That document has been freely available for five months. What the market discusses is a different text: an engineering blog article published a month later, one sentence of which, from the second paragraph, travelled across the web. The sentence is accurate. In isolation, it suggests the feed is ranked by a language model. The same article explains further down that this model operates before ranking, not during it.
We read both documents in full, plus two earlier publications. This article reports what they describe and flags what they leave unsaid. It sets aside advertising, company pages and job recommendations, which belong to separate systems.
What actually ranks the LinkedIn feed?
Two stages, not one. Retrieval narrows millions of posts down to a few hundred candidates. Ranking then orders them. Conflate the two and everything that follows goes wrong.
Retrieval does rely on a language model, fine-tuned for the purpose and built as a dual encoder: the member and the post are each turned into comparable descriptions, which allows a reader to be matched with content on meaning rather than on shared keywords.
Ranking relies on an entirely different model. The research paper calls it Feed SR, the engineering blog calls it Generative Recommender. It is a causal-attention sequential transformer: it reads your interactions in chronological order, never anticipating what comes next. It is not a language model.
The research paper devotes a full section to the architectures that were considered and dropped, which is rare enough to be worth noting. The team did build a ranker based on a language model, then tested it in live conditions. Three obstacles are documented. The values that matter for ranking a post, reaction counts, age, closeness to the author, translate poorly into sentences. Then cost: each post consumed hundreds of text units, against two in Feed SR. Performance settled the rest, since that ranker did not outperform the model already in service on online metrics.
One caveat applies to everything that follows. LinkedIn describes a result obtained on a given date, not a doctrine. The company has announced no abandonment of that path, and nothing indicates where it stands today.
What is the ranking system trying to trigger?
This is the most directly usable information in either document, and the least reported.
The model predicts six actions. Three passive: the click, the scroll-past, and Long Dwell, meaning time spent on a post beyond a threshold that depends on its type. Three active: the reaction, the comment, the share. The research paper groups these last three under the name Contribution. The two metrics on which LinkedIn reports its gains are Long Dwell and Contribution.
Its memory is bounded: one thousand viewed posts, drawn from roughly a year. It also decays, through two mechanisms the paper separates. Training data loses half its influence every sixty days. Within a single member’s sequence, older interactions count about half as much as recent ones.
One less expected signal deserves a mention: how previous readers distribute across reading-time bands, from zero to five seconds up to more than sixty, improves Long Dwell prediction by 2.5 points on its own.
The live experiment reports 2.10% more time spent, with the largest gains among the most active members and no measurable effect on new ones.
A common claim holds that saving a post has become the primary ranking factor. The paper does not list that action among the optimised objectives. This does not prove it carries no weight. It means the claim rests on no primary source, while the other six are named one by one.
How did a research model become the name of the algorithm?
360Brew exists. It is a research paper filed on arXiv in January 2025 by a LinkedIn team, describing a 150-billion-parameter model capable in principle of handling feed ranking, job recommendations and connection suggestions in a single system.
Three facts, and nothing beyond them.
Its abstract describes it as a research pre-production model.
We found no confirmation of deployment by LinkedIn.
The March 2026 engineering blog article does not mention the name.
In January 2026, Forbes published an interview in which a creator followed by 1.2 million people stated that LinkedIn’s latest update was called 360 Brew. Within weeks the claim had become settled fact, and dozens of publications built detailed tactical recommendations on top of it.
One detail measures how far the chain travelled from its source. Several articles state that 360Brew is built on LLaMA 3, whereas the research paper describes a Mixtral architecture. The information appears in plain text in the original. An error of that kind does not survive a reading.
Are external links penalised?
This is the most frequently asked question, and the one on which we do not conclude.
We found no rule in LinkedIn’s public documentation penalising posts that contain a link. That documentation describes quality and safety criteria, and states that content judged low-quality or spam-like receives reduced distribution. Links are not listed.
Independent studies diverge, and the gap is too wide to come down to measurement differences.
Ordinal reports a 26.5% drop in reach across more than 900,000 posts over thirty-seven months, with a Mann-Whitney test and a p-value below 0.001. It is the only available study to publish that kind of statistical check. Two reservations come with it. The company sells a LinkedIn publishing tool with a feature that automatically moves links into the first comment, so it has a commercial interest in the penalty being real. And on another page of its site, it presents that feature as a way to avoid « the 60% external link penalty », precisely the unsourced figure discussed below.
Richard van der Blom’s Algorithm Insights report, based on 1.3 million posts, records an 18.8% drop in median reach for a link placed in the body text.
The State of the Algorithm report for Q1 2026, published by Saywhat from 397,605 posts, reaches the opposite conclusion: posts containing several external links perform markedly better there than those containing none. The authors attribute this to content quality.
One hypothesis reconciles these observations without assuming any penalty at all. A reader clicks and leaves the platform. They are no longer reading the post, which collapses Long Dwell. They neither comment nor share, which removes Contribution. The model observes content that does not hold attention and does not prompt reaction, and distributes it less. The link would not be punished: it would mechanically produce the signals of mediocre content.
That hypothesis is not demonstrated. It is consistent with all the observations, which is not the same thing.
Which figures trace back to no source at all?
Four claims, selected because each carries a precise number and none is verifiable.
Engagement rates by reading-time band, said to rise from 1.2% below three seconds to 15.6% beyond sixty. The bands exist, the February paper confirms it. The rates do not: we found none in LinkedIn’s publications.
The « Depth Score », presented since late 2025 as the platform’s new criterion. The term appears in none of the LinkedIn publications we consulted. The idea behind it, measuring how long someone stops on a piece of content, is real and documented since 2020 under the name dwell time. The label comes from elsewhere.
The three percentages circulating in French-language content, always without a reference: commenting under your own post before anyone else would cost 20% of reach, replying within the first hour would gain 35%, commenting elsewhere fifteen to thirty minutes before publishing would add 21%. None can be traced, neither in LinkedIn’s publications nor in any study whose method and sample are public.
Engagement rates by format, finally, which cannot be reconciled with one another. For carousels, depending on the source: 1.44%, 6.60% or 24.42%. None of those tools publishes its calculation method, or what it puts in the denominator.
A precise figure travels better than honest uncertainty, and it is never asked to justify itself. That is probably why « minus 18.8% according to a study of 1.3 million posts » spreads less well than « minus 60% ».
Four questions are enough to sort the noise, and none requires technical knowledge.
Does the source cite a verifiable primary publication?
Does it give its data volume and time window?
Does it sell something its conclusion serves?
Is the figure too precise for its origin?
Why is this a messaging problem before it is an algorithm problem?
Back to the starting point. A technical article describes a two-stage system. One sentence in its second paragraph summarises part of it, accurately. That sentence is the one that travelled, and it shaped the collective understanding of the subject for months.
Nobody lied. The author wrote an accurate sentence, readers repeated it faithfully, and the overall meaning shifted.
The message that survives is almost never the one that was written. It is the one that repeats most easily.
That gap has a price. Sales cycles that stretch because the prospect understood something other than what was written, sales teams that each end up telling their own version. And differentiation that the market rephrases on your behalf.
This is what we audit at Fast Growth Advisors: the sentence your market will remember, the traceability of every promise, and the distance between the two. The white paper this article draws from applies the same method to an outside subject, where every claim is either backed by a public source or flagged as untraceable.
Full white paper: LinkedIn Algorithm 2026: what is proven, what is measured, what is invented. Public document, citable with attribution. Read it.
Sources
- Hertel, L., Srivastava, S. et al. (2026). An Industrial-Scale Sequential Recommender for LinkedIn Feed Ranking. arXiv:2602.12354, 12 February 2026. arxiv.org/abs/2602.12354
- Danchev, H. (2026). Engineering the next generation of LinkedIn’s Feed. LinkedIn engineering blog, 12 March 2026. linkedin.com
- Firooz, H. et al. (2025). 360Brew: A Decoder-only Foundation Model for Personalized Ranking and Recommendation. arXiv:2501.16450, 27 January 2025. Described as pre-production, since withdrawn from arXiv. Available version
- Borisyuk, F. et al. (2024). LiRank: Industrial Large Scale Ranking Models at LinkedIn. ACM SIGKDD, Barcelona, pp. 4804 to 4815. arXiv:2402.06859. arxiv.org/abs/2402.06859
- LinkedIn Engineering. Understanding dwell time to improve LinkedIn feed ranking (2020). linkedin.com
- Jurka, T. (2026). Updates to The LinkedIn Feed Focusing on Authentic, Relevant Conversations. March 2026. linkedin.com
- Ordinal. LinkedIn Link Penalty Study. More than 900,000 posts, February 2023 to February 2026. The page was consulted for this work and has since become unreachable (404 as of 27 July 2026). Publisher site: tryordinal.com
- van der Blom, R. Algorithm Insights, 2026 edition. 1.3 million posts, 50,000 creators. richardvanderblom.com
- Saywhat. State of the Algorithm, Q1 2026. 397,605 posts. saywhat.ai
- Cook, J. (2026). The LinkedIn Algorithm Changed Again. Here’s What’s New For 2026. Forbes, 12 January 2026. forbes.com
FAQ
Is the LinkedIn feed ranked by generative artificial intelligence?
In the two documents LinkedIn published in the first quarter of 2026, no. Ranking is handled by a sequential transformer, called Feed SR in the research paper and Generative Recommender on the engineering blog. A language model operates at the previous stage, retrieving candidate posts, and in representing the member’s profile.
What is Long Dwell?
Time spent on a post beyond a threshold, one that varies by post type. It is one of the two objectives on which LinkedIn reports its gains, alongside Contribution, which groups reactions, comments and shares. The threshold values are not published.
Should links go in the first comment?
Public data does not settle it. Three serious studies reach incompatible conclusions, one of them the reverse of the other two, and the fate of links placed in comments remains poorly documented. The widely circulated 60% penalty figure traces back to no identifiable source.
Does the author’s profile headline influence how a post is distributed?
The engineering blog lists what makes up the description of a post sent to the model: its text, its format, its counters, and information about its author, including profile headline, company and industry. The headline is therefore part of what situates a piece of content. LinkedIn does not quantify the effect.
Is this information still current?
It describes the state of the February and March 2026 publications. We found no publication on feed ranking since. Referring to the current algorithm therefore assumes nothing significant has changed since, which is likely and unverifiable.
If you remember one thing, what should it be?
That the ranking system predicts six named actions, grouped into two families, and that any claim acting on none of them deserves scrutiny.