How the LinkedIn algorithm works in 2026, and what actually moves reach
Contents
- What LinkedIn rebuilt, in its own words
- Why your reach fell, and why the published numbers disagree
- The four things the ranking model reads: the A.U.T.H. framework
- Formats in 2026: carousels won, video slipped
- The external link rule reversed this year
- How often to post, and what daily posting costs you
- When to post: two large studies that contradict each other
- Does LinkedIn downrank AI-written posts?
- What to measure instead of impressions
- A four-week plan you can start Monday
- What I don't know
- FAQ
- Sources
On March 12, 2026, a LinkedIn engineer named Hristo Danchev published the most detailed public account the company has ever given of how its feed decides what you see. I have been on LinkedIn since April 2004, its first year, and this is the largest ranking change I have watched.
It also explains why your reach collapsed, a figure circulating in a dozen conflicting versions. What follows is what the evidence supports, what it does not, and what a B2B leader does with it Monday morning. Every number is attributed, and where studies disagree, I say so.
In short: LinkedIn replaced its feed pipeline with a 150-billion-parameter foundation model called 360Brew, plus a transformer ranking model that reads over 1,000 of your prior interactions. Reach per post is down across every independent study, but engagement per post is up 13.90%. Two pieces of 2025 advice stopped being true in 2026: external links now cost personal profiles reach, and video is no longer the format to bet on.
What LinkedIn rebuilt, in its own words
LinkedIn's Foundation AI Technologies team published the 360Brew paper on arXiv in January 2025 (lead author Hamed Firooz): 150 billion parameters, fine-tuned on LinkedIn's own member activity, handling more than 30 prediction jobs that used to need separate models. A team of 23 people built it in nine months.
One detail matters to anyone reading this from Europe and almost nobody reports it: the model was trained on first-party data from members outside the European Union. LinkedIn has not published what that boundary implies for EU members' feeds.
Selection now works by meaning: every post and member becomes a numerical fingerprint, and the system pulls the closest matches out of millions of posts in under 50 milliseconds. Ordering runs on the Generative Recommender, which reads more than 1,000 of your past interactions before ranking. Reported engineering gains: 3.6% better retrieval, a 30x correlation improvement, 37% less memory used, trained on eight graphics processors.
Dwell time sits alongside likes and comments as a first-class signal: a fast like is a noisier indicator of value than a long read, LinkedIn argued years earlier. LinkedIn moved from a relationship graph to an interest graph: your follower count is no longer a distribution guarantee, the model decides who your post is useful to.
Why your reach fell, and why the published numbers disagree
There is no single reliable figure for the reach decline, and anyone quoting one to the decimal is guessing.
| Study | Sample | Reported decline |
|---|---|---|
| Richard van der Blom, Algorithm Insights 2026 | 1,300,000 posts, 50,000 creators | Reach down 60% for active creators over two years |
| Richard van der Blom, Algorithm Insights 2025 | 1,800,000 posts | Views down about 50%, engagement down about 25%, follower growth down about 59% |
| AuthoredUp | 3,000,000+ posts | 98% of accounts declined, median impressions down 47% mid-2024 to mid-2025 |
| Saywhat | 329,000+ posts | Median impressions down 66% from the 2023 peak |
| Metricool, LinkedIn Study 2026 | 673,658 posts, 63,108 accounts | Average impressions down 10% year over year |
Five serious studies, five different answers, from down 10% to down 66%. Van der Blom's own report shows the decline far steeper at the top: accounts above 50,000 followers saw reach fall 62%, engagement 67% and follower growth 83%.
The number I would build a plan around runs the other way: Metricool measured likes down 13%, comments down 17% and shares down 10%, while total engagement rose from 12.21% to 13.90% and clicks rose 4.90% per post.
Fewer people see your post; the ones who do, do more with it. If you have been measuring impressions, you have been watching the metric the platform stopped optimizing.
The four things the ranking model reads: the A.U.T.H. framework
We publish this framework in French under the same four letters, a reading grid mapping four properties of a post onto signals the ranking model consumes.
A: Authority on a narrow topic. The interest graph ranks topical fit before network proximity. Van der Blom's 2026 data, via consultant Melanie Goodman (May 2026), puts the effect of holding two to three core topics for ninety days at +27% reach and +41% topic-based follows: 8,000 focused followers can beat 80,000 unfocused ones.
U: Usefulness that survives the scroll. Dwell time plus saves. Independent analyses put a save at roughly 5x the weight of a like and 2x a comment, and van der Blom's data shows carousels hitting a 29.2% save rate, the highest of any format.
T: The first ninety minutes. Metricool measured that 50% of a post's total impressions land within the first 48 hours. Van der Blom's 2025 guidance: most reach is determined in the first 90 minutes, and engaging with other people 15 to 30 minutes before and after you publish lifts reach by up to 20%.
H: Human proof. Generic writing is not penalized because a machine wrote it, but because it produces no dwell time and no saves. Van der Blom's 2026 report puts AI signals in 44% of LinkedIn content, with the reach penalty covered below. The scarce thing is not writing, it is a number, a date and a name that only you have.
Formats in 2026: carousels won, video slipped
Video was the format every 2025 LinkedIn guide told you to bet on, advice that aged badly in about ten months: Metricool's 2026 study, on 673,658 posts from 63,108 accounts, found video now the most-used format on personal profiles at 40.11% of posts, yet it underperforms carousels, images and multi-image posts on impressions and interactions.
| Format | What Metricool measured in 2026 | Verdict for a B2B leader |
|---|---|---|
| Carousel (document) | 1,451 average impressions per post; 11x the interactions of a single image | The format to default to |
| Multi-image | Highest engagement rate measured, 3.71% | Cheap to produce, badly underused |
| Poll | Highest average impressions on Company Pages, at under 1% of posts published | Works because it is rare; stops working if you repeat it |
| Image | Roughly half of all Company Page posts | Reliable, undifferentiated |
| Text only | 1.75 average comments per post | Fine when the argument carries it |
| Video | 605 average impressions; 1.80% engagement rate | Highest effort, currently the weakest return |
Van der Blom's own cut agrees on direction: documents at +6.92% reach and +46.73% engagement, native articles at +47.92% reach. One caution: the poll's top ranking is a scarcity effect that fades once copied.
The carousel wins because it is the only common format that physically requires dwell time: a reader cannot swipe eight slides in half a second.
The external link rule reversed this year
This deserves its own section: the most confidently repeated wrong answer on LinkedIn, and I was on the wrong side of it last year.
Through 2025, study after study reported that external links no longer cost you reach. The 2026 data says something more specific: it depends entirely on which kind of account you are posting from.
| Account type | Impact of an external link in the post body (Metricool, 2026) |
|---|---|
| Personal profile | 27% fewer impressions, 20% fewer interactions |
| Company Page | 51% more impressions, 41% more interactions |
Van der Blom's 2026 report lands in the same place from a different sample: one external link in the post body cuts median reach by 18.8%.
A third study complicates it: Saywhat, on roughly 400,000 posts, found posts with three or more links performed considerably better than link-free ones, and AuthoredUp reports the same shape, a single link worst, several better.
The most plausible reading: a lone link reads as a departure, a resource list reads as substance.
What I actually do: on a personal profile, keep the link out of the body, put it in the first comment, and say so in the post. On a Company Page, put the link in the body. This flipped inside twelve months and can flip again.
How often to post, and what daily posting costs you
Daily posting is expensive bad advice, and now there is a price tag on it.
Van der Blom's 2026 data, reported in May 2026, puts daily publishing at a 26% drop in average reach per post, worsening toward a 45% penalty over time. The opposite move: going from one post a week to two or four adds roughly 1,234 impressions per post.
Metricool's measured reality: personal profiles averaged 3.05 posts per week in 2026, Company Pages 2.34, posting frequency fell about 10% while engagement rose.
Two to three posts a week from a personal profile, at least 24 hours apart, outperforms the alternative.
Saywhat found the top 1% of creators leave 286 comment replies per week against 34 for the average creator, a 741% gap, with a 0.545 correlation between commenting frequency and follower growth. Correlation, not cause, but with three hours a week, spend one writing and two replying.
When to post: two large studies that contradict each other
Most articles hand you a table of golden hours. Here are two of the largest datasets available, and they disagree almost perfectly.
| Source | Sample | Best window found |
|---|---|---|
| Sprout Social, 2026 | Aggregated platform data | Mid-morning: Tuesday 11 a.m. to 5 p.m., Wednesday 11 a.m. to 4 p.m., Thursday 11 a.m. and 1 to 5 p.m. |
| Buffer, 2026 | 4,800,000 posts sent through Buffer, normalized to each account's local time zone | Late afternoon and evening: 3 p.m. to 8 p.m., peaking Wednesday at 4 p.m. |
Both agree on Wednesday and that weekends are dead. The hour of day is five hours apart.
A global average posting time is close to meaningless for your audience.
What I do instead: open your LinkedIn analytics, check when your own followers were active over the last 90 days, pick two windows inside that, and hold them for a month. If your buyers sit on the US East Coast, the two studies overlap in the late morning and again around 4 p.m. Eastern.
Does LinkedIn downrank AI-written posts?
Not for being AI-written. Yes for being generic. LinkedIn has not published a policy penalizing assisted writing, only a ranking model that reads dwell time, comment depth and repeated templated patterns, blind to who held the pen.
Penalties on generic content are real and independently measured. Van der Blom's 2026 report has fully generated posts taking 2.8x less reach and close to 5x less engagement. SocialPilot, updating on August 5, 2026, reports low-effort AI content receiving roughly 30% less reach and 55% less engagement.
I produce with AI and say so. What I do not do is publish a paragraph that could have appeared under anyone else's name. The 41% of long-form LinkedIn posts that Pangram Labs measured as fully generated in July 2026 are not failing because a model wrote them, they are failing because nothing in them is checkable.
The longer argument, including why LinkedIn's reporting button measures the wrong property of a text, is in the AI slop piece.
What to measure instead of impressions
LinkedIn added saves and sends to post analytics in late 2025. Almost nobody in the B2B teams I talk to has changed their reporting since, the cheapest fix available this quarter.
| Metric | What it tells you | Target for a B2B leader |
|---|---|---|
| Save rate | Reference value: someone wants this back later | Above 1% of views |
| Sends (private shares) | Someone put their own credibility on your post inside their company | Rising month over month |
| Substantive comments (over five words) | A real professional conversation started | Above 2% of views |
| Profile clicks | Curiosity converted into a look at who you are | Above 3% of views |
| Connection requests after a post | The audience you attracted, not the one you have | Rising month over month |
Sends are the one I would watch hardest. A post forwarded privately inside a buying committee is worth more than four hundred likes from people who will never buy anything. We track how these numbers move quarter by quarter on the All In blog.
A four-week plan you can start Monday
No tool required: the cadence I run and teach, rebuilt around what the 2026 evidence rewards.
Week 1, pick your two topics. Write down the two subjects you want to be the obvious answer on, not five. Check your last twelve posts against them: everything outside those two topics costs you topical fit under the interest graph.
Every week, one proprietary-data post. A number nobody else has, your win rate on inbound versus outbound, or the real reply rate on a sequence you ran last quarter. Carousel, eight to twelve slides. Your save engine.
Every week, one field-observation post. Something you saw this month, dated and located: "on a tender in Austin" can be pictured and checked, "recently, with a client" is filler. Text or text plus image, 900 to 1,200 characters. Your comment engine.
Every other week, one position post. A stance a competent peer could argue against. Text only, your send engine.
Every other week, one human post. The person behind the title, including the expensive mistake with the amount attached. Photo plus text.
The part most people skip: book two blocks of 45 minutes a week to reply to other people's posts, 35 to 120 words, on four to seven different profiles. We break the cadence down week by week on the All In podcast.
What I don't know
This section is where the piece should be judged.
I don't know the actual weights. Nobody outside LinkedIn does. The save-at-5x-a-like figure is an inference from observed post performance, published by practitioners, not a coefficient anyone has seen. I use it because it is directionally consistent across several samples, and I would drop it tomorrow if LinkedIn published something better.
I don't know why the studies disagree so widely on reach decline, from 10% to 66%. Sampling explains some of it. Follower-tier composition explains more, since the drop is far steeper above 50,000 followers. It does not explain all of it.
I don't know what 360Brew's training-data boundary means for European members, and the paper does not say.
I don't know whether the external-link finding holds. It flipped between 2025 and 2026, and one credible study on 400,000 posts still reports the opposite for multi-link posts. Anyone telling you this is settled has not read all three.
And I have no way to verify van der Blom's 2026 figures at source: the Algorithm Insights report is paywalled, and the numbers in this piece reach me through his public statements and secondary coverage dated May and June 2026. That is a real limitation and I would rather write it than hide it behind a footnote. As of August 5, 2026, SocialPilot reports no confirmed ranking change announced for August. If that changes, it will be covered here and in the weekly All In newsletter.
FAQ
How does the LinkedIn algorithm actually work in 2026?
LinkedIn uses two rebuilt systems, described publicly on March 12, 2026. A dual-encoder retrieval layer built on language-model embeddings selects candidate posts from millions in under 50 milliseconds by reading meaning rather than matching keywords. A transformer ranking model, the Generative Recommender, then orders them using more than 1,000 of the member's past interactions, dwell time, engagement counts, profile and work history, post format and device type.
Why did my LinkedIn reach drop so much?
Because distribution moved from a relationship graph to an interest graph: follower count no longer guarantees delivery. Independent studies put the decline between 10% and 66% depending on sample and period, with the steepest drops on accounts above 50,000 followers (down 62% per van der Blom's 2026 data). Engagement per post moved the other way, rising to 13.90% in Metricool's 2026 study.
What is the best post format on LinkedIn in 2026?
Carousels, also called document posts. Metricool's 2026 study on 673,658 posts found they generate 11 times more interactions than single images and average 1,451 impressions per post against 605 for video. Van der Blom's data shows a 29.2% save rate on carousels, the highest of any format. Video is now the most-used format on personal profiles and one of the weakest performing.
Do external links still reduce LinkedIn reach?
For personal profiles, yes, and that reversed in 2026. Metricool measured 27% fewer impressions and 20% fewer interactions on personal-profile posts containing a link, while Company Page posts with links gained 51% impressions. Van der Blom separately reports an 18.8% median reach cut from one link in the body. A Saywhat study of 400,000 posts contradicts this for multi-link posts, so treat it as unsettled.
How often should a B2B executive post on LinkedIn?
Two to three times per week from a personal profile, at least 24 hours apart. Metricool measured personal profiles averaging 3.05 posts weekly in 2026. Daily posting carries a reported 26% drop in average reach per post, worsening toward 45% over time as audience fatigue builds. Time saved is better spent replying to other people's posts.
Does LinkedIn penalize AI-generated content?
Not for being AI-assisted. LinkedIn's ranking model measures how long people read and how deep the replies go, neither of which detects authorship. Generic content is downranked whether a human or a model produced it: van der Blom's 2026 report shows fully generated posts taking 2.8x less reach, and SocialPilot reports roughly 30% less reach for low-effort AI content. An edited draft carrying your own data and voice performs like any other good post.
What LinkedIn metrics should I track instead of impressions?
Saves, sends, substantive comments, profile clicks and post-publication connection requests. LinkedIn added saves and sends to post analytics in late 2025 and most B2B teams have not updated their reporting. Sends matter most for a sales organization: a post forwarded privately inside a buying committee signals more commercial intent than several hundred public likes.
Sources
- LinkedIn Engineering, Hristo Danchev, "Engineering the next generation of LinkedIn's Feed", March 12, 2026: retrieval and ranking architecture, signals used, measured engineering gains.
- PPC Land coverage of the same post, March 2026: sub-50ms latency, 1,000+ historical interactions, recall and memory figures.
- Firooz et al., "360Brew: A Decoder-only Foundation Model for Personalized Ranking and Recommendation", arXiv 2501.16450: 150B parameters, Mixtral 8x22 base, 30+ tasks, training data from members outside the EU, cold-start performance.
- LinkedIn Engineering, "Understanding feed dwell time": why dwell time is used as a ranking signal.
- Metricool, 2026 LinkedIn Study press release, April 2026: 673,658 posts, 63,108 accounts, likes down 13%, comments down 17%, shares down 10%, engagement up nearly 14%.
- Metricool, "LinkedIn Trends: 6 Strategy Insights from Our 2026 Study", April 16, 2026: format impressions and engagement rates, link impact by account type, 50% of impressions inside 48 hours, posting frequency.
- Net Influencer coverage of the Metricool 2026 study: personal profiles versus Company Pages, links, hashtags, questions and comment CTAs.
- Richard van der Blom, Algorithm Insights 2026 (paywalled report): 1,300,000 posts, 50,000 creators, reach down 60% for active creators over two years.
- Melanie Goodman, "LinkedIn Algorithm 2026: Why Your Reach Dropped", May 7, 2026: secondary reporting of van der Blom's 2026 figures on links, daily posting, topic authority and feed composition.
- Kara Redman, "I Read Every LinkedIn Algorithm Report So You Don't Have To": side-by-side comparison of van der Blom, Saywhat, AuthoredUp, Socialinsider, Trust Insights and Ocean Labs, including save rates, comment behavior and the multi-link contradiction.
- Sprout Social, "Best Times to Post on LinkedIn in 2026": mid-morning windows, Tuesday through Thursday.
- Buffer, "Best Time to Post on LinkedIn in 2026", 4.8 million posts: 3 p.m. to 8 p.m. windows, Wednesday 4 p.m. peak, local time zone normalization.
- Pangram Labs, "AI in your feed", July 9, 2026: 41% of LinkedIn posts over 250 words measured as fully AI-generated, on a declared sample of 1,002,627 posts collected through an opt-in browser extension.
- SocialPilot, "LinkedIn Algorithm June 2026: What Actually Drives Reach Now", updated August 5, 2026: AI content reach penalties, dwell time engagement bands, no confirmed ranking change announced for August 2026.
All In: publish against the mechanism, not against last year's tactic
Every ranking tactic published in 2025 had a shelf life of about ten months. What the model measures, attention and usefulness, has not moved since LinkedIn started measuring it, and it is the only thing worth building a content calendar on.
All In is the B2B media that decodes LinkedIn, expert blog, weekly podcast and newsletter for SME leaders and sales directors who want to turn LinkedIn into measurable growth. An original creation by Patrick de Carvalho, on LinkedIn since 2004. Motto: "I Never Lose."
Discover All In and get what matters on LinkedIn, every week.