Part two · The platforms · 05 of 08
LinkedIn: the only feed whose ranking system is published in academic papers
The only platform here that publishes its ranking system in peer-reviewed papers, naming the model, the objectives and the candidate count. It is also the platform whose most repeated marketing claim has no source at all.
-
1.3bn
Members, per LinkedIn's newsroom. Its own About page says 1 billion. Neither is dated
LinkedIn 1
-
512
Candidate posts the ranker scores for each member request
LinkedIn, Feb 2026 2
-
2
Named headline objectives: Long Dwell and Contribution
LinkedIn, Feb 2026 2
-
+12%
LinkedIn revenue growth, Q4 FY26, 10 percent in constant currency
Microsoft, 29 July 2026 3
-
5.20%
Engagement rate by impressions, business Pages only
Socialinsider, 2026 4
-
The link penalty has no source
LinkedIn has never said links are penalised. A named LinkedIn product director says on record they are not. 5
-
LinkedIn tried an LLM ranker and rejected it
Its own paper: "The LLM-Ranker never achieved superior online performance." 2
-
No stated format preference
LinkedIn's Feed product lead tells members to "play with all the different creation types." 6
-
Three member numbers, no dates
1.3 billion, 1 billion and 1.2 billion appear across LinkedIn's own properties. None of the pages is dated. 1 , 2
- LinkedIn, About Us (news.linkedin.com) and About page (about.linkedin.com). Neither carries a date. Retrieved 8 August 2026. 1st
- Hertel et al. (LinkedIn), An Industrial-Scale Sequential Recommender for LinkedIn Feed Ranking , arXiv 2602.12354, 12 February 2026, revised 29 May 2026. 1st
- Microsoft, FY26 Q4 press release and earnings call, 29 July 2026. 1st
- Socialinsider, LinkedIn Organic Benchmarks 2026 . 1.3m posts, 16,645 business Pages. 2nd
- Rishi Jobanputra, Senior Director of Product Management, LinkedIn, reported August 2025 and July 2026. 2nd
- PR Daily, quoting Rishi Jobanputra, 25 August 2025. 2nd
LinkedIn publishes its ranker in academic papers.
No other platform in this document does this. The papers are LinkedIn-authored, dated, peer reviewed in two cases, and they name the production system.
-
01
First Pass Rankers
Multiple rankers each "create a preliminary ranking of their inventories based on predicted relevance" to the viewer. 7
-
02
Second Pass Ranker
"Combines the output of all FPRs and creates a single personalized ranked list." 7 Historically a logistic regression model outputting a probability of interaction. 8
-
03
Reranker
"Modifies the output of the SPR and creates the final feed that is sent to the appropriate front-end." 7
What is running now
In February 2026 LinkedIn published Feed SR, "a transformer-based sequential ranking model for LinkedIn Feed that replaces a DCNv2-based ranker." It reports it "has been serving the majority of LinkedIn's Feed traffic for over three months" with "+2.10% time spent" and "+3.52% like, comments, or reshares" against the previous production model. 9
The scale of the shortlist
The ranker "scores N candidates (typically N=512) for each member request." Candidates are retrieved from multiple sources first, then ranked. 9 Facebook narrows its inventory to roughly 500 posts before heavy ranking. These are the only two published shortlist sizes in this document and they are almost the same number, arrived at independently.
- LinkedIn Engineering, Feed , AI team page. No date on the page. Retrieved 8 August 2026. 1st
- LinkedIn Engineering blog, Making Your Feed More Relevant, Part I , November 2015. 1st
- Hertel et al. (LinkedIn), An Industrial-Scale Sequential Recommender for LinkedIn Feed Ranking , arXiv 2602.12354, February 2026. 1st
What the model is actually predicting.
LinkedIn names the outcomes its ranker tries to predict, and how they are combined. Dwell time is one of them, and has been since 2020.
The predicted actions, named
"Predicting multiple action probabilities including like, comment, share, vote, and long dwell and click for each member, candidate post pair. These predictions are linearly combined to generate the final post score."
By February 2026 LinkedIn had reduced the headline framing to two objectives. "Long Dwell: dwelling on a post longer than a specified threshold of time." And "Contribution: a like, comment, or share on a post." 11
Dwell time, and the skip penalty
LinkedIn published its dwell time work in May 2020. It measures two kinds: dwell "on the feed," which "starts measuring when at least half of a feed update is visible as a member scrolls," and dwell after a click. 12
"We incorporated a new 'P(skip)' objective into our feed value function that predicts the probability that the member will 'skip' an update. We incorporated the new model into our final ranking function by reducing the score of all updates by an amount proportional to the predicted P(skip) value."
A fast scroll past does not merely fail to help. It is a published, named term that subtracts from the score.
- Borisyuk et al. (LinkedIn), LiRank: Industrial Large Scale Ranking Models at LinkedIn , arXiv 2402.06859, KDD 2024. 1st
- Hertel et al. (LinkedIn), arXiv 2602.12354, February 2026. 1st
- LinkedIn Engineering, Understanding dwell time to improve LinkedIn feed ranking , May 2020. The page itself renders no date; May 2020 established from contemporaneous reporting. 1st
The link penalty.
The single most repeated claim in B2B marketing. It has no primary source, and the most widely circulated version of it publishes no sample, no method and no denominator.
What is claimed
A Forbes column published 30 July 2026 states: "Every time you put a link in a LinkedIn post, the algorithm cuts its reach by around 60%," and adds that "the old trick of dropping it in the first comment gets the same cut now." 13
What LinkedIn has actually said
A LinkedIn Senior Director of Product Management for Feed has stated publicly that external links are not penalised. 14 A LinkedIn Senior Director and Executive Editor put it as: "Our algorithm is really focusing on the value and content that's being shared in the post." 15
No LinkedIn source retrieved has ever stated that link posts are demoted. The claim exists only in marketing writing.
What is true, and much smaller
LinkedIn did reduce the size of the link preview image in organic posts during 2024, so a link post occupies less of the screen unless it is promoted. 16 That is a layout change, not a ranking one, and it is the only verifiable adverse change to links found.
- Forbes, The LinkedIn Link Penalty Cutting Your Reach By 60% , 30 July 2026. No sample, method or denominator published. 2nd
- Rishi Jobanputra, LinkedIn, video statement August 2025, reported 4 July 2026. 2nd
- Buffer, quoting Laura Lorenzetti, LinkedIn News, 19 December 2025. 2nd
- Social Media Today, Where Each of the Big Social Platforms Stands on External Links , 25 November 2024. 2nd
- Socialinsider, LinkedIn Organic Benchmarks 2026 . Engagement divided by impressions. 1.3m posts, 16,645 business Pages, January to December 2025. 2nd
What LinkedIn says it reduces, and what it has never said.
Three things are on record from named LinkedIn staff. Three more are attributed to LinkedIn constantly and appear in no LinkedIn source.
| Behaviour | Status | The evidence |
|---|---|---|
| Spam and bot content | On record | LinkedIn's Senior Director of Product Management for Feed: "We actively detect and limit the reach of spammy or low-quality content, including bot-generated posts and comments." August 2025. 18 |
| Automated comments | On record | LinkedIn Help states that "if we detect excessive comment creation or use of an automation tool, we may limit the visibility of those comments." 19 In February 2026 a LinkedIn VP of Product announced removal from Most Relevant, reach limited to the commenter's own network, and account restrictions for repeat offenders. 20 |
| AI slop | On record, July 2026 | LinkedIn's Chief Product Officer announced a "seems like AI slop" reporting control, new classifiers, and reduced visibility in out-of-network recommendations. He was explicit about the mechanism: "this lets us tune our models and make better feeds," and "we want members to get feedback from real humans on what sounds authentic, not just have an AI detector review it and get it wrong." 21 |
| Engagement bait | No LinkedIn source | No LinkedIn page or statement retrieved uses the term or describes a demotion for it. The published "engagement bait" demotion policy belongs to Meta, and appears to have been transferred across platforms by marketing blogs. The nearest LinkedIn statement is about withdrawing one gamed signal: "we just stopped using that as a signal." 22 |
| Engagement pods | No LinkedIn source | LinkedIn's enforcement statements are about automated comments and fake engagement. No retrievable LinkedIn source describes a penalty for manual reciprocal commenting groups. 18 , 19 , 20 |
| Reposts without comment | No LinkedIn source | No LinkedIn statement on the distribution of a plain repost was found. LinkedIn did change how it counts reposts in its transparency reporting from July 2025, but that is a measurement change in a safety report, not a ranking statement. 23 |
- PR Daily, quoting Rishi Jobanputra, LinkedIn, 25 August 2025. 2nd
- LinkedIn Help, Comments on posts , quoted by Social Media Today, 20 August 2025. Help page not directly retrievable. 2nd
- Social Media Today, quoting Gyanda Sachdeva, VP Product Management, LinkedIn, 16 February 2026. 2nd
- TechCrunch 30 July 2026 and Social Media Today 2 August 2026, quoting Hari Srinivasan, Chief Product Officer, LinkedIn. 2nd
- Social Media Today, quoting Tim Jurka, Senior Director of Engineering, LinkedIn, 27 February 2024. 2nd
- LinkedIn, Community Report , July to December 2025 reporting period. 1st
Knowledge, format and hashtags.
LinkedIn has been consistent on one of these for three years, silent on the second, and has withdrawn the third.
Knowledge and advice, since 2023
LinkedIn retuned the feed toward knowledge and advice in June 2023 and reported a roughly 40 percent rise in members seeing knowledge-grounded content from outside their network. 24
Its Editor in Chief on virality: "When things go viral on LinkedIn, usually that's a sign to us that we need to look into this." 24
Still the stated position in December 2025: "Think about what kind of knowledge you have to offer to help people." 25
Format: no stated preference
LinkedIn has never published a ranking preference for text, document, image, video, newsletter or poll. Its Feed product lead's position is the opposite: LinkedIn wants members to "play with all the different creation types available on LinkedIn." 26
Every format ranking table in circulation comes from third-party business Page datasets, not from LinkedIn. Those are worth reading, and they are on the next page, but they are not policy.
Hashtags: withdrawn
LinkedIn's Feed product lead, July 2025: "Since hashtags were introduced over a decade ago, our feed algorithm has really evolved." LinkedIn has been retiring hashtag feeds. 27
A LinkedIn editorial executive in December 2025 called them "a nice to have, not a need to have." 25
- Entrepreneur, quoting Alice Xiong, Director of Product Management, and Dan Roth, Editor in Chief, LinkedIn, 26 June 2023. 2nd
- Buffer, quoting Alice Xiong and Laura Lorenzetti, LinkedIn, 19 December 2025. 2nd
- PR Daily, quoting Rishi Jobanputra, LinkedIn, 25 August 2025. 2nd
- Social Media Today, quoting Rishi Jobanputra, LinkedIn, 20 July 2025. 2nd
The measured numbers, and the one that does not exist.
Two large 2026 panels. One publishes its denominator, one does not. Neither publishes the number most people are looking for.
Engagement rate by format, business Pages
| Format | Engagement rate, by impressions |
|---|---|
| Native documents | 7.00% |
| Multi-image | 6.45% |
| Video | 6.00% |
| Image | 5.30% |
| Text | 4.50% |
| Poll | 4.20% |
| Link | 3.25% |
Socialinsider. Engagement divided by impressions. 1.3 million posts from 16,645 business Pages, January to December 2025. 28 Platform average 5.20 percent. Business Pages only, so it says nothing about personal profile posts, which is where most owner-led content sits.
The second panel, and its gap
Metricool measured 673,658 posts from 63,108 accounts, comparing January and February 2025 against the same months in 2026. Overall engagement up nearly 14 percent, while likes fell 13 percent, comments 17 percent and shares 10 percent. Personal profiles out-engaged company Pages by 63 percent. Nearly 40 percent of all interactions land on the first day. 29
- Socialinsider, LinkedIn Organic Benchmarks 2026 , published 16 March 2026, updated 15 July 2026. 2nd
- Metricool, 2026 LinkedIn Study , 14 April 2026. 673,658 posts, 63,108 accounts. Denominator not published. 2nd
Six claims checked against the source.
Two are true and specific. Four are not sourced to LinkedIn at all, and one of those four was published in a national business title in the last fortnight.
| The claim | Verdict | What the source actually says |
|---|---|---|
| Links cut your reach by 60 percent | No source | LinkedIn has never said links are penalised, and a named LinkedIn product director has said publicly that they are not. The 60 percent figure traces to a vendor blog with no sample, no method and no denominator. 30 , 31 |
| Dwell time is a ranking signal | True, and published | LinkedIn published the mechanism in May 2020, including a named P(skip) objective that reduces a post's score in proportion to predicted skip probability. Long Dwell remains one of two named headline objectives in its February 2026 paper. 32 , 33 |
| LinkedIn's feed now runs on an LLM | Contradicted | LinkedIn tested an LLM ranker for Feed and rejected it. Its own paper: encoding a member's history "took tens of thousands of tokens," and "the LLM-Ranker never achieved superior online performance." The shipped system is a transformer-based sequential recommender, which is not the same claim. 33 |
| LinkedIn demotes engagement bait | Wrong platform | No LinkedIn source uses the term or describes the demotion. The published engagement bait policy belongs to Meta. It appears to have been carried across platforms by marketing writing. 34 |
| Documents and carousels outrank other formats | Measured, not policy | LinkedIn has stated no format preference and tells members to use all creation types. Native documents do lead one 2026 business Pages panel at 7.00 percent engagement by impressions against a 5.20 percent average, which is a measurement of business Page behaviour and not a ranking rule. 35 , 36 |
| LinkedIn has 1.3 billion members | Pick a number | LinkedIn's newsroom says 1.3 billion. Its own About page says more than 1 billion. Its February 2026 engineering paper says 1.2 billion. The regional figures on the newsroom page sum to roughly 1.356 billion. None of these pages carries a date. 37 , 33 |
- Forbes, 30 July 2026. 2nd
- Rishi Jobanputra, LinkedIn, August 2025. 2nd
- LinkedIn Engineering, dwell time, May 2020. 1st
- LinkedIn, arXiv 2602.12354, February 2026. 1st
- Meta, Content Distribution Guidelines , September 2021. Absence in LinkedIn sources verified 8 August 2026. 1st
- PR Daily, quoting Rishi Jobanputra, 25 August 2025. 2nd
- Socialinsider, LinkedIn Organic Benchmarks 2026 . 2nd
- LinkedIn, About Us and About page. No dates. 1st