Danchev, Engineering the next generation of LinkedIn's Feed, LinkedIn Engineering Blog, March 12, 2026. Accessed November 2026.
- Lars Hertel, Gaurav Srivastava et al. (LinkedIn), An Industrial-Scale Sequential Recommender for LinkedIn Feed Ranking, arXiv, February 2026. Accessed November 2026.
- Srinivasa Ramanujam et al. (LinkedIn), Large Scale Retrieval for the LinkedIn Feed using Causal Language Models, arXiv, October 16, 2025. Accessed November 2026.
- AuthoredUp, Vanishing Audience, newsletter of July 9, 2025: 3 million posts from more than 30,000 profiles. Accessed November 2026.
- Andrew Hutchinson, LinkedIn Vows to Take Action Against Engagement Pods, Social Media Today, November 6, 2025 (statement by Gyanda Sachdeva). Accessed November 2026.
- Metricool, LinkedIn study 2026: 673,658 posts from nearly 65,000 accounts. Accessed November 2026.
- Post lifespans: engagement observation by Falia, revised in November 2026.
Gabriel almost always takes your first call and carries out your audit. He builds the strategy starting from your growth goal: where to put your budget, which market to test and how to connect each lead to a real sale in your CRM. He mainly leads engagements for three goals: Optimize the profitability of your digital campaigns, Develop a new market, and Generate demand and growth. With Geneviève, he also works on organic search (SEO), AI visibility (GEO) and conversion rate optimization (CRO). The sales a Google Ads or Meta Ads campaign brings in depend on the page that receives the click. He writes mainly about marketing strategy, paid advertising and measurement.
About Falia →
Strategy
SEO
Paid advertising
Conversion
AI visibility
Web design