LinkedIn stopped reading your post and started reading you

LinkedIn stopped evaluating your post as a thing in itself. On 12 March 2026 its engineering team published the most detailed public account of the feed it has ever released, and the system described there runs in an order almost no advice assumed. It builds a representation of a person first, then asks which posts fit that person. The post is the candidate answer. The person is the question.

That sounds like a small reordering of the same machinery. It is not. Almost everything written about this platform before that date assumes a post walks into an arena and competes on its own merits: the hook, the format, the first sixty minutes, the comment count. The new architecture makes your post one possible answer to a question the system already formed about a reader who has never heard of you. If it has not formed a question you can answer, your hook is irrelevant, because the hook is only read after retrieval has let you into the room.

So the useful question changed shape. It is no longer how do I make this post perform. It is what does the system currently believe I am consistently about, and is that belief specific enough to match a stranger to. What follows is the evidence, including the parts that cut against it.

What actually changed on 12 March 2026?

The document is called Engineering the next generation of LinkedIn's Feed, written by Hristo Danchev, on LinkedIn's own engineering blog. It describes retiring four separate retrieval systems (chronological, trending, collaborative filtering and embeddings) that each pushed candidates towards you, and replacing all of them with one retrieval layer built on large language model embeddings. Above that sits a new ranking model LinkedIn calls the Generative Recommender. Two changes shipped together, and the one that decides whether you are seen at all got the least attention.

Ranking is the part everyone argues about. Retrieval is the part that decides your fate. Retrieval is the shortlist: out of millions of indexed posts, which handful are even worth scoring for this member. Ranking only orders what retrieval already chose. That shortlist used to come from proximity and behaviour, meaning who you follow, what people like you clicked, and which words matched which words. The new layer works on meaning. LinkedIn's stated intent is to surface content from authors you do not follow, on topics you care about, expressed in different terminology than you might use. Read that last clause twice. Keyword overlap is no longer the bridge between your writing and a stranger. Similarity of meaning is.

LinkedIn published its working too, which is rarer than it sounds for a company this size. Converting raw engagement counts into percentile buckets improved the correlation between popularity signals and embedding similarity by 30 times, and improved Recall@10 by 15%. Two hard negatives per member added a further 3.6%. Retrieval now returns in under 50 milliseconds across millions of indexed posts. I am not repeating that to sound technical. A company that publishes its recall improvements is telling you where it spent the money, and it spent it on the step that decides whether you are considered at all.

1. 27 Jan 2025
   360Brew paper publishedLinkedIn's Foundation AI Technologies team puts a 150 billion parameter decoder-only ranking model on arXiv, thirteen months before the feed rebuild ships
2. 3 Nov 2025
   Default AI training on member dataLinkedIn begins using member profile data and public posts to train generative AI models by default, with manual opt-out; private messages excluded
3. 12 Mar 2026
   The feed rebuild goes publicFour retrieval systems replaced by one LLM semantic layer, plus a Generative Recommender ranking over more than a thousand ordered reader interactions
4. 18 Mar 2026
   Engagement pods crackdown on recordLinkedIn's VP of Trust Product confirms pod groups were removed and that suspected pod activity limits how far a post travels outside the author's network
5. 14 Apr 2026
   Metricool publishes the awkward data673,658 posts: likes, comments and shares all down year on year, while clicks and total engagement rose
6. 30 Jul 2026
   The slop button and the AI writer's funeralLinkedIn ships 'Seems like AI slop' reporting and discontinues its own 'Enhance your post' AI writing feature on the same day

LinkedIn's March 2026 rebuild replaced network and keyword based retrieval with meaning based retrieval, which makes what you are consistently about more decisive than how any single post is written.

Who is the 'you' that LinkedIn is now reading?

Here's the thing about that phrase. The you being read, in the engineering document itself, is mostly the reader and not the writer. LinkedIn's ranking model takes more than a thousand of a member's past interactions and feeds them in as an ordered sequence with causal attention, not as a bag of independent events. Order carries meaning. What somebody engaged with this month sits in relation to what they engaged with last year, and the model learns from that relation.

Long dwells appearing first in that list settles an argument the industry has run on vibes for years. Dwell time is a confirmed ranking input, stated by the company, in writing, on its own engineering blog. Not a like, and not a comment, but the silent seconds where somebody stopped scrolling and actually read. You cannot see it in your analytics, you cannot ask a colleague to give you one, and no growth tactic manufactures it. It comes from writing something a person wants to finish, and nothing else.

So where does the writer sit in this. In two places, both less about tactics than most people would like. First, a retrieval layer matching meaning needs a stable meaning to match against. If your last twenty posts point in twenty directions, there is nothing consistent to align with anyone's interest profile. You are not punished for that. You are simply never a confident answer to any question. Second, and colder, LinkedIn says it uses the language model's world knowledge to solve cold start, reasoning about a new member's likely interests from a profile headline and job title alone. Your headline stopped being a label a human skims. It is an input a model reasons from.

• Retrieval basisFour systems: chronological, trending, collaborative filtering, embeddingsOne unified LLM semantic retrieval layer
• Unit of analysisThe individual post and its own engagementAn ordered sequence of 1,000+ reader interactions
• Reader historyEvents scored independentlyCausal attention over ordered events, positive engagement only
• New member cold startLittle to go on until behaviour accumulatesLLM reasoning from a profile headline and job title
• Discovery intentMostly people you already followAuthors you do not follow, on topics you care about

Because LinkedIn's retrieval layer matches meaning rather than keywords, a scattered posting history is not penalised, it is ignored, because there is no stable subject to match a reader against.

Why does 'positives only' change what it costs to have an opinion?

One line in that engineering post changed my thinking more than everything else in it combined, and I have not seen a single algorithm guide quote it. The member interaction sequence is built from positive engagement only. Content a member scrolled past is excluded entirely. It is not stored in your reader's history as a negative signal. It is not stored there at all.

The engineers had a practical reason. Dropping the scrolled-past data cut memory per sequence by 37% and made training 2.6 times faster, which on a platform of 1.3 billion members is the difference between shipping a model and writing a paper about one. The consequence for anybody publishing is closer to philosophical. The system holds a detailed memory of what people loved and none at all of what bored them. The fear that quietly shapes most professional writing here, that a stranger scrolls past your sharp opinion and thinks less of you, has no representation inside the model.

Being ignored leaves no scar on the machine. It leaves no mark at all.

Be careful here, because this is where advice of this kind turns reckless. Positives-only history does not mean LinkedIn has no negative signals anywhere. On 30 July 2026 it shipped a button letting members report a post as seeming like AI slop, and reported posts get reduced reach. That is deliberate. What positives-only removes is the passive penalty, the slow accumulation of having been dull in public.

I have written safely and I know exactly why I did it. A few years and several hundred articles into publishing regularly, I started cutting the sharpest line out of a draft at the last minute, because a sharp line invites an argument and an argument feels like exposure. What I was deleting was the only sentence anybody would have remembered. The architecture has made that cost explicit. Careful and generic gets scrolled, and the scroll is not even recorded. Specific and arguable gets read to the end, and reading to the end is the signal the ranker names first.

Do my followers still decide who sees my work?

Less than you think, in both directions, and this is where the news gets good for anyone starting from nothing. LinkedIn's stated design intent is more out-of-network discovery, and Richard van der Blom's Algorithm Insights 2026, built on 1.3 million posts from 50,000 creators, measures suggested posts from outside a member's network at roughly 9% of the feed. Not a takeover, not a rounding error, and LinkedIn has confirmed the intent to grow it while declining to say by how much.

The more uncomfortable number is the split inside that same dataset. Reach fell around 60% over two years for active creators, while inactive and occasional posters saw reach rise 10 to 15%. Engagement fell 20 to 25%, less steeply than reach, meaning engagement per view went up. Read those three together and the platform is not shrinking. It is redistributing away from high-volume posting. Van der Blom cut his own recommendation from five or six posts a week to two to four, the first time I have watched a LinkedIn consultant tell people to do less of the thing he sells.

That data carries a caveat I keep applying to everything here. Van der Blom sells LinkedIn consulting and his methodology is not independently auditable. I cite him because his sample is large and disclosed, not because his incentives are clean. Nobody measuring this market has clean incentives, which is why the corrections below matter more than the tactics.

• 1,000+Reader interactions in the ranking sequence
• under 50msRetrieval time across millions of posts
• ~9%Feed share coming from outside your network
• 1.3 billionLinkedIn members

Follower count and reach are decoupling in both directions: active creators lost roughly 60% of their reach over two years while occasional posters gained 10 to 15%.

If reach fell, why did total engagement rise?

Because the engagement moved somewhere you cannot screenshot it. Metricool's LinkedIn Study 2026 compared 673,658 posts from 63,108 accounts across January and February 2025 against the same window in 2026. Likes fell 13%, comments fell 17%, shares fell 10%. Clicks rose 5%, and overall engagement rose nearly 14%. Every number that goes into a client report went down, and the total went up.

Set that beside the engineering post and it stops being a contradiction. The visible signals fell, the invisible ones rose, and the ranker's list of embedded engagement types opens with long dwells. Metricool calls this the rise of invisible interactions: clicks, carousel swipes, video views, link taps. I would put it more bluntly. The signals that were easy to manufacture fell, and the ones that are hard to manufacture are what the system reads. If you are still reporting likes, you are telling somebody the wrong story about work that may well have succeeded.

The format data points the same direction. Socialinsider's benchmarks, from 1.3 million posts across 16,645 business pages, put native documents top of the engagement table at 7.00%, then multi-image at 6.45%, video at 6.00%, images at 5.30%, plain text at 4.50%, and links last at 3.25%. State the caveat before acting on it: that dataset covers business pages, not personal profiles. Metricool measured personal profiles at 2.60% against 1.74% for company pages, a real gap and a far smaller one than the folklore claims.

• Native documents7.00%
• Multi-image6.45%
• Video6.00%
• Image5.30%
• Text only4.50%
• Link post3.25%

Read that ordering through dwell and it explains itself. Documents and multi-image posts hold a person in place while they swipe. Links do the opposite, because their purpose is to make the reader leave. Which answers the link question honestly. LinkedIn's Senior Director of Product Management said on record in September 2025 that there is no penalty for posts containing external links, provided the post delivers value on its own, and van der Blom measures an 18.8% median reach reduction for one in-body link. Both are true. There is no penalty flag, and a post whose job is to send you away generates no dwell. The fix is writing a post that is complete without the link.

Which LinkedIn numbers are actually real?

Now the part that matters more than any tactic, because it is a method rather than a tip. Much of what circulates as algorithm knowledge has no primary source behind it, and the fastest way to become the trusted person in your field is to be the one who checks. Here is what I would delete from your notes today.

The first sixty to ninety minutes determine up to 70% of your reach. LinkedIn has never published a time window or that percentage. The mechanism underneath is real, since long dwells and early engagement are confirmed inputs, but every number attached to the golden hour is folklore that hardened through repetition. The tell is that it never arrives with a sample size.

Comments of five or more words count twice as much as a like. No primary source exists. The March 2026 post names which engagement types are embedded and deliberately withholds their weights. That omission is a decision, not an oversight, and I treat any published weighting as invented until LinkedIn reverses it.

LinkedIn's 2026 Authenticity Update, with its 97% detection accuracy and its claim that engagement drops 30 to 45% for four to eight weeks. There is no LinkedIn product or announcement by that name. The real event underneath is an on-record interview with Oscar Rodriguez, LinkedIn's VP of Trust Product, in Forbes on 18 March 2026. The rest came from SEO content citing other SEO content, and the giveaway is the precision itself, because a fabricated statistic is almost always suspiciously exact.

LinkedIn has 175 million Premium subscribers. This one is arithmetic rather than opinion, which is what makes it worth working through. LinkedIn's newsroom announced on 7 March 2024 that Premium had passed $1.7 billion in annual revenue with 25% year-over-year subscriber growth, and it has never published a subscriber count. Divide $1.7 billion by 175 million and every subscriber pays $9.71 a year, against pricing of roughly $30 to $70 a month. It is almost certainly a corrupted echo of the revenue figure, and it sits in dozens of roundups that professionals cite in real decks in front of real clients.

Personal profiles get 561% more reach than company pages. Metricool's transparent measurement says personal profiles run a 63% higher engagement rate, 2.60% against 1.74%, with comparable average impressions per post. Personal profiles do win, by an amount you can defend in a room containing one sceptic, which beats a number that collapses the second somebody asks where it came from.

And the one closest to this piece's own subject: 360Brew is the LinkedIn algorithm. 360Brew is a research paper, submitted to arXiv on 27 January 2025 by Hamed Firooz and LinkedIn's Foundation AI Technologies team, describing a 150 billion parameter decoder-only model built on Mixtral 8x22B. What LinkedIn described shipping in March 2026 is a unified LLM retrieval layer plus a Generative Recommender. Related lineage, different artefacts. LinkedIn has never said 360Brew runs your feed.

	Named primary source	Sample or method published	Confirmed by LinkedIn
'First 60 to 90 minutes decide 70% of reach'	×	×	×
'Comments of 5+ words score 2x a like'	×	×	×
'AI detection is 97% accurate'	×	×	×
'175 million Premium subscribers'	×	×	×
'Personal profiles get 561% more reach'	×	×	×
'Long dwells are a ranking input'	✓	−	✓

LinkedIn has published which engagement types its ranker embeds and has never published their relative weights, so every specific weighting figure in circulation is unsourced.

Is this just personalisation with a new name?

That is the strongest objection to everything above and it deserves an answer. Feeds have been personalised for a decade, and consultants have been telling founders to be consistent since long before any of this shipped. If the conclusion looks the same as it always looked, why does the mechanism matter.

Three reasons, and the first carries the weight. What changed is the retrieval step, not the ranking step. Personalisation used to reorder a shortlist your network largely determined, so a bigger network meant a bigger shortlist, and the old advice was mechanically correct: connect more, post more, get people engaging early. Now meaning determines the shortlist and the network reorders inside it. Different machine, different lever, and pulling the old lever harder is how most accounts spent the last two years.

Second, this version of consistency makes different predictions from the old version, and predictions can be checked. Old consistency advice predicted volume, because more posts meant more chances to hit. The published architecture predicts the opposite for most people, and van der Blom's split between active creators and occasional posters is the exact shape you would expect if that were true. It predicts narrower rather than broader, and dwell rather than likes, which is the shape of Metricool's finding too. Third, this is published mechanism rather than reverse-engineered guesswork, and that gap is the gap between a discipline and a superstition.

Here is what I cannot tell you, and I would rather say it than paper over it. LinkedIn published which signals go into the sequence and withheld the weights. It stated the intent to show more out-of-network content and withheld the magnitude. It announced expanded classifiers for low-quality AI posts on 30 July 2026 without publishing thresholds, accuracy, or even whether the rollout is a test. Two of my third-party sources sell things, so weigh them accordingly. What would settle the open questions is LinkedIn publishing signal weights, or publishing the real out-of-network share of the feed.

What does a machine-readable professional identity look like?

Not a niche in the marketing sense of picking a smaller pond. A subject you return to, holding a position you actually hold, repeated often enough that a system reading a thousand of somebody's past interactions can place you without straining. The test is embarrassingly simple. Read your last twenty posts in one sitting and write down the single sentence they all agree on. If you cannot find it, the machine cannot either, and no hook repairs that.

There is a second reason this matters now, and it comes from a measurement I find uncomfortable. Pangram scanned 1,002,627 posts across LinkedIn, Medium, Substack, X and Reddit, with data from 24 April 2026. LinkedIn was one third of the items scanned and 62% of all AI content flagged. More than 40% of LinkedIn longform came back as fully AI-generated, against a 13.8% average across those platforms and 4.4% on Reddit. State the caveat rather than bury it: Pangram sells AI detection and its sample came from an opt-in browser extension, so it is not a random draw. Discount it heavily and the direction still holds.

LinkedIn's own response says more than the study does. On 30 July 2026 it shipped the reporting button for AI slop and discontinued Enhance your post, its own AI writing feature, on the same day, replacing it with a proofreading tool that corrects grammar without altering voice. Its Chief Product Officer, Hari Srinivasan, said AI slop is a top priority for all of us. That is a company acknowledging its own product helped manufacture the problem it is now selling the cure for, and I respect shipping the admission rather than quietly deprecating the feature.

Follow those two facts to the same destination. A retrieval layer that matches meaning cannot tell you apart from the median if you write the median sentence, and the median sentence here is now frequently machine-written. Writing that obviously came from a specific human being is now a differentiated position by default rather than by effort, which is the least demanding advantage I have seen in nine years of this work.

It also explains why the shortcuts stopped paying. Oscar Rodriguez confirmed on record that LinkedIn removed the groups used to coordinate engagement pods and went after the automation tools, and that suspected pod activity limits how far a post travels outside the author's own network while it stays visible to existing connections. The trigger, he said, is a quid pro quo expectation, not asking a colleague to take a look. Notice what that penalty does. It leaves you visible to people who already know you and removes you from the pool of strangers, which is the only pool that grows anything.

If somebody handed me an account tomorrow and asked what to change first, this is the list, in this order.

The shift underneath all of this is not really technical, even though it arrived as an engineering post. For most of the last decade, being visible online rewarded range, because more surface area meant more chances to be found. A system that reads a thousand of a reader's prior actions and matches on meaning rewards the reverse. It rewards being one thing clearly enough, and long enough, that a machine can predict you.

That's the heart of everything. You are not writing posts for an audience any more, you are teaching a system what you are for, and every scattered post is a contradictory training example. Nine years and 2,500 articles into this, I find that clarifying rather than depressing, because the unglamorous discipline of saying the same true thing repeatedly finally has a published mechanism behind it.

So do not start with your next post. Open your profile, read your last twenty posts end to end, and write down the one sentence a stranger would say you are about. Then read your headline and ask whether it says the same thing, because that line is what the system reasons from when it knows nothing else. If those two sentences do not match, you have found the real problem, and it was never the hook.

Questions I get asked about this

Did LinkedIn actually change its algorithm in 2026?

Yes, and it documented the change itself. On 12 March 2026 LinkedIn's engineering blog published 'Engineering the next generation of LinkedIn's Feed', describing the replacement of four separate retrieval systems (chronological, trending, collaborative filtering and embeddings) with a single LLM based semantic retrieval layer, plus a new ranking model called the Generative Recommender that processes more than a thousand of a member's past interactions as an ordered sequence.

Is 360Brew the LinkedIn algorithm?

No. 360Brew is a research paper submitted to arXiv on 27 January 2025 by LinkedIn's Foundation AI Technologies team, describing a 150 billion parameter decoder-only model built on Mixtral 8x22B. What LinkedIn described shipping in March 2026 is a unified LLM retrieval layer plus a Generative Recommender. Related lineage, different artefacts. LinkedIn has never said 360Brew runs the feed.

Does dwell time affect LinkedIn reach?

Yes, and it is now confirmed by LinkedIn rather than inferred. The 12 March 2026 engineering post names the engagement types embedded into a member's interaction sequence as 'long dwells, likes, comments, shares', with long dwells listed first. LinkedIn has not published the relative weight of any of those signals.

Do the first 60 to 90 minutes decide 70% of my reach?

The mechanism is real and the number is folklore. Long dwells and early engagement are confirmed ranking inputs, but LinkedIn has never published a time window or that percentage, and no source carrying the claim shows a sample size or a method. Treat it as unsourced.

How often should I post on LinkedIn now?

Richard van der Blom's Algorithm Insights 2026, based on 1.3 million posts from 50,000 creators, cut the recommendation from five or six posts a week to two to four. His data shows reach down around 60% for active creators over two years, while inactive and occasional posters saw reach rise 10 to 15%. He sells LinkedIn consulting, so weigh it accordingly, but the sample is large and disclosed.

Sources

1. Engineering the next generation of LinkedIn's Feed, LinkedIn Engineering Blog, 12 March 2026.
2. 360Brew: A Decoder-only Foundation Model for Personalized Ranking and Recommendation (arXiv:2501.16450), arXiv, LinkedIn Foundation AI Technologies, 27 January 2025.
3. LinkedIn Study 2026, Metricool, 14 April 2026.
4. LinkedIn Organic Benchmarks 2026, Socialinsider, 2026.
5. Richard van der Blom on Algorithm Insights 2026, Creator Science episode 307, Creator Science, 2 June 2026.
6. LinkedIn adds a button to report AI-generated slop, TechCrunch, 30 July 2026.
7. LinkedIn is testing a 'seems like AI slop' reporting tool, Engadget, 30 July 2026.
8. AI in your feed, Pangram, 24 April 2026.
9. LinkedIn Premium subscription tops $1.7B in revenue, fueled by AI features, LinkedIn Newsroom, 7 March 2024.
10. LinkedIn Just Killed Engagement Pods. Here's What To Do Instead, Forbes (contributor), 18 March 2026.
11. Does LinkedIn intentionally limit the reach of posts with links, Matt Navarra, relaying LinkedIn Sr. Director of Product Management, September 2025.
12. LinkedIn updates its Feed algorithm, Social Media Today, 12 March 2026.
13. Microsoft FY2026 Q3 earnings, Microsoft, April 2026.

Every figure above is linked to its original source. Where I am predicting rather than reporting, I say so in the panels marked My take.