Something strange is happening in the LinkedIn advice economy. Search for what changed in the feed this year and you will find dozens of confident posts quoting precise figures: reach down 47 per cent, sixty per cent of distribution routed to interest clusters, follower-based reach collapsed from forty per cent to fifteen. The numbers are specific. They are also, as far as anyone can trace, unsourced.
Meanwhile LinkedIn published a long and unusually detailed technical account of what it actually rebuilt. Comparatively few people quote that.
This matters if you are a founder with limited hours and no marketing team. Optimising for an algorithm nobody has described accurately is an expensive way to spend a Tuesday. So here is the separation: what LinkedIn has published, what remains unconfirmed, and what you can reasonably act on.
What did LinkedIn actually change in its feed?
LinkedIn replaced its multi-source retrieval system with a single language-model-based one, and replaced independent post-by-post scoring with a sequential ranking model. On 12 March 2026 the company published an engineering post, “Engineering the next generation of LinkedIn’s Feed”, written by Hristo Danchev, describing the rebuild across a platform LinkedIn says now reaches more than 1.3 billion members.
The old system pulled candidate posts from several separate pipelines running in parallel: a chronological index of network activity, trending content filtered by geography, collaborative filtering based on similar members, industry-level trending, and various embedding-based retrieval systems. Each carried its own infrastructure and its own optimisation logic. LinkedIn’s stated problem was maintenance cost and the difficulty of tuning across sources that no single team controlled.
The replacement is a dual encoder. One language model reads both sides of the match. On the member side it takes work history, skills, education and the ordered sequence of posts that member previously engaged with. On the content side it takes format, author headline, company, industry, engagement counts and the post text itself. Both are placed in the same embedding space, and relevance becomes semantic proximity rather than a negotiated blend of separate signals.
Ranking changed as well. The previous approach, according to the post, treated each impression independently: it judged whether to show you something without reference to what you had just read. The new ranking model, which LinkedIn calls the Generative Recommender, processes over a thousand of your recent interactions as an ordered sequence, using causal attention so that each moment can only see what came before it. The stated aim is to understand, in LinkedIn’s phrasing, “what a post is actually about” and how that relates to where a member’s interests are heading rather than only where they have been.
What is 360Brew, and is it running your feed?
360Brew is a LinkedIn research paper, and it has not been confirmed as the model ranking your feed. This is worth stating plainly, because a great deal of current LinkedIn advice is built on the assumption that it is.
The paper, arXiv 2501.16450, was submitted on 27 January 2025 by Hamed Firooz and colleagues at LinkedIn. It describes a 150-billion-parameter, decoder-only model capable of handling more than thirty predictive tasks across the platform. Two details in its own abstract are routinely left out when it is cited. The authors describe 360Brew V1.0 as a “research pre-production model”. And the performance claims are based on offline metrics, not live traffic.
There is a further detail that almost nobody mentions. The paper has been withdrawn from arXiv. The administrative note attached to it states that the version was removed because the submitter did not have the right to agree to the licence at the time of submission. That is a rights matter, not a judgement on the quality of the research. But it does mean the single most-cited source in LinkedIn strategy content right now is a withdrawn preprint about a model its own authors called pre-production.
LinkedIn’s March 2026 engineering post does not use the name 360Brew at all. It calls the ranking model the Generative Recommender. Coverage of the rebuild notes that the two share a decoder-only architecture and a similar parameter count, so a lineage is entirely plausible. Plausible is not the same as confirmed, and that distinction should change how confidently anyone writes about it.
Why are the reach statistics you keep seeing unreliable?
They are unreliable because they have no traceable origin and they contradict each other. Set the widely circulated figures side by side and you will find one article claiming average reach fell by roughly half, another claiming forty-seven per cent, and a third giving a range that does not overlap with either. None link to a dataset, a methodology, or a sample size.
This is not an argument that reach did not fall. Many people’s reach genuinely did, and that experience is real. It is an argument about what you can build a strategy on. A number you cannot verify is not evidence. It is a mood, converted into a statistic to make it sound like evidence.
You Creatives is a Stockholm-based personal branding and content studio for founders, and this is the pattern we see most often in the Nordics: a founder reads a confident number, concludes they are failing against a benchmark that was never measured, and quietly stops posting. The number did the damage, not the algorithm.
What does the rebuilt feed appear to reward?
Three things, as far as the published material actually supports: topical consistency, a profile that matches what you write about, and sustained interest rather than a spike of reactions. None of that is a hack. It is close to the advice you would have been given ten years ago, which is either reassuring or annoying depending on the kind of week you are having. The difference now is that there is a documented mechanism behind it, so it is worth understanding why each one matters.
Why does topical consistency now matter more than network size?
Because the system is matching meaning, not counting connections. If both your profile and your posts are placed in the same semantic space as the people who should be reading you, distribution can reach beyond your immediate network. LinkedIn’s engineering post makes this explicit in its description of cold-start handling: a new member with nothing but a headline and job title can still be understood, because a model trained on a large corpus knows that an electrical engineer writing about grid optimisation likely has latent interest in renewable infrastructure, even with no engagement history to confirm it.
For a founder with eight hundred connections and a specific point of view, this is the most useful thing in the whole document. Semantic clarity is something you can actually build. Network size takes years.
Why does your profile need to match what you write about?
Because your profile is part of the input. The member-side prompt includes headline, work history and skills, and it is encoded into the same space as the content. A profile that describes one thing while your posts consistently discuss another gives the model a muddled signal about who you are and who should see you. Aligning the two is unglamorous work that costs an afternoon rather than a budget.
What happens to posts written mainly to trigger engagement?
Honestly, we do not know with confidence, and neither does anyone claiming otherwise. A sequential model that reads a thousand interactions in order is structurally better placed to distinguish sustained interest from a burst of shallow reaction than a model scoring each impression alone. Whether LinkedIn actively penalises engagement-bait, or simply stops rewarding it as much, is not something the published material settles. Treat any post that tells you exactly how the penalty works as speculation.
What should a founder with a small network do differently?
Narrow the topic before widening the audience. If your last twenty posts span hiring, funding, climate policy, productivity and a holiday photo, there is no coherent signal for a semantic system to match against. Three or four adjacent themes, held for months, give the model something to work with and give a reader a reason to remember you for something.
Then make the profile agree with the posts. Then keep the voice recognisably yours, which is harder than it sounds once you start writing for distribution rather than for a person. We have written separately about keeping your own voice when someone else drafts your posts, because that is where most founder content quietly goes wrong.
The same logic now applies well beyond LinkedIn. AI answer engines match meaning rather than keywords, which is why we covered how GEO and AEO change discoverability for founders. The mechanism is the same one described above, arriving on a different surface.
What is the honest trade-off here?
Narrowing your topic means saying less about things you find interesting. That is a real cost, and it is felt sharply by founders who are genuinely curious people. A tighter topic may make your feed less fun to write and, for a while, quieter than it was.
There is a second caveat. Everything above describes an architecture, not an outcome. LinkedIn has told us how the system reads content. It has not told us how much reach any given approach produces, and it never will. Anyone converting an architecture description into a promised percentage lift is doing the same thing as the unsourced statistics at the top of this article, just with better sources.
What we would say with confidence is narrower and more durable. A system that reads meaning rewards people who mean something specific. That was true before this rebuild and it will outlast it. If you want help finding what that specific thing is for you, that is what our brand and content services exist to do.
Frequently asked questions
Has LinkedIn confirmed that 360Brew powers the feed?
No. LinkedIn’s March 2026 engineering post describes the ranking model as the Generative Recommender and does not use the name 360Brew. The 360Brew paper itself, published on arXiv in January 2025, describes a research pre-production model evaluated on offline metrics, and that paper has since been withdrawn from arXiv over a licensing rights issue. The architectures resemble each other closely enough that a connection is likely, but LinkedIn has not stated one.
Does a small network still limit how far my posts travel?
Less than it used to, based on what LinkedIn has published. The retrieval system matches members and posts by semantic similarity in a shared embedding space, and LinkedIn specifically describes inferring likely interests for members with no engagement history. That said, your network still shapes early engagement signals, so a small network is a slower start rather than no start.
Should I change my posting frequency because of this update?
Nothing in LinkedIn’s published material addresses frequency, so any specific recommendation about it is guesswork. What the material does support is consistency of subject matter. If you are choosing between posting five scattered times a week and twice a week on a tight theme, the second is better aligned with how the system is described as working.
