How the LinkedIn feed actually ranks in 2026

The short answer

Retrieval reads the text. Ranking reads your history. That split is the clearest way to hold the whole system in your head, and it explains most of what people find confusing about the feed.

Stage one: retrieval reads the text

Before 2025, LinkedIn assembled candidates from several separate sources: trending posts, your network, and collaborative filtering. Since the rebuild, a single language model — a fine-tuned Llama 3, per LinkedIn's retrieval paper — turns both the post's text and the text describing a member into embeddings, then matches them.

LinkedIn draws an explicit promise from this: you can be shown content from authors you do not follow, on topics you care about, even when the wording differs. That is a genuine improvement in discovery. It is also the mechanism by which roughly half your feed can come from people you did not choose.

Stage two: ranking reads the history

LinkedIn calls the ranker the "Generative Recommender". It treats a member's history as an ordered sequence of more than 1,000 past interactions — what you read at length, liked, commented on, shared — and predicts what you will do with each candidate post: read it at length, skip it, like it, comment, or share it.

The February 2026 paper behind the ranker reports the measured result: after more than three months serving the majority of feed traffic across 1.2 billion members, the sequential model gained 2.10% in time spent and 3.52% in likes, comments and shares over the previous model — and beat the more expensive language-model-only approaches on live results and on compute cost alike.

What changed about a broad post

Everything above collapses to one practical rule. A post aimed at a nameless crowd has no member it is precisely relevant to, so the matcher has nothing to tie it to. A post aimed at an identifiable audience is easy to match, and easy to predict engagement for. This is the mechanism behind the widely reported reach decline — we measured it against the studies in LinkedIn reach fell 47%.

What LinkedIn says it is reducing

The same announcement names the tactics it is actively suppressing, which is unusually explicit for a platform that publishes no visibility guideline:

• Engagement pods. LinkedIn's VP of Product Management said the goal is to make them "entirely ineffective" (Social Media Today, November 2025).
• Third-party software and browser extensions that automate comments. Not allowed on LinkedIn.
• Posts that fish for reactions, such as "Comment 'Yes' if you agree".
• Videos unrelated to the text that accompanies them.
• Recycled thought-leadership posts.

What about 360Brew?

360Brew is the model most articles still name as the algorithm. LinkedIn tested it with a small group of members in early 2025, judged it not the right fit, and shut the test down. On 14 April 2026 LinkedIn's VP of Engineering answered the question publicly — "Is 360Brew used as part of your ranking system? Short answer: No!" — and the original paper has been withdrawn from arXiv over a licensing issue. None of the affected versions of that paper mention the system that actually ranks the feed.

What to do with this

1. Name one audience per post, with a job title and a situation attached.
2. Keep your profile headline descriptive rather than aspirational — it is part of the text the system reads.
3. Be present in the first hour. Ranking predicts comments and long reads, and a live exchange feeds both.
4. Expect minutes, not hours, for distribution. A post's useful life is still one to two days, so there is no scheduled time to optimise for.
5. Stop paying for reach tactics. They cost the same as before and return markedly less.

The part that should bother you

Read the two stages again and notice what is missing: your intent. The matcher knows the text of a post and the text of your profile. The ranker knows what you clicked. Neither of them knows who you decided to pay attention to — that you added someone on purpose, and would like to see what they said.

That is the gap a chosen list fills. Not a better ranking, but a different question: not "what will hold my attention" but "what did the people I chose publish". Surfolks collects those public posts on a schedule and scores them against your goals, and it never posts, comments, or sends anything on your behalf — the boundary is in what we collect. For the X side of the same story, see X's ranking weights.