The 2026 LinkedIn algorithm is described by two in-house publications: a research paper in February and an engineering blog post in March. They show a two-stage system whose ranking does not rely on a language model, and six actions the feed tries to trigger. At Fast Growth Advisors, we read both in full to separate the proven from the invented. Several widely shared figures trace back to no source.
Our methodology: read those two documents and two earlier publications, report what they describe, flag what they leave unsaid. Advertising, company pages and job recommendations belong to other systems and stay out of scope.
What actually ranks the LinkedIn feed?
Two stages, not one: retrieval narrows millions of posts down to a few hundred candidates, then ranking orders them.
Conflate the two and everything that follows goes wrong.
Retrieval does rely on a language model, built as a dual encoder. Member and post are each turned separately into two comparable descriptions, which matches a reader with content on meaning rather than on the keywords they share. One sentence of the engineering post, accurate but isolated, made people believe the whole feed worked this way.
Ranking uses another model. Its name: Feed SR (Sequential Recommender) in the research paper, Generative Recommender in the blog post. It is a causal-attention sequential transformer, reading your interactions in chronological order without ever anticipating what comes next.
No language model at this stage.
Rarely for such papers, it also details the architectures that were dropped. A ranker based on a language model was indeed tested in live conditions. Three obstacles. First, the values that matter for ranking (reactions, age, closeness to the author) translate poorly into sentences. Then cost: hundreds of text units per post, against 2 in Feed SR. And no online gain over the model in service.
One caveat applies to what follows: LinkedIn describes a result obtained on a given date, not a doctrine. Fast Growth Advisors found no announcement that this path was abandoned.
What is the ranking system trying to trigger?
Six named actions, the most directly usable information in either document and the least reported.
On the passive side sit the click, the scroll-past and Long Dwell, that is, reading time beyond a threshold specific to each format. On the active side, the reaction, the comment and the share, which LinkedIn files under the label Contribution. Those two signals, Long Dwell and Contribution, are where the company reports its gains.
The model’s memory is bounded: 1,000 viewed posts, drawn from roughly a year.
It also decays, in two ways. On the training side, data loses half its influence every sixty days, and within a single member’s sequence, older interactions count about half as much as recent ones.
A less expected signal: how previous readers spread 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.
Result of the live experiment: 2.10% more time spent, mostly among very active members, with no measurable effect on new ones.
Which leaves saving a post, often presented as the primary ranking factor. No mention among the optimised objectives. For Fast Growth Advisors, this does not prove it carries no weight: no primary source supports it, while the other six actions are named one by one.
How did a research model become the LinkedIn algorithm?
Through a claim repeated without checking. 360Brew exists: it is a paper filed on arXiv in January 2025 by a LinkedIn team, describing a 150-billion-parameter model capable in principle of handling the feed, job recommendations and connection suggestions in a single system.
Three facts, and nothing beyond them.
- Its abstract describes it as a “pre-production” model.
- We found no confirmation of deployment by LinkedIn.
- The March 2026 engineering post does not mention the name.
In January 2026, Forbes published an interview in which a creator followed by 1.2 million people stated that the latest update was called 360 Brew. Within weeks the claim had become settled fact, and dozens of publications built tactical advice 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; the research paper describes a Mixtral architecture, as Fast Growth Advisors checked. An error of that kind does not survive a reading.
Are external links penalised?
Public sources do not allow anyone to say so, and Fast Growth Advisors does not conclude.
We found no rule in LinkedIn’s public documentation penalising posts that contain a link. It states that content judged low-quality or spam-like receives reduced distribution. Links are not mentioned.
As for independent studies, they diverge, too widely for a mere measurement difference.
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: the only published statistical check. Two reservations, though. Ordinal sells a tool that moves links into the first comment, so it has an interest in the penalty being real. Another page of its site even presents that feature as a way to avoid “the 60% external link penalty”, the unsourced figure discussed below.
Richard van der Blom (Algorithm Insights, 1.3 million posts) records an 18.8% drop in median reach for a link in the body text.
Saywhat’s State of the Algorithm (first quarter of 2026, 397,605 posts) reaches the opposite conclusion: posts with several external links do markedly better than those with none. Its authors attribute this to content quality.
One hypothesis reconciles these observations without assuming any penalty. One click, and the reader leaves the platform: they stop reading, and Long Dwell collapses. They neither comment nor share, and Contribution disappears. Faced with content that does not hold attention, the model 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 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 that Fast Growth Advisors could not trace to any source.
First case: engagement rates rising from 1.2% below three seconds of reading to 15.6% beyond sixty. The bands exist, the rates do not.
Second case, the “Depth Score”, presented since late 2025 as the platform’s new criterion. The term appears in none of the LinkedIn publications we consulted. Its idea, measuring how long someone stops on a piece of content, is real and documented since 2020 under the name dwell time. Only the label comes from elsewhere.
Third case, three percentages circulating in French-language content without a reference: commenting on your own post first would cost 20% of reach, replying within the hour would gain 35%, commenting elsewhere fifteen to thirty minutes before publishing would add 21%. Untraceable at LinkedIn and in any study that publishes its method and sample.
Last case, engagement rates by format, which cannot be reconciled. For carousels, depending on the source: 1.44%, 6.60% or 24.42%, with no published calculation method or denominator.
A precise figure travels better than honest uncertainty, and it is never asked to justify itself. Hence, probably, the success of “minus 60%” over “minus 18.8% according to a study of 1.3 million posts”.
Four questions are enough to sort the noise, with no technical knowledge required:
- 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?
Because one accurate sentence, taken out of context, was enough to skew the collective understanding of a system for months.
Nobody lied. Its 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 else, sales teams that each tell their own version, differentiation that the market rephrases on your behalf.
This is what the Fast Growth Advisors messaging audit measures: the sentence your market will remember, the traceability of every promise, and the distance between the two. Same method in the white paper this article draws from: every claim is either backed by a public source or flagged as untraceable.
That gap has a name: Message-Market Fit.
How many companies hold on to the sentence they want remembered?
Across the 369 French post-funding startups in our Observatory (edition No. 1, second quarter of 2026), the average score stays at 5.33 out of 10.
Full white paper: LinkedIn Algorithm 2026: what is proven, what is measured, what is invented. Public document, citable with attribution. Read the white paper.
Sources
Primary sources first. Fast Growth Advisors lists the documents published by LinkedIn at the top, then the third-party studies, each with the reservations that its measurement method or its commercial interest calls for.
- Hertel, L., Srivastava, G. et al. (2026). An Industrial-Scale Sequential Recommender for LinkedIn Feed Ranking. arXiv, 12 February 2026.
- Danchev, H. (2026). Engineering the next generation of LinkedIn’s Feed. LinkedIn engineering blog, 12 March 2026.
- Firooz, H. et al. (2025). 360Brew: A Decoder-only Foundation Model for Personalized Ranking and Recommendation. arXiv, 27 January 2025, since withdrawn.
- Borisyuk, F. et al. (2024). LiRank: Industrial Large Scale Ranking Models at LinkedIn. arXiv, 2024.
- LinkedIn Engineering (2020). Understanding dwell time to improve LinkedIn feed ranking.
- Jurka, T. (2026). Updates to The LinkedIn Feed Focusing on Authentic, Relevant Conversations. March 2026.
- Ordinal. LinkedIn Link Penalty Study, February 2023 to February 2026. Page unreachable (404) as of 27 July 2026. Publisher site.
- van der Blom, R. Algorithm Insights, 2026 edition. richardvanderblom.com.
- Saywhat. State of the Algorithm, first quarter of 2026. saywhat.ai.
- Cook, J. (2026). The LinkedIn Algorithm Changed Again. Here’s What’s New For 2026. Forbes, 12 January 2026.
FAQ
Is the LinkedIn feed ranked by generative artificial intelligence?
No, according to the 2 documents LinkedIn published in early 2026.
To rank the LinkedIn feed, the platform uses 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?
It is time spent on a post beyond a given time threshold.
That threshold varies by post type, and its values are not published. LinkedIn reports its gains on two objectives, Long Dwell and Contribution, which groups reactions, comments and shares.
Should links go in the first comment?
Nothing in the public data settles it.
Three serious studies of links on LinkedIn reach incompatible conclusions, one of them the reverse of the other two, and the fate of links placed in comments remains poorly documented. Widely circulated, the 60% penalty figure traces back to no identifiable source.
What does LinkedIn’s model see of a post’s author?
Three things: their profile headline, company and industry.
According to LinkedIn’s engineering blog, the description of a post sent to the model includes its text, its format, its counters, and information about its author, including profile headline, company and industry. The headline therefore helps situate a piece of content, but LinkedIn does not quantify the effect.
Is this information still current?
It describes the state of the February and March 2026 publications.
Fast Growth Advisors found no LinkedIn publication on how the feed is ranked since then. Referring to the current algorithm therefore assumes nothing significant has changed, which is likely but unverifiable.
If you remember one thing, what should it be?
Remember this: LinkedIn’s ranking predicts 6 named actions.
Those six actions form two families, passive (click, scroll-past, Long Dwell) and active (reaction, comment, share). Any recommendation that acts on none of them deserves scrutiny before you apply it to your own posts.