The LinkedIn algorithm in 2026: what actually changed

The LinkedIn algorithm changed in a documented, specific way in 2026: on 12 March, LinkedIn Engineering published a description of a rebuilt feed that 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. That is the whole confirmed change. Everything else you have read this year, including the reach-drop percentages and the lists of newly penalised behaviours, is extrapolation on top of that one post.

What does the primary source actually say?

The document is “Engineering the next generation of LinkedIn's Feed” by Hristo Danchev, published on LinkedIn's engineering blog on 12 March 2026. It is written for engineers, which is probably why so few people writing about it seem to have read past the summary. Two systems were replaced.

Retrieval: one encoder instead of many sources

Previously, candidate posts were gathered from several separate retrieval sources stitched together. The rebuild uses a single unified system built on an LLM-based dual encoder. One side builds a representation of the member from profile information, skills, work history, education and the chronologically ordered sequence of posts they have engaged with. The other builds a representation of candidate content. Retrieval is then a similarity lookup between the two, trained with an InfoNCE objective using a mix of easy and hard negative examples.

The reported engineering results: bucketing numerical features by percentile improved Recall@10 by 15%, using one easy and two hard negatives improved recall by a further 3.6%, and the whole thing serves under 50ms across a platform with more than 1.3 billion members.

Ranking: a sequence model instead of independent scores

The ranker is a transformer-based Generative Recommender. Rather than scoring each candidate post on its own, it treats a member's feed history as a sequence and reads more than a thousand prior interactions in order, with a Multi-gate Mixture-of-Experts prediction head over DCNv2 experts. LinkedIn also reports training efficiency work: a 37% memory reduction, 40% more training sequences per batch, 2.6x faster training, and a further 2x speedup from custom attention kernels.

Those last figures are about training cost. They say nothing about your reach, and several summaries have quietly converted them into implied audience gains.

What the technical terms mean, in plain words

The engineering post is precise and largely unread, which is a bad combination for a subject this widely discussed. Five terms carry most of the meaning.

A dual encoder is two models trained together to put related things in the same place. One reads you and produces a list of numbers. The other reads a post and produces a list of numbers. If the two lists are close together, the post is considered relevant to you. The useful property is that once every post has been turned into numbers ahead of time, finding your candidates is a geometry problem rather than a reading problem, which is how it runs in under 50ms.

An embedding is that list of numbers. The reason this matters for writers is that an embedding represents meaning rather than spelling. Two posts using completely different vocabulary about the same subject land near each other, and a post that repeats a keyword twelve times does not move closer to anything. That is the technical death of keyword stuffing.

Hard negatives are training examples that look relevant but are not. Teaching a model with obviously unrelated pairs is easy and produces a lazy model. Teaching it to separate a post you engaged with from a very similar post you ignored is what produces fine distinctions. LinkedIn reports that adding one easy and two hard negatives improved recall by 3.6%. In practical terms, the system is specifically trained to tell near-misses apart, which means generic content in a crowded subject area competes against a model tuned to notice the difference.

A Generative Recommender here means a transformer applied to your interaction history the way a language model is applied to a sentence. Your last thousand feed interactions are the sequence, and the model predicts what comes next. The consequence is order-sensitivity: the same post can rank differently for two people with identical interests but different recent behaviour.

Multi-gate Mixture-of-Experts is how one model predicts several things at once, such as the chance you click, comment or dwell, without those predictions degrading each other. Each prediction gets its own gate deciding which internal experts to consult. It is worth knowing this exists because it is direct evidence that the system still predicts multiple distinct actions. Anyone claiming the feed now optimises for a single metric is contradicted by the architecture.

How did the old system work, and why does that matter?

The comparison is what makes the change legible. The previous stack gathered candidates from several separate sources, each with its own logic: things your connections engaged with, things from pages you follow, things matching your declared interests. Those pools were merged, then each post was scored on its own by a collection of specialised models predicting individual outcomes, and the scores were combined into a final order.

That design has a structural property worth naming: your network was the gate. Almost everything you could see had to arrive through a social path. It is the reason the standard growth advice for a decade was to build the connection graph first and worry about content second, and it is the reason a good post from an account with 300 connections had a hard ceiling.

Replacing that with unified embedding retrieval removes the gate. Not entirely, since network signals still feed the member representation, but the path from a stranger to your post no longer requires an intermediary. This is the single most consequential difference between the two systems, and it is why so much pre-2026 growth advice now reads as slightly beside the point.

Why did so many people report a reach drop?

Redistribution produces loud losers and quiet winners. When the basis for candidate selection changes across an entire platform, some accounts gain and some lose, and only one of those groups writes a post about it. That asymmetry alone guarantees that any change of this size reads publicly as a decline, independent of what happened to the average.

Three further things make the reported numbers unreliable. Self-selected panels recruit from people who noticed a change, which is a sample of the affected. Nothing isolates the platform change from seasonality, from the reporter's own shifting posting habits, or from ordinary week-to-week variance, which in our data is very large. And impressions are a lifetime-cumulative number, so a comparison of a recent post against an older one is partly measuring how long each has existed.

The honest position: a large change happened, it plausibly moved reach around, and any precise figure for what it did to accounts in general is fabricated. Our guide to organic reach decline works through why panel estimates in this niche run consistently pessimistic, and why reach drops covers the mundane account-level explanations that are usually the real answer.

What did not change?

This is the more useful list, because it is where most of the year's advice went wrong.

• No published weighting. LinkedIn described an architecture, not an objective function. Nothing states that a comment is worth four likes or that saves outrank shares. Every such number in circulation is invented.
• No stated penalties. The post does not announce a penalty for links, hashtags, edits, emojis, or anything else. Pages presenting a 2026 penalty list are describing folklore.
• No posting-time guidance. Retrieval by relevance weakens the case for precision timing rather than strengthening it, but LinkedIn published nothing on this. Our timing guide covers what the published studies actually measured.
• No follower-count mechanics. Nothing in the rebuild says small accounts are favoured or disadvantaged.

What actually changes for you?

Three consequences follow from the architecture. Everything beyond these three is speculation, including ours.

1. Strangers can find you, and irrelevance costs more

Embedding-based retrieval means a reader with no connection path to you can be shown your post because their interests match its content. That is a genuine expansion of the ceiling for a small account writing clearly about one thing. The same mechanism raises the floor cost of writing about nothing in particular: a post with no identifiable subject has no interest to be matched against, so it falls back on your immediate network and stops there.

2. Topical coherence became mechanical

“Pick a lane” used to be positioning advice you could reasonably ignore. Under a system that builds a representation of what you write about, a consistent body of work produces a cleaner signal than the same volume scattered across unrelated subjects. This is the strongest argument for content pillarsthat does not depend on anyone's taste.

3. Judging a post by its first hour makes even less sense

If ranking depends on a thousand prior interactions per reader, then whether your post surfaces for a given person depends heavily on what that person did recently. Post-level variance was already high, and this widens it. Our golden hour analysis reached the same conclusion before the rewrite, from the evidence rather than the architecture.

Which tactics does this genuinely retire?

Note how short the “retired” column is. A structural rewrite of the serving stack turns out to invalidate fewer writing habits than a typical algorithm explainer claims, because most writing habits were about people rather than machinery. For the two rows with live debates behind them, see hashtags and engagement pods.

What does the data say about what still works?

We analysed 12,988 posts from 65 creators, comparing the top decile by engagement rate against the bottom half. These are pre-rewrite measurements, which we will come back to, but they describe reader behaviour rather than serving mechanics.

The pattern worth noticing is that none of these are algorithm exploits. Writing in the first person, addressing the reader directly and using native video are all things that change how a human responds. That is why we expect them to survive a ranking change: the ranker is trained to predict human response, so anything that genuinely improves human response keeps working through architecture changes that invalidate tricks.

The comment gap is the sharpest signal in the data. Top-decile posts show a median of 72 comments against 5 in the bottom half, a far wider separation than reactions (235 against 80). Full method in the engagement study.

The limitation, stated plainly: this data predates the March 2026 rewrite. It describes what earned engagement under the previous distribution system. We think the hook and format findings age well because they are about readers, and we would rather flag the caveat than present pre-rewrite measurements as a description of the current feed, which is what most pages citing 2024 and 2025 studies are currently doing without saying so.

How should you actually respond?

1. Change nothing for a month. You cannot measure a platform change against your own account. You do not have the sample size, and everything else moved at the same time.
2. Narrow your subject range. This is the one adjustment with a mechanism behind it. Three recurring themes, not nine.
3. Name things plainly. Write the actual noun for the category you work in rather than gesturing at it. This helps retrieval and it helps readers.
4. Keep publishing on a cadence you can sustain. Consistency remains the variable most people fail on, and it has nothing to do with architecture. Our posting frequency guide covers what the evidence supports.
5. Judge posts over weeks. Sequence-aware ranking makes single-post readings even noisier than before.

Six claims about the 2026 update, checked

Each of these appears on multiple first-page results. We traced each to its furthest reachable source.

One of six is supported by a primary source. That ratio is roughly what we find every time we audit this niche, and it is the reason we date every platform claim we publish. For the full breakdown of the first two rows, see our 360Brew analysis; for the link question, see external links on LinkedIn.

Does this change anything for company pages?

Probably, and in a direction that is easy to get wrong. The member representation described in the engineering post is built from profile information, skills, work history, education and engagement history. A company page has no work history and no education, and its posts are not authored by a person with those attributes. Nothing in the post says pages are disadvantaged, so this is inference rather than documentation.

What is documented is that retrieval matches content representations to member representations. That mechanism does not obviously care whether the author is a person or an organisation. The gap between personal and page reach that most B2B teams observe has never had a published mechanical explanation, and it still does not have one. Treat anyone offering a confident number here with the same suspicion as the reach-drop figures.

How do you audit your own content against this?

A short exercise that takes twenty minutes and is more useful than any tactic in this article. Take your last fifteen posts and answer three questions for each.

1. Could a stranger tell what this post is about from the first two lines? Not whether it is interesting. Whether the subject is identifiable. Count how many fail.
2. How many distinct subjects appear across the fifteen? If the answer is more than four, retrieval has a blurry picture of you. This is the number most people are surprised by.
3. Does the post name its category in plain words anywhere? Posts that only gesture at a subject through metaphor or in-group shorthand give the encoder very little to work with.

The typical result is that a third of posts have no identifiable subject and the subject count is around seven. Fixing that is a larger reach intervention than any posting-time adjustment, and unlike posting time it also makes the writing better. Our content audit guide covers the fuller version of this exercise.

What is worth watching next?

Three things would genuinely change the picture, and none of them is another blog post about weightings.

Any published figure for out-of-network share. The rebuild makes stranger-to-you distribution possible. Whether it accounts for 5% or 40% of the feed is the difference between an interesting architectural note and a reason to rethink how anyone grows an audience here. LinkedIn has not said, and it is the number most worth having.

Whether ranking and retrieval get unified. The current design keeps them separate, which is why the 360Brew idea of one model doing everything is interesting rather than academic. If a future post describes a single model handling both, that would be the change that actually matches what people have already been writing about for a year.

Anything LinkedIn says about content quality directly. The engineering post is about machinery, not editorial standards. Statements about generic or automated content would carry real weight, and would matter more to most readers than any retrieval detail. Until then, claims that the feed penalises AI writing are unsourced, which we cover in AI content detection on LinkedIn.

Until one of those arrives, the practical picture is stable, and it is a disappointing picture for anyone hoping for a lever. Write about a recognisable subject, say it in a way that earns the second line, and keep going long enough for the sample size to mean something.

The one-paragraph version

LinkedIn rebuilt feed retrieval around an LLM dual encoder and feed ranking around a sequential transformer, and published the details on 12 March 2026. The practical effect is that relevance to a reader's interests now matters more than proximity in your network, which helps clear writing about a specific subject and hurts posts that are about nothing in particular. No weights were published, no penalties were announced, and no reach data was released. For the separate question of whether the model everyone names is the one that shipped, see our breakdown of what 360Brew is and is not. For the fundamentals that did not change, start with the algorithm guide.