What actually changed
On 12 March 2026 LinkedIn published the most detailed account of its feed it has ever released, written by Hristo Danchev on the LinkedIn Engineering blog. It describes replacing a fragmented pipeline of specialised ranking models with one system built on a foundation model.
That model is 360Brew, described in an accompanying paper as a decoder-only foundation model for personalised ranking and recommendation. It is roughly 150 billion parameters, built on LLaMA 3 and fine-tuned on LinkedIn data, and it replaced five separate content retrieval systems that had each been scoring a different slice of the problem.
The structural point is the same one that happened on X. Both platforms spent a decade adding specialised models and hand-tuned features, and both have now collapsed that into a single large model that reads what a post means rather than counting the signals attached to it.
If you have read how the X algorithm works, the shape will be familiar even though the details differ.
The pipeline, as published
Two stages, and the division of labour between them matters more than any individual ranking factor.
Stage one, retrieval. An LLM-powered system narrows millions of candidate posts to roughly two thousand, using embeddings and cosine similarity rather than keyword matching. This is the stage that decides whether your post is even considered for a given person.
Stage two, ranking. A generative recommender takes those two thousand and orders them, processing over a thousand of that reader's recent interactions as a chronological sequence. It is predicting relevance to one specific person given everything they have recently done, rather than scoring your post in the abstract.
The consequence worth internalising is that there is no longer a single feed ranking. There is a prediction about one reader at a time, and the same post can be ranked completely differently for two people with identical job titles.
What it weighs, and why saves beat likes
Interactions are weighted by how much intent they demonstrate. A save is the strongest signal in the published account, a substantive comment sits above a short one, and a like is near the bottom because it costs the reader almost nothing.
This is a meaningful departure from the old model, where reaction volume drove distribution and the optimisation was to collect as many cheap responses as possible. Under a system that reads comments rather than counting them, four people writing paragraphs is worth more than forty people tapping an icon.
The practical implication for a launch post is that you should write something a reader would save to send to a colleague, which is a much higher bar than something they would agree with.
The numbers everyone quotes
Search this topic and you will find confident figures: dwell time is the number one factor, sixty-one seconds is the threshold for maximum distribution, the golden hour is exactly sixty to ninety minutes, polls now get 0.07 per cent engagement.
We would treat all of those the way we treat the engagement weights people quote for X. Dwell time being a real and confirmed signal is well established, and LinkedIn Engineering has said so directly.
A specific second-count threshold is not something LinkedIn has published, and a threshold that precise is the kind of claim that gets repeated between blogs until it acquires the appearance of a source.
The direction is reliable. Time spent reading matters, early response matters, intent-heavy interactions matter more than cheap ones. The exact ratios are inference presented as documentation, and building a content plan on a number nobody can source is how teams end up optimising for something that was never measured.
Why pods stopped working
Engagement pods, comment rings and the reciprocal-liking economy degraded sharply through 2026, and the mechanism is worth understanding because it generalises.
A system that counts comments cannot tell a real one from a manufactured one. A system that reads them can.
When the ranking model is a language model processing the actual text, twenty comments saying "great post" from the same rotating group are legible as exactly what they are, and the same applies to engagement bait phrased as a question the author does not want answered.
This is the same reason identical copy across five creator accounts fails on X. Amplification that looks manufactured gets treated as manipulation, and the fix on both platforms is the same: different people, different angles, in their own words.
How this differs from X
Both platforms now run on a single large model, and it is tempting to conclude the playbook transfers. It mostly does not.
| X | ||
|---|---|---|
| Decides on | Engagement velocity against your recent baseline | Predicted relevance to one reader, given their history |
| Window | Minutes. The first stretch is close to everything | Hours, and distribution can keep widening for days |
| Best signal | Replies and reposts from credible accounts | Saves, then substantive comments |
| Audience | Practitioners, in public, arguing | The person who signs, plus their committee |
The most expensive mistake is treating the two windows as equivalent. A launch on X lives or dies in the first hour, so amplification has to be compressed.
A LinkedIn post that starts slowly can still travel two days later if the right person saves it, which means the same urgency does not apply and the same amplification is wasted effort.
What it means for a launch
Our order has not changed: X first on launch day, LinkedIn second, and only where it earns its place. The reasoning is in LinkedIn launch strategy, and it is about who is on each platform rather than about which algorithm is better.
What changes under 360Brew is the LinkedIn half. Write for one reader rather than for a feed, because the model is predicting for one reader. Write something worth saving rather than something worth agreeing with.
Expect the post to keep moving for days rather than dying in an hour, and do not judge it at the sixty-minute mark. And stop cross-posting the X version, which was already failing and now fails against a model that can tell it was written for somewhere else.
Questions we get asked
What is 360Brew?
LinkedIn's foundation model for feed ranking, described in a paper as a decoder-only model for personalised ranking and recommendation. It is roughly 150 billion parameters, built on LLaMA 3 and fine-tuned on LinkedIn data, and it replaced five separate content retrieval systems. LinkedIn's engineering blog set out the architecture publicly in March 2026.
Is dwell time the most important LinkedIn ranking factor?
Dwell time is a real and confirmed signal, but the specific thresholds circulating are not something LinkedIn has published. Figures like sixty-one seconds for maximum distribution get repeated between blogs until they look sourced.
The direction is reliable: time spent reading matters, and intent-heavy interactions like saves outweigh likes. The exact ratios are inference.
Why did LinkedIn engagement pods stop working?
Because the ranking model reads comments rather than counting them. A system that counts cannot tell a genuine comment from a manufactured one, and a language model processing the actual text can. The same logic explains why identical copy across several creator accounts fails: amplification that looks manufactured is treated as manipulation.