How the LinkedIn Algorithm Works in 2026

LinkedIn’s feed ranker scores every candidate post by predicting several things you might do with it — like, comment, share, vote, long dwell and click — and combining those predictions linearly into one score. That is not an inference. It is in LiRank, LinkedIn’s own research paper, published February 2024. What LinkedIn does not publish is the coefficients, which means a weighting between engagement types genuinely exists and no public source can tell you what it is.

Start there, because it settles a whole genre of advice. Every article claiming a comment is worth five likes, or seven, or ten, is filling in a number LinkedIn has never released.

This page carries a different rule from most guides on this topic: every ranking claim below has a source, a date, and a confidence label. Where LinkedIn has published something, you get the quote and when it said it. Where only outside measurement exists, it is labelled as measurement and attributed to whoever measured it. All sources were read on 24 September 2026.

The Short Version

LinkedIn’s documentation of its own ranking is thin and old. The substantive engineering write-up on dwell time is from May 2020. The Help pages say “hundreds of signals” and name none. The one genuinely specific disclosure is a research paper from February 2024. Everything granular you have read about LinkedIn ranking is third-party measurement, not platform disclosure — and that distinction is the most useful thing on this page.

The honest asymmetry: Instagram’s head narrates ranking changes on camera. LinkedIn does not. So a LinkedIn guide that sounds as confident as an Instagram guide is telling you something it cannot know.

Table of Contents

• The Short Version
• What Is the LinkedIn Algorithm?
• How Thin LinkedIn’s Own Documentation Actually Is
• The Ranking Pipeline, Stage by Stage
• Dwell Time: The One Signal LinkedIn Explained Properly
• Signal, Source, Date, Confidence
• A Dated Changelog of What LinkedIn Has Actually Published
• Do Links Cost You Reach? What the Studies Measured
• Which Formats Perform Best — and Why We Will Not Give You a Number
• Eight LinkedIn Algorithm Claims That Do Not Survive Checking
• What Nobody Can Tell You
• How to Work With What Is Actually Documented
• Frequently Asked Questions
• Sources and Check Dates

What Is the LinkedIn Algorithm?

The LinkedIn algorithm is a multi-stage ranking system that predicts how each member will respond to each candidate post, then orders the feed by a combined score of those predictions. It is not one model. LinkedIn’s Feed AI team page describes first pass rankers running per content type, feeding a second pass ranker that merges everything into one personalised list, and states the system learns from implicit actions — clicks, likes, comments, shares, time viewed — rather than explicit ratings.

One line on that page is worth more than most ranking-factor lists: the system optimises for engaging interactive sessions, not per-post relevance at the top of the feed. The objective is the session, not the post.

LinkedIn’s Help page on how the feed ranks content adds the official framing: “Our AI systems and algorithms consider hundreds of signals to determine what content appears in each member’s Feed,” drawing on post context plus signals from your profile, network and activity. It also makes one commitment worth quoting — demographic information such as age, race and gender is excluded from content visibility decisions — and names not a single ordered signal.

How Thin LinkedIn’s Own Documentation Actually Is

This section exists because every other guide skips it, and skipping it is how confident nonsense gets published. Here is the complete inventory of substantive LinkedIn first-party material on feed ranking, as of 24 September 2026:

• One engineering post on dwell time, May 2020.
• One research paper, LiRank, February 2024 — genuinely the most specific disclosure any major platform has made about a feed objective function.
• An undated AI team page describing ranker architecture.
• Help pages that speak in generalities.
• One named executive statement about link reach, roughly September 2025, on a permalink that is not reliably readable without a logged-in session.

That is the lot. There is no LinkedIn equivalent of Instagram’s per-surface ranking explainer, no annual ranking post, and no executive who narrates changes the way Adam Mosseri does — a contrast our companion guide to how the Instagram algorithm works makes obvious.

So when you read that LinkedIn “now favours” some format, or that an action carries some percentage more weight, ask which of those five sources it came from. Usually none of them: it came from an independent study with the attribution stripped along the chain of articles quoting each other.

The Ranking Pipeline, Stage by Stage

LinkedIn’s published architecture has four stages, and each one is sourced below.

1. Candidate retrieval per content type. First pass rankers run separately by content type, according to LinkedIn’s Feed AI team page.
2. Merge into one list. A second pass ranker combines the per-type candidates into a single personalised ranking.
3. Predict several actions per post. From LiRank, LinkedIn’s ranking-models paper: the system predicts “multiple action probabilities including like, comment, share, vote, and long dwell and click” for each member and candidate post pair.
4. Combine linearly into a final score. LiRank again: “These predictions are linearly combined to generate the final post score.”

Two things follow from step 4, and both matter more than any tactic. A weighting exists — engagement types are not equal in the score, so anyone telling you they all count the same is wrong. And the weighting is unknowable from outside, because LinkedIn publishes no coefficients. That makes “a comment is worth N likes” not merely unsourced but structurally unanswerable from public information. LiRank reports the feed outcome of the work it describes as a 0.5% increase in member sessions, which is a useful reminder of how far real optimisation sits from the language of viral hacks.

Notably absent from LinkedIn’s published pipeline is the tidy spam-filter-then-test-pool ladder that most LinkedIn algorithm explainers describe. LinkedIn does act against engagement bait under its policies, but the four-stage funnel — filter, small test audience, scoring, wider distribution — appears in no LinkedIn source we could read.

Dwell Time: The One Signal LinkedIn Explained Properly

Dwell time is the one LinkedIn ranking signal with a real first-party explanation attached, and it is from May 2020 — not 2024, and not 2026. LinkedIn Engineering’s post on dwell time, by Siddharth Dangi, Johnson Jia, Manas Somaiya and Ying Xuan, defines two measurements:

• Dwell on the feed, which “starts measuring when at least half of a feed update is visible as a member scrolls through their feed.”
• Dwell after the click, meaning time spent on the content once opened.

That 50%-visible threshold is the most concretely actionable detail LinkedIn has ever published about its feed, and it is worth reading exactly as written. It is a viewport condition, not a statement about which half of your writing gets measured: the clock starts once at least half of the post’s card is on screen, and what accumulates from that moment is elapsed time. A post that is scrolled past before it is half visible accumulates nothing at all. LinkedIn says nothing about the first half of a post’s content — that inference is everywhere and it is not in the source.

LinkedIn is explicit about why it prefers dwell to clicks. Clicks are “not always measurable,” a “binary indicator,” “noisy,” and their positive signals “rather sparse” — a member can click, find the update irrelevant and bounce in seconds. Dwell is “always measurable” and a “real-valued measure of engagement.” The 2020 model predicted the probability of a skip, and the reported A/B result was a large decrease in skipped updates alongside more clicks and viral actions.

LiRank updates the mechanism in 2024: long dwell is “a ‘long dwell’ binary classifier predicting whether there is more time spent on a post than a specific percentile (e.g., 90th percentile)”, and that threshold is contextual — clustered by features including ranking position, content type and platform.

The practical consequence is the opposite of what most advice implies. There is no number of seconds to beat, by design. The bar is relative to comparable posts in comparable slots, which makes advice like “aim for seven seconds of dwell” meaningless. If you want to influence this signal, the lever is whether the post is worth staying with once it is on screen — which is why formatting is not cosmetic. Our free LinkedIn text formatter exists for that: line breaks and emphasis survive the composer, which is the difference between a scannable opening and a wall of text.

Signal, Source, Date, Confidence

Here is every LinkedIn ranking claim this page makes, with its provenance. First-party means LinkedIn, its engineering blog, its research, or a named executive on the record.

Note the two rows that are deliberately empty. They are the rows other guides fill in.

A Dated Changelog of What LinkedIn Has Actually Published

Almost every “LinkedIn algorithm update” article dates its claims wrong, so here is the timeline with real dates.

That last row is the honest state of 2026. We will not describe an architecture we could not confirm LinkedIn shipped, and any page stating it as fact has not opened a first-party source — because as far as we can establish, there is not one to open.

Do Links Cost You Reach? What the Studies Measured

LinkedIn denies intent, independent samples measure a real gap, and nobody has established the cause. All three of those are true simultaneously, and most articles report only whichever one suits their argument.

The denial is first-party and named: Rishi Jobanputra, Senior Director of Product Management at LinkedIn, has stated on the record that LinkedIn does not intentionally limit the reach of posts merely for containing an external link, qualified to posts that lead with value for the reader. We attribute that by name and title rather than quoting it, because LinkedIn feed permalinks are not reliably readable without a session, so we could not verify the wording ourselves.

The measurement is secondary, and substantial. From Ordinal’s study of the LinkedIn link penalty, which aggregates its own 2023–2026 series with other samples:

• Median reach loss around 26.5% for posts containing a link.
• A widening trend: roughly a 5% gap in 2023, roughly 42% by 2025.
• LinkPost, 358,000 posts: 448 median impressions with a link versus 705 without, a 36% gap.
• Richard van der Blom’s figures: about 11% loss with one link, 7% with two or three, 24% with four or more.

That last set is worth pausing on, because it is non-monotonic — two or three links measure better than one. A clean platform penalty would not behave like that, which is itself evidence that something other than a link rule is driving the numbers, and most write-ups quietly smooth it away.

Ordinal then says the decisive thing outright: “LinkedIn has never confirmed a deliberate downranking mechanism… this is an observed correlation.” The same study notes that no public research has cleanly separated whether the effect attaches to the URL or to the preview card — which makes the popular advice that pasting the link in the body removes the cost a hypothesis, not a finding.

The usable conclusion: if reach on a post matters more to you than the click, the measured gap is a real reason to keep the link out of it — a decision about measured odds, not obedience to a rule LinkedIn says does not exist.

Which Formats Perform Best — and Why We Will Not Give You a Number

LinkedIn has never published anything about how post format affects ranking, so every format reach delta in circulation is outside measurement. No first-party LinkedIn source we could read ranks text against image against video against document, or states a reach multiplier for any of them.

The numbers you have seen almost certainly originate from Richard van der Blom’s Algorithm Insights Report, the most-cited independent LinkedIn study in existence. The 2026 edition is a seventh annual edition of roughly 200 pages, stating a sample of about 1.3 million posts from around 50,000 creators. That is a serious undertaking, and it deserves a serious assessment:

• What is stated: a named author and a headline sample size, which already puts it well above the typical algorithm blog post.
• What is not available: how posts were sampled, whether comparisons were controlled, and how individual per-feature percentages were derived. The report itself was not fetchable in our pass, so even the sample figures come from the author’s own promotion rather than the document.

So we cite it as a large observational study by a named researcher and do not reproduce its per-format percentages as if they were ranking weights. If you use those figures, attribute them to him by name — never to LinkedIn. They circulate everywhere with the attribution stripped, which is how they end up read as platform disclosure.

Two things you can act on without a fabricated number. Whatever the format, the documented signal is attention: dwell starts at 50% visible and long dwell is judged against comparable posts. And timing is an audience question rather than a ranking signal, with its own evidence base — our guide to the best time to post on LinkedIn covers it separately for exactly that reason.

Eight LinkedIn Algorithm Claims That Do Not Survive Checking

This is the section neither of the top-ranking LinkedIn algorithm guides has. Each claim below was found asserted as fact, usually without attribution.

Two of those errors are pure date failures, which is worth noticing: a large share of what is wrong in this topic is not invented so much as undated, then re-dated by whoever refreshed the article for a new year.

What Nobody Can Tell You

Publishing the gaps is part of doing this honestly. As of 24 September 2026, these have no published answer:

• The weight of comments versus reactions. LiRank confirms both are predicted and combined linearly, and releases no coefficients.
• Any early-engagement window. Checked across Help, the Feed AI page, the dwell-time post and LiRank. None names one, and the term is not LinkedIn’s.
• Whether the measured link gap is caused by LinkedIn at all, and whether it attaches to the URL or the preview card. The study itself says the question is open.
• Per-claim methodology for the most-cited independent report, and confirmation of its sample from the document rather than from its promotion.
• Whether LinkedIn shipped LLM-based feed ranking in 2026, and what such a system reportedly downranks.
• Whether AI-generated content is downranked. No ranking statement exists from LinkedIn on this.

If a guide answers one of those with confidence, that confidence is the problem.

How to Work With What Is Actually Documented

No promises here, because LinkedIn publishes no weights and anyone forecasting a reach outcome from a ranking signal is guessing. These are practices that follow directly from something LinkedIn has documented, with the source named so you can judge each one:

• Hold the scroll once the post is on screen. From the 2020 dwell definition: the clock starts once at least half of the post’s card is visible, and what it records from then on is time spent. A post scrolled past before it is half on screen accumulates no dwell at all. LinkedIn’s definition is a viewport condition, so it says nothing about which half of your content is measured.
• Stop chasing a dwell target. From LiRank: the long-dwell threshold is contextual by design, so there is no seconds number to hit.
• Expect a reach cost when you include a link, and decide whether the click is worth it. From the measured gap, not from a platform rule — LinkedIn says there is no deliberate rule.
• Give people more than one way to respond. From LiRank’s list of predicted actions: like, comment, share, vote, long dwell and click are each predicted separately and then combined. LiRank does not rank them and publishes no coefficients, so this is not a claim that a reply beats a reaction — only that a post earning several of those actions feeds more of the predictions the score is built from than one that earns a scroll-past.
• Drop the obsolete controls. Creator Mode and profile hashtags are gone. In-post hashtags still work as topic and search signals rather than reach levers, covered in our guide to LinkedIn hashtags; for a starting set, our free LinkedIn hashtag generator is ungated.
• Date every claim you act on. Two of the most repeated LinkedIn ranking claims are misdated sources, one by over four years.

Running this across several networks is where the time goes, and it is where what we build fits. Outfeed AI is chat-first rather than dashboard-first: describe a post once and the assistant drafts a version per platform in your brand voice, previews each, then schedules or publishes across nine networks from one conversation. Starter is $29/month flat, Pro is $59/month flat with 3 team members, and per-seat billing starts only at the $149 per seat Agency tier. Start a free 3-day trial — 100 AI credits, card required.

Frequently Asked Questions

How does the LinkedIn algorithm work in 2026?

LinkedIn’s own research paper LiRank, published February 2024, describes the production feed ranker as predicting several action probabilities for each member and post pair — like, comment, share, vote, long dwell and click — then combining those predictions linearly into one score. Upstream, first pass rankers run per content type and a second pass ranker merges them into a single personalised list. LinkedIn does not publish the coefficients, so a weighting exists and is unknowable from public material.

Does LinkedIn penalise posts with external links?

Both things are true and most articles only tell you one. LinkedIn denies intent: Rishi Jobanputra, Senior Director of Product Management, has stated on the record that LinkedIn does not deliberately limit reach for posts containing a link. Independent samples nonetheless measure a real gap — Ordinal reports a median reach loss around 26.5%, and a 358,000-post LinkPost sample found 448 versus 705 median impressions. The gap is well measured, the cause is not established.

Is dwell time a LinkedIn ranking factor?

Yes, and LinkedIn documented it in May 2020, not recently. LinkedIn Engineering measures dwell on the feed, which starts counting once at least half of an update is visible as you scroll, and dwell after the click. The 2024 LiRank paper adds that long dwell is a binary classifier against a percentile threshold that is set contextually by ranking position, content type and platform, so no universal seconds target exists.

Is a comment worth more than a like on LinkedIn?

Nobody outside LinkedIn can say — not even the direction. LiRank confirms the shape of the answer — the ranker predicts like, comment, share, vote, long dwell and click probabilities and combines them linearly — but it lists those actions without ranking them, and LinkedIn publishes no coefficients. So every specific multiplier you read, whether 5x, 7x or 10x, is invented, and so is the assumption that comments come out above reactions at all. The fact that the number differs between articles is the giveaway.

Is there a LinkedIn golden hour?

Not in anything LinkedIn has published. The golden hour is creator folklore, not LinkedIn terminology, and no LinkedIn source defines or confirms any time window — not its Help pages, not the Feed AI team page, not the dwell-time post, not LiRank. Early engagement plausibly matters because the ranker learns from engagement, but the specific 60 or 90 minute cliff has no published basis.

Does turning on LinkedIn Creator Mode increase your reach?

No, and the toggle no longer exists. LinkedIn retired Creator Mode in March 2024 and removed profile hashtags and topics in February 2024. The features it used to gate — LinkedIn Live, Audio Events, Newsletters, creator analytics, the follow link and custom profile buttons — remain available without it. It was never a ranking boost, so advice to switch it on describes a control that is gone.

Did LinkedIn change its algorithm in 2026?

No 2026 change could be verified first-party. Reports circulated during 2026 of an overhaul using large language models for retrieval and ranking, but no LinkedIn engineering post or paper confirming it could be read, and one key report was unreachable. LiRank, from February 2024, remains the newest verified first-party disclosure about how the feed ranks. Treat any confident 2026 architecture claim as unverified.

Sources and Check Dates

Every claim on this page traces to one of the following, all read on 24 September 2026. Where no source exists, the text says so.

• LiRank: Industrial Large Scale Ranking Models at LinkedIn (Borisyuk et al., LinkedIn), 10 February 2024, revised 7 August 2024. Source for the predicted actions, the linear combination, the long-dwell percentile classifier and the reported session lift.
• LinkedIn Engineering — Understanding dwell time to improve feed ranking, 12 May 2020, by Siddharth Dangi, Johnson Jia, Manas Somaiya and Ying Xuan. Source for both dwell measurements, the 50%-visible threshold, the clicks-versus-dwell reasoning and the skip model.
• LinkedIn Engineering — Feed AI, undated team page. Source for first and second pass rankers, learning from implicit actions, and the session-level objective.
• LinkedIn Help — How the Feed ranks content, page undated. Source for “hundreds of signals”, the profile/network/activity framing, the demographic exclusion, and the Top versus Recent control.
• Ordinal — LinkedIn link penalty study, 2026. Source for the 26.5% median, the 2023-to-2025 trend, the LinkPost 448 versus 705 figures, the van der Blom per-link percentages, and the stated limitation that this is an observed correlation.
• LinkedIn Help — Updates to Creator Mode, February and March 2024, for the retirement of the toggle and of profile hashtags. Rishi Jobanputra, Senior Director of Product Management, LinkedIn, roughly September 2025, for the link-reach statement — attributed by name, not quoted, because the permalink is not readable without a session. Richard van der Blom, Algorithm Insights Report 2026, cited by name as a large observational study with partially available methodology.

Deliberately not stated here: any engagement multiplier; any dwell-time seconds target; any golden-hour window; any per-format reach percentage presented as a ranking weight; any 2026 LLM ranking architecture. None of those is established, and this page will not supply them.

Re-verification: this page is re-checked against primary sources every 90 days. The load-bearing sources to watch are the 2020 dwell-time post, which is overdue a successor, and any new LinkedIn engineering publication on feed ranking — that is where a genuine 2026-or-later change would be confirmable.