How the LinkedIn algorithm works in 2026, and what it actually rewards

The LinkedIn algorithm is a ranking system, not a broadcast system. When you publish, LinkedIn does not push your post to your followers. It adds the post to a pool of candidates, shows it to a subset of your connections and followers, measures what those people do (including how long they stop scrolling before moving on), and then either widens distribution or quietly lets the post die. LinkedIn's own help pages say the feed weighs “hundreds of signals” and that its systems “filter out or taper distribution of low-quality and unsafe content.” Nearly everything else you have read about the LinkedIn algorithm, including the famous test audience of exactly 100 people, is inference. Some of it is wrong.

Updated for the March 2026 feed rebuild. LinkedIn replaced its retrieval and ranking systems that month and published the architecture. This guide covers the fundamentals, which did not change. For what the rebuild specifically altered, see what actually changed in the 2026 algorithm, and for the widely repeated claim that a model called 360Brew now ranks your feed, see what 360Brew is and is not.

What does the LinkedIn algorithm actually do with a new post?

Start with what LinkedIn says in writing, because it is more specific than most people assume. On the help page Distribution of your content on LinkedIn, LinkedIn describes what happens the moment you post: it appears in the Activity section of your profile, and “it's shared with a subset of your connections and followers, depending on your connection strength, your connection's notification settings, and notification state.”

That one sentence carries three facts worth pulling apart. First, initial distribution is a subset, not your whole audience. Second, the subset is chosen by connection strength, so the people who already interact with you get seen first. Third, notification state is in the mix, meaning who happens to be on the app matters.

From there the ranking machinery takes over. LinkedIn's engineering team published a March 2026 post, Engineering the next generation of LinkedIn's Feed, describing the current architecture as two stages. A retrieval stage assembles candidates for each member. A ranking stage scores them. The retrieval stage used to run several separate systems (LinkedIn names “trending content, collaborative filtering, and embedding-based systems, each maintaining separate infrastructure”) and now runs a single dual encoder built on large language model embeddings.

The ranking stage is where the interesting change happened. LinkedIn calls the new model the Generative Recommender, and describes it this way: “Instead of scoring each post in isolation, GR processes more than a thousand of your historical interactions to understand temporal patterns and long-term interests.” The model is trained on both “passive tasks (click, skip, long-dwell) and active tasks (like, comment, share).”

Read that list again, because it is the closest thing to an official ranking objective anyone outside LinkedIn has. Six named outcomes: click, skip, long dwell, like, comment, share. Three of them are things a reader does without touching a button.

The retrieval stage is the part nobody writes about

Almost every algorithm guide describes ranking and stops. But a post that never enters the candidate set for a given member cannot be ranked at all, and retrieval is where the 2026 change is largest.

One detail from that engineering post is worth dwelling on because of what it implies. The team found that feeding raw engagement numbers into a language model worked badly, so they converted metrics into percentile buckets with special tokens, turning something like “views:12345” into a percentile token instead. That change alone produced a 15% recall improvement. LinkedIn's explanation: “LLMs don't inherently understand magnitude, so raw numerical features tokenize poorly and lose ordinal meaning.”

The takeaway for a writer is not about tokens. It is that your post is being represented as meaningplus a relative-performance percentile, and matched against a member's inferred interests. A post that is clearly about one identifiable subject gets a cleaner embedding than a post that gestures at four. Vagueness is now a retrieval problem, not only a reading problem.

Is there really a “test audience” for every LinkedIn post?

Yes and no, and the difference matters if you are trying to diagnose a bad week.

The version of the story you usually read goes: LinkedIn shows your post to about 100 people, and if a certain percentage engage within 60 or 90 minutes, it graduates to a wider audience, then a wider one after that. Numbers and thresholds vary by whoever is telling it. None of those numbers appear in any LinkedIn publication we could find.

What is documented is the shape, not the sizes. LinkedIn says the first audience is a subset chosen partly by connection strength. LinkedIn's ranking model is trained to predict skips and dwell, which only makes sense if early observed behaviour changes later distribution. And Sprout Social's breakdown of the LinkedIn algorithm describes the same mechanic in general terms: the algorithm tests distribution across portions of a network and expands or contracts it based on sustained interaction rather than a single window.

So the practical model is right. The arithmetic is invented. If someone tells you the threshold is 2% engagement in 90 minutes, ask where the number came from. It came from nowhere.

Which ranking signals does LinkedIn confirm it uses?

LinkedIn's help page How the Feed ranks content is short but precise: “Our AI systems and algorithms consider hundreds of signals to determine what content appears in each member's Feed.” It then names the families: the context of a post (LinkedIn's examples are “whether it's a helpful insight, job opportunity, or career milestone”), signals from your profile, signals from your network, and signals from your activity.

A second page, LinkedIn relevance, adds the part everyone argues about: “These algorithms also filter out or taper distribution of low-quality and unsafe content to enhance the value of what you see on your Feed.” That is LinkedIn confirming, in its own words, that reduced distribution is a thing it does deliberately. We come back to that in the section on shadowbans, and in more depth in why LinkedIn reach drops and how to diagnose it.

Here is the signal inventory, with the strength of evidence attached to each row. This is the part most algorithm guides skip, and it is the only part that lets you tell advice from guessing.

Do comments count more than reactions?

Almost certainly yes, and no published weighting exists. Anyone quoting you a ratio (“a comment is worth 7 likes”) invented it or inherited it from someone who did.

What can be said honestly is that comments carry more information and more downstream distribution. A comment takes seconds of composition, so it is a stronger attention signal than a tap. It also puts the post in front of the commenter's own network, which reactions do more weakly. And it produces replies, which produce more comments, which is the only self-reinforcing loop available to an organic post.

We chose to weight one comment as four reactions in our own scoring, and we chose that number because it is defensible, not because LinkedIn published it. What the data then showed is that comments discriminate between good and bad posts far more sharply than reactions do, which is at least consistent with them being scarcer and harder to earn.

What is dwell time, and why does it matter more than likes?

This is the most under-read document in the entire LinkedIn SEO field, and it is sitting on LinkedIn's own engineering blog. In Understanding dwell time to improve LinkedIn feed ranking, LinkedIn explains why it stopped treating clicks and likes as the whole picture.

LinkedIn defines two dwell measurements: time on the feed, counted when at least half of an update is visible while a member scrolls, and time after the click, counted once someone opens the content. It then lists three problems with using clicks and reactions alone. They are sparse (“click and viral actions can be rare, especially for passive consumers of the feed”). They are binary, so they carry no intensity. And they are noisy, because of “click bounces” where a member opens something and closes it immediately.

Dwell time fixed all three at once. LinkedIn calls it an “always measurable” and “real-valued measure of engagement.” Then they went further and built a model of skipping. They found a threshold below which members had effectively skipped an update, and trained a model to estimate the probability that your dwell time on a given post would fall under it. Posts with a high predicted skip probability get their ranking score reduced.

In offline testing the model improved area under the ROC curve by as much as 10%. In A/B testing, LinkedIn reported fewer skipped updates and more clicks and viral actions, meaning the passive signal made the active signals better too.

The writing implication is blunt. A post that makes 500 people stop scrolling for four seconds and then move on may do more for your distribution than a post that makes 40 people tap Like on their way past. Anything that buys attention inside the feed itself counts: a first line that creates a question, a document post that people swipe through, a short video that is watched rather than scrolled over. This is why the first 200 characters of a LinkedIn post carry so much weight. They are not just persuading a human, they are feeding a skip-prediction model.

Why does LinkedIn suppress virality by design?

Because the platform's stated objective is not attention, it is professional conversation, and those two things want different feeds.

The clearest evidence is a 2019 LinkedIn engineering post, Community-focused Feed optimization. LinkedIn added “contribution” as a ranking objective alongside click prediction, and explicitly optimised for what it called a creator-side effect: “While viewers are less sensitive to the freshness, content creators do care about receiving prompt responses.” The goal was that reactions be “given to a broader range of creators so that they have additional ways to engage with content on the platform,” and that all members “have a chance to feel heard.”

Translate that out of engineering language: LinkedIn deliberately spreads engagement across more creators instead of concentrating it on the ones already winning. A feed tuned purely for clicks converges on a small set of viral accounts. LinkedIn tuned against that.

The measured consequence shows up in third-party data. AuthoredUp's analysis of 476,465 personal-profile posts between September 2025 and February 2026 puts the median post at roughly 840 impressions. Not 84,000. Not 8,400. The median LinkedIn post is seen by fewer people than fit in a mid-sized conference hall, and 1,000 impressions already beats 56% of everything posted.

The same study found something that should reframe how you read viral posts: engagement rate falls as reach rises. Posts in the typical band (bottom 80%) took a median 496 impressions at a 2.86% engagement rate. Viral posts (top 5%) took a median 34,121 impressions at 0.89%. Reach multiplied roughly 69 times; the rate collapsed to under a third. Wide distribution means the post left the audience that cares.

If your business model needs 40 right people rather than 40,000 people, that is good news rather than bad. We unpack which metrics actually predict outcomes in the guide to LinkedIn impressions versus reach.

What do 12,988 posts say about what the algorithm rewards?

We took a public dataset of 34,012 LinkedIn influencer posts, filtered to 12,988 English posts from authors with at least 1,000 followers and at least 25 reactions, and scored each one by engagement rate: reactions plus four times comments, divided by the author's follower count. Then we compared the top decile (1,298 posts) against the bottom half (6,494 posts). The full method is in our 34,000-post LinkedIn engagement study.

The media mix is the part that speaks directly to distribution, because format is one of the few things the algorithm can evaluate before a single human reacts.

Two clean patterns. Native video over-indexes about 2.6x among top posts. Shared article and link posts over-index about 1.5x among the weakest. That is exactly the shape you would predict from a system optimising for dwell: video holds attention inside the feed, a link asks the reader to leave.

The gap that surprised us was on comments rather than reactions. The top-decile median post earned 235 reactions and 72 comments. The bottom-half median earned 80 reactions and 5 comments. Reactions differ by roughly 3x. Comments differ by more than 14x. Comments are the scarce signal, which is why we gave them a full comment strategy playbook of their own.

Two things people insist are algorithmic levers did nothing in our data. Hook length: the median visible hook was 206 characters for top posts and 205 for the bottom half. Numbers in the hook: 32.1% of top posts against 32.8% of weak posts. If a rule survives on repetition rather than evidence, this is usually how it looks under measurement.

Do personal profiles beat company pages in the LinkedIn algorithm?

Personal profiles win on engagement. Whether they win on reach depends entirely on which denominator you use, and this is one of the places where two large, honest studies appear to contradict each other.

Metricool's study of 673,658 posts across 63,108 accounts found personal profiles beat company pages on engagement by 63%, and generate 238% more comments per post. Company pages, meanwhile, got shared up to 17 times more often across all formats, which is a real advantage that almost never gets mentioned.

AuthoredUp's engagement rate analysis of 476,781 posts reports the opposite ordering: company pages at a 2.60% median engagement rate against 2.38% for personal profiles. The reason is the denominator. AuthoredUp divides by impressions, LinkedIn's own method for Pages. Metricool's comparison is follower-relative. Company pages get fewer impressions per follower, so dividing by impressions flatters them and dividing by followers does not.

Neither study is wrong. They answer different questions. “Which account type earns more attention per person who follows it” and “which account type converts an impression into a tap” are separate, and the second one is not the one your marketing plan cares about.

AuthoredUp also found engagement rate declines as follower count rises: a median 2.68% for profiles with 1,001 to 5,000 followers, dropping to 1.53% above 100,000. Large accounts reach further and convert worse. Anyone benchmarking their small account against a creator with 200,000 followers is comparing across a gradient.

Does the LinkedIn algorithm treat new or small accounts differently?

It treats them differently in retrieval and, on the evidence, more kindly in engagement rate than most people expect.

LinkedIn named cold start explicitly as a motivation for moving retrieval onto language model embeddings: the new system helps “especially” with “cold-start users and sparse content niches.” A collaborative-filtering system has nothing to work with when an account has no interaction history. An embedding system can still match a post about warehouse automation to people who work in warehouse automation.

AuthoredUp's engagement rate breakdown by follower count is the useful benchmark here, because it runs from tiny accounts upward.

The curve peaks in the 1,001 to 5,000 band and declines from there. Small accounts are not being throttled. They are converting attention better than large ones and simply getting less of it, which is what you would expect when initial distribution runs on connection strength and a small account has fewer strong connections to start with.

The practical consequence: at 800 followers your job is to earn a reply from the 40 people who see the post, not to chase impressions. Reach is downstream of that. If you want the equivalent numbers for your role, we broke the cohort out by audience in the LinkedIn benchmarks for founders and CEOs.

Which post formats does the algorithm distribute most in 2026?

Here the honest answer is that the big studies disagree, and knowing why is more useful than picking a winner.

The video row is the interesting one. AuthoredUp's analysis of more than 3 million posts from March 2025 to February 2026 puts video below baseline for reach and down 36% year over year. Socialinsider's benchmarks, drawn from 1.3 million posts on 16,645 business pages, put video engagement up 7%. Our own data, which comes from an older influencer dataset, shows video over-indexing 2.6x among top posts.

Three different populations, three different metrics, three different answers. AuthoredUp measures personal profiles by reach. Socialinsider measures business pages by interactions per impression. Ours measures influencer accounts by engagement per follower. A format can lose reach and gain engagement rate at the same time, which is precisely what happens when a platform floods a format and then shows it to fewer, better-matched people.

The two claims that survive all three datasets: documents and carousels over-perform relative to how rarely they are used (AuthoredUp puts them at under 5% of all posts), and link-led posts underperform everywhere. Everything else is population-dependent. We cover the link question in detail in do external links kill your LinkedIn reach.

How much do posting time and frequency really matter?

Less than the industry implies, and in a specific direction: frequency has a measurable effect, day of week barely does.

AuthoredUp's timing analysis of over 3 million posts found the gap between the best and worst day of the week is under 10% for personal profiles, with an 8am to 11am local window doing best. For company pages the day effect is stronger, with Wednesday roughly 1.7x Sunday. Their summary is worth quoting because it cuts against a whole genre of blog post: a carousel published at a mediocre hour usually beats a plain text update sent at the perfect time.

Frequency does more work. AuthoredUp found 4 to 5 posts a week produced a 2.60% engagement rate and 28% higher impressions per post than posting weekly. Buffer's analysis of over 2 million posts lands in the same region, recommending 2 to 5 posts a week. Note that these are correlations on creators who choose to post often, and people who post five times a week are not a random sample of people.

The mechanism is plausible independent of the numbers. Consistency builds the connection strength that decides who is in your initial subset. It also gives the ranking model more interaction history to work with, and the 2026 feed model explicitly reads over a thousand past interactions per member.

How long does a LinkedIn post keep getting distribution?

Days, not hours, and the tail is longer than the folklore suggests.

Metricool's study found roughly 40% of a post's interactions arrive on day one and about 50% of lifetime impressions land in the first two days. Flip that around: half the impressions arrive after day two. A post that looks dead at bedtime is often still being served the following afternoon.

LinkedIn's own analytics retention windows hint at the same thing. Its help page on post analytics says discovery and social engagement counts remain available for 1,000 days and members reached data for 400 days, which is not the retention schedule of a system that considers a post finished after 90 minutes.

In-network versus out-of-network is the number to watch

LinkedIn's post analytics report two figures most people scroll past: in-network impressions, defined as the “percentage of impressions from members who follow or are connected to you,” and out-of-network, the percentage from members who do not.

That split is the single best available proxy for whether the algorithm expanded a post beyond its starting subset. A post that is 95% in-network never left the room. A post that is 60% out-of-network was picked up by retrieval and pushed to strangers. If you are trying to work out whether a post “worked” algorithmically rather than socially, that ratio tells you more than the impression count does.

Two caveats from LinkedIn's own documentation. Impressions are described as an estimate that “may not be precise.” And your own views and engagements are counted toward your post's analytics, which quietly inflates small numbers.

What has changed in the LinkedIn algorithm recently?

Three documented changes, in reverse order.

March 2026: the Generative Recommender

LinkedIn replaced its multi-source retrieval stack with a single LLM-based dual encoder, and replaced independent post scoring with a sequential transformer that reads a member's interaction history as a sequence. LinkedIn's stated reason for the retrieval change is coverage: “LLM embeddings generalize beyond observed behavior, assessing latent interests from world knowledge where collaborative filtering and keyword matching fall short, especially critical for cold-start users and sparse content niches.”

That has a practical edge. A retrieval system built on collaborative filtering can only surface your post to people similar to those who already engaged with it. A system built on language embeddings can surface it to people whose stated interests match what the post is about, even with no engagement history to go on. Writing clearly about a specific subject is now a retrieval strategy, not just a readability preference.

November 2025: fairness testing published

In Putting members first, LinkedIn described two evaluations it runs before major model launches: creator allocation testing, which checks whether posts from different creator groups get systematically lower rankings when competing for the same slot at similar quality, and viewer quality-of-service testing, which checks whether some viewer groups get a worse feed. LinkedIn also used the post to make a point about why reach feels harder: content volume has grown, which increases competition for the same attention.

Ongoing: quality tapering

Both Hootsuite's and Sprout Social's current algorithm guides describe LinkedIn deprioritising low-value content built around engagement mechanics rather than original insight. Hootsuite's breakdown frames it as a spam and engagement-bait filter running before ranking. That is consistent with LinkedIn's own language about tapering low-quality content, though the specific implementations described in those guides are the authors' inference, not LinkedIn's statements.

How does the LinkedIn algorithm handle AI-generated content?

There is no published AI-detection ranking signal. Anyone telling you LinkedIn demotes posts it detects as AI-written is describing something no source confirms. We looked specifically for one, because we sell an AI writing product and would rather know.

What LinkedIn does have is authenticity rules and a quality taper, and those bite for reasons that have nothing to do with detection. The Professional Community Policies target spam, artificial engagement, and content that misleads. The relevance page commits LinkedIn to tapering “low-quality” content. Neither mentions how the text was produced.

The practical risk from generic AI writing is therefore mechanical rather than punitive: a post assembled out of interchangeable observations gives readers no reason to stop, so it loses on dwell and skip prediction, which are documented signals. The failure mode is the writing, not the tool. A post that reads like an internal memo written by nobody in particular fails whether a human or a model produced it.

The one thing we would treat as settled: disclosure norms are moving faster than detection. If your posts are drafted with AI, the durable defence is that they contain things only you know. Specific numbers, specific clients, specific mistakes. That also happens to be the profile of the top decile in our data, where first-person openers appear about twice as often as in the bottom half.

What does LinkedIn say it does not do?

Worth stating plainly, because the negative claims are more falsifiable than the positive ones.

• No demographic ranking.“Our algorithm and systems do not use demographic information, such as age, race, or gender, as a signal to determine the visibility of content, profile, or posts in the Feed.” LinkedIn adds that self-identification fields such as gender and pronouns are not signals in feed visibility.
• No pay-to-rank in organic.On the relevance page: “Feed distribution isn't influenced by payments from third parties to LinkedIn, except for promoted (paid) content, which is clearly labeled.”
• No published link penalty. LinkedIn has never confirmed one. That neither proves nor disproves the effect independent studies keep measuring; it just means the mechanism is unknown.
• No tolerance for coordinated engagement. The Professional Community Policies say: “Don't do things to artificially increase engagement with your content. Respond authentically to others' content and don't agree with others ahead of time to like or re-share each other's content.” LinkedIn says it may limit visibility, apply labels, or remove content, and that repeated violations can restrict the account.

Which LinkedIn algorithm “rules” are folklore?

Not all of these are false. They are unsourced, which is different, and worth flagging because they get repeated as if LinkedIn had published them.

• “The golden hour decides everything.”No published window exists. What is measured: Metricool found roughly 40% of a post's interactions arrive on day one and 50% of lifetime impressions land in the first two days. Early matters. Sixty minutes is a made-up boundary.
• “Never edit a post after publishing.” No source in any LinkedIn documentation.
• “Reply to every comment within the first hour or lose reach.” The 2019 engineering post does say freshness of responses matters to creators, and replies obviously generate more comments. The hard deadline is invented.
• “Use exactly three hashtags.” Our data shows four or more hashtags appears about 1.6x as often in bottom-half posts (41.9% vs 25.9%), which argues against stuffing, not for a magic number. More in do hashtags still work on LinkedIn.
• “LinkedIn caps you at one post per 24 hours.” No published cap. Frequency studies find more posting correlates with more total reach, though not linearly.
• “Dwell time is three seconds.” LinkedIn describes a skip threshold but has never published its value.

How do you optimize for the LinkedIn algorithm?

Everything above collapses into a fairly short list. Each item maps to a documented signal rather than a rumour.

1. Write the first 200 characters as if they are the whole post. The skip-prediction model watches whether people stop. Nothing else you do matters if they do not. Practical detail in how to start a LinkedIn post.
2. Buy dwell, not taps. A story with an unresolved middle, a carousel worth swiping, a list someone reads to the end. Dwell is measured on every impression; reactions are measured on the few percent who bother.
3. Engineer for comments specifically. Comments are the scarcest signal in our data by a wide margin, and Metricool found posts containing a question earned 77% more comments while posts with a clear call to comment earned 80% more.
4. Keep the post native. If a link must exist, put it in the first comment and make the post work with the link deleted.
5. Post on a schedule you can hold. Two to five times a week is where the large studies cluster. Consistency compounds connection strength, which decides your starting audience.
6. Be recognisably about something. LLM-based retrieval matches meaning. A feed of unrelated takes gives the model nothing stable to match you against.
7. Be present for an hour after you publish. Not because of a timer, because replying is the cheapest way to turn one comment into three.
8. Stop stuffing hashtags. Zero to three. Four or more travels with weak posts.
9. Do not buy or trade engagement. It is explicitly against policy and the downside is account-level.
10. Judge yourself on members reached and comments, not impressions. Impressions inflate with repeat views and LinkedIn describes them as an estimate.

A 30-minute audit you can run on your own account

You do not need a tool for this. LinkedIn gives every member the numbers, they are just buried one tap deep on each post.

1. Export or transcribe your last 30 posts into a spreadsheet. For each one record: date, format, whether it contained an outbound link, members reached, impressions, out-of-network percentage, reactions, comments.
2. Compute members reached divided by impressions. AuthoredUp measured a median ratio of about 46.5% across 42,493 posts. Much higher than that and your post is reaching new people once each. Much lower and the same audience is seeing it repeatedly.
3. Sort by out-of-network percentage, not by impressions. Your top five rows are the posts retrieval actually picked up. Read their first lines back to back. The pattern is usually obvious and usually not what you expected.
4. Group by format.Compare the median members reached for link posts against everything else. If the gap is large in your own account, you have your own answer to the link debate and do not need anyone else's study.
5. Compute comments per 1,000 members reached. This is the number that moves distribution and the number most people never look at. Our top-decile posts sat at a median of 72 comments against 5 for the bottom half.
6. Split the 30 posts into two blocks of 15 and compare medians, not averages. One viral post will wreck an average and tell you nothing about your baseline.

The output of that audit is a shortlist of two or three habits to change. It beats any general advice, including ours, because it is measured on the only audience that matters to you.

How do you tell whether the algorithm changed or your content did?

Almost every reported algorithm change is one of four things: a content shift you made and forgot, a format shift, an audience mismatch, or normal variance on a small sample. The distribution of LinkedIn post performance is heavily skewed, so a run of five flat posts after one good one is what the base rate looks like, not evidence of punishment.

The test is boring but it works. Pull your last 30 posts. Tag each by format, whether it contained a link, and whether the first line was first-person. Compare members reached (not impressions) across those groups, and compare the last 15 posts against the previous 15. If one format or one habit collapsed, that is your answer. If everything fell at once on the same date, look for a platform-wide report before assuming you were singled out. We walk the whole diagnostic in why did my LinkedIn reach drop.

One caveat on third-party reports of decline. Agorapulse's summary of Richard van der Blom's Algorithm Insights report cites views down 50%, engagement down 25%, and follower growth down 59%. That report is widely quoted in this niche. We could not access the primary document ourselves, so treat those figures as reported rather than verified. The direction is corroborated by AuthoredUp's 36% year-over-year drop in median video reach and by LinkedIn's own note that content volume has grown.

The LinkedIn algorithm, in one paragraph

LinkedIn retrieves candidates using language-model embeddings of what your post is about, ranks them with a model that reads each viewer's history and predicts six outcomes (click, skip, long dwell, like, comment, share), starts with a subset of your closest connections, tapers anything it scores as low quality, and deliberately spreads engagement across creators rather than concentrating it. The winning move under that system is not clever. Write something specific enough that a stranger stops scrolling, and interesting enough that a few of them argue with you in the comments.