LinkedIn's 360Brew AI algorithm: what changed, and what LinkedIn actually confirmed
Contents
- What LinkedIn actually shipped, and what it never confirmed
- How the feed pipeline works now: retrieval, then ranking
- The signals that decide your reach now
- Why median reach fell by roughly half
- What gets suppressed now: pods, AI slop, and link posts
- The All In 360Brew Readiness Audit
- What I don't know
- FAQ
- Sources
I have been on LinkedIn since April 2004, the platform's first year. I sat through the B2B pivot, the algorithmic feed, and now this: a feed that reads your post the way a person would, before it decides who gets to see it. In January 2025, a LinkedIn research team led by Hamed Firooz published a paper describing 360Brew, a 150-billion-parameter model built to rank and recommend content by reasoning about it in plain language rather than counting likes. Fourteen months later, LinkedIn's own engineering blog confirmed a rebuilt feed pipeline, and never once used the name 360Brew. Both facts are true, they point in different directions, and almost nothing written about this topic mentions the second one.
In short: 360Brew is a 150-billion-parameter, decoder-only AI model that LinkedIn described in a January 2025 research paper as capable of handling more than 30 ranking and recommendation tasks by reasoning over text instead of counting engagement signals. LinkedIn's own engineering blog, published March 12, 2026, confirmed a full rebuild of the feed around a dual-encoder retrieval system and a model it calls the "Generative Recommender," without naming 360Brew once. Median reach on personal profiles fell roughly 50% year over year, according to AuthoredUp's analysis of more than 3 million posts, while saves, delayed comments, and dwell time now carry more weight than likes.
What LinkedIn actually shipped, and what it never confirmed
Start with what is on the record. On January 27, 2025, Hamed Firooz and 22 co-authors at LinkedIn published "360Brew: A Decoder-only Foundation Model for Personalized Ranking and Recommendation" on arXiv, paper number 2501.16450. The claim: a single 150-billion-parameter model, decoder-only, meaning it generates its output step by step rather than scoring a full text at once the way LinkedIn's older BERT-based systems did, built on the Mixtral 8x22 mixture-of-experts architecture (a design that activates only a fraction of those 150 billion parameters for any given request, to keep inference affordable). The model replaces the "dedicated models that have been developed and maintained over multiple years by teams of a similar or larger size" for more than 30 separate tasks across the platform. The paper reports results measured, in its own words, "based on offline metrics," a distinction worth sitting with: offline means benchmarked against historical data, not proven on live traffic in front of real members.
That paper was later withdrawn from arXiv. The listed reason has nothing to do with the science: administrators pulled it because "the submitter did not have the right to agree to the licence at the time of submission," a licensing dispute, not a retraction of the findings. I mention it because half the articles about 360Brew treat the paper as a live product announcement, and it was a research paper, pulled for a paperwork reason, describing a model tested offline.
Then, on March 12, 2026, LinkedIn engineer Hristo Danchev published "Engineering the next generation of LinkedIn's Feed" on the company's own engineering blog. It describes a genuine rebuild: a unified dual-encoder retrieval system replacing five separate pipelines, and a ranking model LinkedIn calls the Generative Recommender, built on a transformer with causal attention (the same family of architecture as most modern chatbots, adapted here to predict what you will engage with next rather than what word comes next). Read the post closely and one thing stands out by its absence: the word "360Brew" does not appear anywhere in it.
I am not accusing LinkedIn of hiding anything. Product names change between a research paper and a shipped feature all the time, and the architectures plausibly share lineage: both are large decoder-only models processing text about members and content. But plausible is not the same as confirmed, and I would rather tell you that plainly than repeat, as most competing articles do, that "360Brew now ranks your feed" as an established fact. What I can state as fact: LinkedIn rebuilt its ranking pipeline in March 2026 around the Generative Recommender, and it has never, in its own words, called that system 360Brew.
How the feed pipeline works now: retrieval, then ranking
Whatever you call the underlying model, the pipeline itself is well documented, and it works in two stages.
Stage one: retrieval. A companion paper, "Large Scale Retrieval for the LinkedIn Feed using Causal Language Models" (arXiv 2510.14223, submitted October 16, 2025, led by Sudarshan Srinivasa Ramanujam with more than 20 LinkedIn co-authors), describes a causal language model fine-tuned from Meta's LLaMA 3, working as a dual encoder. It generates embeddings, meaning numeric representations of what a text is about, for both members and posts, using text alone, and narrows a pool the paper describes as "hundreds of millions" of candidates down to roughly 2,000 per member. The latency budget: a few milliseconds, at a throughput of several thousand requests per second.
Stage two: ranking. The Generative Recommender then works through those 2,000 candidates. According to Danchev's March 2026 post, it processes more than a thousand of a member's past interactions as an ordered sequence rather than as isolated data points, which lets it read a trajectory: someone drifting from hiring-practice posts toward workforce-automation posts is treated as a signal about what to serve next, not two unrelated clicks. The model reports sub-50-millisecond retrieval latency across an index of millions of posts, and uses a mixture-of-experts prediction head to weigh passive signals (clicks, skips, long dwell) separately from active ones (likes, comments, shares).
The shift that matters for anyone publishing on the platform is architectural, not cosmetic. The old system scored a post as a bundle of numbers: click-through rate, early-engagement velocity, category tags. The new one reads the post, the author's profile, and the reader's history as connected text and reasons about fit. A LinkedIn engineering quote captures the design intent directly:
"What if we could replace this multi-source complexity with a single, unified retrieval system powered by LLM-generated embeddings?" Hristo Danchev, LinkedIn Engineering, March 12, 2026
That is not a marketing line. It is the actual question the team set out to answer, and the technical specifics back it up: hard negative sampling improved retrieval recall by 3.6%, percentile-bucketed numerical features improved the correlation between popularity and embedding similarity by a factor of 30, and a custom attention variant delivered a further 2x speedup over standard implementations. None of that changes how you write a post. All of it changes what the system is capable of noticing when you do.
The signals that decide your reach now
The practical consequence of reading text instead of counting clicks: some signals that used to matter a great deal now matter less, and one signal most people ignored now matters most.
According to AuthoredUp's analysis of more than 3 million LinkedIn posts published between March 2025 and February 2026, a save now drives five times more reach than a like, and twice as much as a standard comment. That ranking is consistent with how a reasoning model would treat the action: a like costs nothing and proves nothing, while a save is a reader deciding a piece of writing is worth finding again. The same dataset found that document-format posts capture 12.9% of all saves while making up only 4.9% of published content, a rate roughly 2.6 times their share.
Dwell time, meaning how long a member's screen actually holds on a post before scrolling past, works on the same logic. LinkedIn's own ranking research, the LiRank paper (arXiv 2402.06859, submitted February 2024 by a team led by Fedor Borisyuk), documents dwell time as a production ranking input across Feed, and reports a 0.5% lift in member sessions once it went live. The team's stated philosophy, from LinkedIn's original dwell-time engineering post: "time well spent is better than more time spent." Because the Generative Recommender weighs a member's most recent behavior heavily, an opening line that earns a few extra seconds of attention compounds into a meaningfully different distribution outcome, not just a nicer read.
Timing of engagement changed too. AuthoredUp and Botdog both describe delayed engagement, meaning saves and substantive comments arriving 24 to 72 hours after publication, as a stronger signal than a first-hour spike, associated with posts performing four to six times better in the following days. The interpretation lines up with the architecture: a post still earning attention two days later looks, to a model reading sequences of behavior, like something with lasting value rather than a five-minute spike engineered by a fast comment round.
| Signal | Old algorithm (pre-2024) | Feed under the Generative Recommender |
|---|---|---|
| Like | Counted directly, high weight | Counted, low weight on its own |
| Save | Minor signal | 5x the reach of a like (AuthoredUp) |
| Comment (short) | Counted as engagement | Near-neutral; length and timing matter more |
| First-hour engagement spike | Primary distribution trigger | One signal among several |
| Delayed engagement (24 to 72 hours) | Largely ignored | 4 to 6x performance association (AuthoredUp, Botdog) |
| Dwell time | Not tracked as a ranking input | Directly measured, production input since LiRank (LinkedIn, arXiv 2402.06859) |
Why median reach fell by roughly half
The most-cited number in every article on this topic is a reach decline, and it deserves the exact source attached to it rather than a rounded headline.
AuthoredUp's Trends 2026 report, drawn from more than 3 million posts published between March 2025 and February 2026 and compared against the prior year, found median reach for the top 5% of profiles falling from 13,711 impressions per post to 6,868, a decline of roughly 50%. Separately, Richard van der Blom's Algorithm Insights 2026 Report, built on an analysis of 1.8 million posts, put organic views down 50% and follower growth down 59% year over year, with average post reach settling at 8% to 12% of a creator's follower count, down from 15% to 20% a year earlier. Two independent datasets, two different methodologies, the same order of magnitude.
Read that number correctly before you panic over it. The Generative Recommender is not showing less content overall; LinkedIn still reports serving more than 1.3 billion members. It is distributing each individual post to a narrower, better-matched slice of that population instead of spraying it broadly and hoping. For a founder using LinkedIn to reach buyers rather than to collect views, 500 impressions among people who actually buy what you sell beats 2,000 impressions among people who never will, and the math on lead generation looks better than the raw reach number suggests.
Video and plain-text posts absorbed the sharpest hits. AuthoredUp's data shows video reach down 36% year over year, landing at 0.86 times a profile's median. Document-format posts, by contrast, carry a 1.39x reach multiplier and a 1.30x engagement multiplier over the same baseline, which tracks with the save-rate data above: a document is a save-shaped object.
What gets suppressed now: pods, AI slop, and link posts
Three practices that once worked, or looked like they worked, now cost you distribution under a model built to read rather than count.
Engagement pods. Groups of members who systematically like and comment on each other's posts leave a pattern the Generative Recommender can read directly in the text and timing of the interactions: reciprocal accounts, comments with no lexical connection to the same professional domain, engagement clustered in narrow time windows. LinkedIn's own public position, stated by VP of Product Management Gyanda Sachdeva, targets any third-party tool, browser extension, or plugin used to automate engagement, and the platform escalated enforcement in 2026 with broader detection and reach limits on flagged accounts, a fight I have tracked since it started on the All In blog.
Generic AI content. Forbes contributor Jodie Cook, writing in July 2026 about content moves LinkedIn started punishing this year, describes posts that read as templated AI output losing distribution because readers "scroll past it in under three seconds," which starves the post of the one signal the new pipeline weighs most: dwell time. AuthoredUp's own trends data corroborates the pattern directly: AI-flagged posts show roughly 30% less reach and 55% less engagement than comparable human-written posts in the same dataset. The mechanism is not a hidden AI detector bolted onto the ranking model; it is the same reasoning process applied to any text that reads as generic, whether a human or a machine wrote it.
Outbound links in the body of a post. The same Forbes piece puts posts carrying an external link at roughly 60% less reach than posts without one, a penalty that predates 360Brew by years but that the new pipeline appears to have kept, since a link click pulls a reader off-platform before dwell time can accumulate. If your team still drops a link in the body copy instead of the first comment, that single habit is worth fixing before anything else in this article.
The All In 360Brew Readiness Audit
Five dimensions, each tied to a mechanism described above, scored 0 to 5. Run it against your last ten posts.
Profile-to-content alignment. The retrieval model reads your headline, your About section, and your recent posts as one continuous text. A headline reading "VP of Sales" attached to posts about parenting advice creates a semantic mismatch the model can detect and that costs you distribution to a relevant audience.
Save-worthy density. Count the elements in your last ten posts that someone would plausibly want to find again: a framework, a comparison table, a checklist, a named benchmark. Below one per post on average, you are not generating the signal that now carries five times the weight of a like.
Opening-line specificity. Your first two lines get disproportionate weight because that is where dwell time either starts accumulating or does not. A specific number or a named situation beats a general statement every time; "vague" and "skipped in three seconds" are the same failure from two different angles.
Comment depth potential. Does the post take a position specific enough that someone could disagree with it, or does it stay safely neutral? Neutral posts draw polite one-word replies. Specific ones draw the substantive, delayed comments the Generative Recommender rewards.
Sector-specific vocabulary. A model reasoning over text picks up precise terminology as a signal of genuine expertise. Generic business language reads the same to the model whether a person or a template produced it, and it gets ranked accordingly.
| Dimension | Score /5 | Fix if below 3 |
|---|---|---|
| Profile-to-content alignment | _ /5 | Rewrite headline and About to match your actual post topics |
| Save-worthy density | _ /5 | Add one reference element per post: a framework, table, or checklist |
| Opening-line specificity | _ /5 | Replace generic openers with a specific number or named situation |
| Comment depth potential | _ /5 | End with a question that requires a real answer, not a reaction |
| Sector-specific vocabulary | _ /5 | Name your tools, your sources, your metrics |
| Total | _ /25 | Below 15: your content and your pipeline are misaligned |
What I don't know
I don't know for certain that the Generative Recommender described in March 2026 is 360Brew under a new name. The architectures line up: both are large decoder-only models reading text about members and content. LinkedIn has confirmed the pipeline and never confirmed the name, and I am not going to paper over that gap to make a cleaner story.
I don't know the exact false-positive rate of whatever suppresses posts that read as generic AI output, and neither does anyone writing about this topic outside LinkedIn; the company has not published one. AuthoredUp's 30%-less-reach, 55%-less-engagement figures describe a correlation drawn from its own dataset of published posts, not a controlled experiment isolating AI writing as the only variable, and I am citing it as the strongest available signal, not as proof of a specific detection mechanism.
I don't know how long the current reach baseline holds. LinkedIn rebuilt this pipeline once already, in under fifteen months, and Danchev's post describes embedding updates happening "within minutes" of a behavior change, which suggests a system built to keep moving rather than settle. Anything in this article describing a fixed multiplier is a snapshot of mid-2026, not a permanent constant.
FAQ
What is LinkedIn's 360Brew algorithm?
360Brew is a 150-billion-parameter, decoder-only AI model that LinkedIn's research team described in a January 2025 paper (arXiv 2501.16450) as capable of handling more than 30 ranking and recommendation tasks through text-based reasoning rather than counting engagement signals. The paper was later withdrawn from arXiv for licensing reasons, unrelated to its findings.
Is 360Brew still the model ranking the LinkedIn feed in 2026?
Not confirmed by LinkedIn. The company's own engineering blog post from March 12, 2026, describes a rebuilt pipeline around a dual-encoder retrieval system and a ranking model it calls the Generative Recommender, and never uses the name 360Brew. The architectures are plausibly related; LinkedIn has not stated they are the same system.
Why has my LinkedIn reach dropped so much in 2026?
AuthoredUp's analysis of over 3 million posts found median reach for top-performing profiles down roughly 50% year over year, and Richard van der Blom's Algorithm Insights 2026 Report independently found organic views down 50% and follower growth down 59%. The pipeline now distributes each post to a narrower, better-matched audience instead of spraying it broadly, which lowers raw impressions while often improving relevance.
Do saves really matter more than likes on LinkedIn?
Yes, according to AuthoredUp's dataset of more than 3 million posts: a save generates five times the reach of a like and twice that of a standard comment. Document-format posts, which naturally invite saving, carry a 1.39x reach multiplier over the platform median in the same analysis.
Does LinkedIn's algorithm detect AI-generated content?
Not through a labeled detector that LinkedIn has described publicly. AuthoredUp's data shows posts flagged as AI-generated getting roughly 30% less reach and 55% less engagement than comparable human-written posts, a pattern consistent with a model that reads generic phrasing as low-value regardless of who or what produced it, rather than a system specifically hunting for AI authorship.
What's the difference between LinkedIn's old algorithm and the Generative Recommender?
The old system scored posts as numbers fed through five separate specialized pipelines, one for the network feed, another for job suggestions, and so on, each hand-built with its own thresholds. The Generative Recommender, confirmed by LinkedIn in March 2026, reads a member's text history as an ordered sequence and a post's text directly, replacing that patchwork with language-based reasoning across a single unified system.
Sources
- Firooz et al., "360Brew: A Decoder-only Foundation Model for Personalized Ranking and Recommendation," arXiv 2501.16450, submitted January 27, 2025, later withdrawn for licensing reasons.
- Srinivasa Ramanujam et al., "Large Scale Retrieval for the LinkedIn Feed using Causal Language Models," arXiv 2510.14223, submitted October 16, 2025.
- Hristo Danchev, "Engineering the next generation of LinkedIn's Feed," LinkedIn Engineering Blog, March 12, 2026. Corroborated by ppc.land and Netinfluencer.
- AuthoredUp, "LinkedIn Trends in 2026: What the Data Shows", analysis of 3M+ posts, March 2025 to February 2026.
- AuthoredUp, "LinkedIn 360Brew: What Actually Changed", analysis of 621,000+ posts, June 2024 to May 2025.
- Richard van der Blom, Algorithm Insights 2026 Report, analysis of 1.8 million posts, cited across multiple 2026 industry write-ups including UpGrowth.
- LinkedIn Engineering, "LiRank: Industrial Large Scale Ranking Models at LinkedIn," arXiv 2402.06859, February 2024; and Siddharth Dangi et al., "Understanding dwell time to improve LinkedIn feed ranking," LinkedIn Engineering Blog.
- Botdog, "5 Biggest LinkedIn Algorithm Changes in 2026: What Salespeople Need to Know".
- Jodie Cook, "5 LinkedIn Content Moves LinkedIn Started Punishing In 2026," Forbes, July 23, 2026.
- Social Media Today, "LinkedIn Vows to Take Action Against Engagement Pods, Fake Engagement", Gyanda Sachdeva statement.
- You Creatives, "What LinkedIn Actually Changed in Its Feed", independent analysis noting the absence of the 360Brew name in LinkedIn's official March 2026 post.
All In: reading LinkedIn's engineering notes so you don't have to guess
Most articles on 360Brew repeat a research paper as if it were a live product announcement. This one reads LinkedIn's own engineering blog against that paper and reports the gap between the two, because a leader deciding where to invest content effort deserves the difference between confirmed and assumed.
All In is the B2B media that decodes LinkedIn, expert blog, weekly podcast and newsletter for SME leaders and sales directors who want to turn LinkedIn into measurable growth. An original creation by Patrick de Carvalho, on LinkedIn since 2004. Motto: "I Never Lose."
Discover All In and get what matters on LinkedIn, every week.