The majority of LinkedIn feed traffic has been ranked by language models and transformer models since 2025. Long dwell time is a training signal in its own right, sitting alongside likes, comments and shares. How heavily LinkedIn weights each signal is not public. For B2B reach, that makes reading attention the lever, and interchangeable content gets throttled.
This article separates three levels: what LinkedIn has published itself, what large third-party datasets observe, and what circulates without a primary source. Every figure carries its source. Where the collection period and sample size have been published, they appear too, and where they are missing, that is stated. This matters because most figures circulating about the LinkedIn algorithm are copied from one another unchecked.
What LinkedIn has published, and when
On 12 March 2026, LinkedIn published two things in parallel. In his feed update , Tim Jurka describes the move to "generative recommenders" and large language models. According to that announcement, the platform reduces repetitive low-substance posts, engagement bait along the lines of "Comment 'Yes' if you agree", videos that do not match the text, and recycled thought leadership.
One point matters for dating all of this: that post summarises changes over the past year and describes a rollout in progress. It marks no cut-off date. Feed SR, the ranking model described in arXiv 2602.12354 , had already been serving the majority of feed traffic for more than three months before the paper appeared in February 2026. Anyone naming a switchover date is compressing a process that spans several quarters.
The engineering post by Hristo Danchev supplies the technical reasoning the same day. A unified LLM retrieval layer replaces several separate systems there, among them trending posts, collaborative filtering and industry-specific trending sources. A day later came the package against inauthentic activity : engagement pods are removed, automated comments are hidden from "Most Relevant", and more than 100 million members now carry a verification.
What these sources do not contain is any weighting of the signals. LinkedIn describes architecture and training objectives while publishing nothing about how strongly an individual signal acts. Every percentage figure about the effect of a ranking factor therefore comes from third-party measurement.
- Oct 2024 Dwell time as an adjustable ranking weight Engineering post describes dwell time as passive consumption with an adjustable weight, explicitly without a universal threshold in seconds.
- Oct 2025 Retrieval via causal language models arXiv 2510.14223, for suggested content from outside a member's network: around 2,000 candidates, latency budget of a few milliseconds.
- from ~Nov 2025 Feed SR serves the majority of traffic According to the February 2026 paper, the model had by then been in production for more than three months.
- Feb 2026 Feed SR replaces the DCNv2 ranker arXiv 2602.12354. Online test: plus 2.10 percent time spent, plus 3.52 percent likes, comments and shares.
- 12 Mar 2026 Generative recommenders in the feed explained Product announcement and engineering post summarise the rollout in progress.
- 13 Mar 2026 Package against inauthentic activity Engagement pods removed, automated comments hidden, more than 100 million verifications.
- May 2026 Reach limits for generic content Laura Lorenzetti confirms the throttling. LinkedIn cites 94 percent accuracy in early tests, with no figure for the error rate.
- 2 Aug 2026 EU AI Act, Article 50 Transparency obligations apply from this date, with different duties for providers and deployers. Marking duty for pre-existing systems only from 2 December 2026.
Source: LinkedIn News, LinkedIn Engineering Blog, arXiv 2510.14223 and 2602.12354, European Commission
Why dwell time matters as a signal
The most solid sentence on the subject sits in LinkedIn's engineering blog. The ranking model trains on "passive tasks (click, skip, long-dwell) and active tasks (like, comment, share)". Long dwell time is therefore a training signal in its own right, standing alongside the active signals.
What does not follow from this is any ranking order. LinkedIn names both groups while publishing no weighting. Likes, comments, shares and popularity scores remain explicit model features. Claiming that dwell time has displaced comments as the most important signal goes beyond the source.
Passive signals
- Click, opening a post
- Skip, scrolling past
- Long dwell, extended reading time
Active signals
- Like, the reaction
- Comment, the reply
- Share, passing it on
passive tasks (click, skip, long-dwell) and active tasks (like, comment, share)
LinkedIn Engineering, Engineering the next generation of LinkedIn's Feed, 12 March 2026
Source: LinkedIn Engineering Blog, 12 March 2026
The Feed SR ranking model delivers 2.10 percent more time spent and 3.52 percent more likes, comments and shares in its online test. That time figure refers to total time members spend on the platform. It says nothing about dwell time on an individual post.
Two details of the measurement method carry practical weight. First, LinkedIn defined feed dwell time in 2020 such that measurement begins once at least half of a post is in view. Whether that still holds for the current model is not published. Second, no universal target in seconds has been published. LinkedIn writes explicitly that 30 seconds may count as long for an image but not for a video. The threshold is relative to format and kept internal.
In practice: a post needs reading attention before it can earn interaction. Someone scrolling straight past produces a skip, and skip is one of the passive tasks the model trains on. Whether and how strongly that counts against a post is not published.
What the format data shows, and what it does not
A note on how much weight these figures can carry, applying to all of them. They are observational data from uncontrolled averages. The collection periods fall mostly before March 2026, and Socialinsider itself notes that the annual values presented as 2026 come from 2025. This data cannot evidence the rollout. It is compatible with the reading that formats generating reading attention perform better, and it carries no claim about causes.
The methodologically cleanest dataset comes from AuthoredUp : more than three million posts from personal profiles between March 2025 and February 2026, reported as a factor against each profile's own median. That normalisation compares every profile with itself, which removes size differences between accounts.
| Category | Reach (factor) | Interaction (factor) |
|---|---|---|
| Document | 1.39 | 1.30 |
| Image | 1.20 | 1.33 |
| Poll | 1.78 | 0.37 |
| Text | 1.07 | 0.78 |
| Video | 0.86 | 0.93 |
| Article | 0.69 | 0.44 |
| Repost | 0.29 | 0.22 |
Source: AuthoredUp 2026, more than 3m posts from personal profiles, March 2025 to February 2026 Values as a factor against each profile's median. 1.00 equals that profile's own median.
The poll deserves particular attention. At a factor of 1.78 it reaches further than any other format, and at 0.37 it produces the weakest interaction. Judged on impressions, that looks like success. Worth noting: this interaction figure combines reactions, comments and shares, and excludes the vote itself. About the conversational value of a poll it therefore says only that polls average fewer interactions beyond voting.
Company pages show a similar pattern from another direction. Socialinsider analysed 1.3 million posts from 16,645 company pages between January 2024 and December 2025. One definition matters for comparing this against other studies: Socialinsider defines engagement rate as interactions divided by impressions. Figures from studies that normalise on followers or on posts are not comparable.
| Category | Value |
|---|---|
| Native documents | 7.00 % |
| Multi-image | 6.45 % |
| Video | 6.00 % |
| Image | 5.30 % |
| Text | 4.50 % |
| Poll | 4.20 % |
| Link post | 3.25 % |
Source: Socialinsider 2026, 1.3m posts from 16,645 company pages, January 2024 to December 2025
Video shows a countervailing movement in the same dataset that needs keeping apart. Average video views fell across every page size, from 190 to 155 for pages with 1,000 to 5,000 followers and from 2,430 to 1,380 for pages with 100,000 to one million followers. Video engagement rate rose 7 percent to 6.00 percent over the same period. Fewer people watch a video, and those who do react more often.
Source: Socialinsider 2026, 1.3m posts from 16,645 company pages, January 2024 to December 2025. Values presented as 2026 come from 2025 per the provider.
No format penalty follows from this, and neither does dropping video. One possible explanation is saturation: according to Metricool , video is now the most-used format on personal profiles. If more video is published while attention stays flat, views per post fall. That explanation is untested, and format shares come out differently in the AuthoredUp dataset, which shows how strongly such claims depend on the sample. The practical conclusion concerns what to expect from view counts, less so the choice of format.
How LinkedIn throttles generic content
Two things get mixed together here. No blanket penalty for AI assistance is documented. What LinkedIn names is the throttling of interchangeable content, and separately from that, the sanctioning of automation.
Laura Lorenzetti, VP and Executive Editor at LinkedIn, explained the measure to Entrepreneur in May 2026. Content creation was up 14 percent year over year: "AI can really help people unlock content creation. But it also means that a lot of people can produce a lot of very low-quality content." According to The Next Web , the detection system correctly flagged generic content in 94 percent of cases in early tests. LinkedIn has published neither a test sample nor the share of wrongly flagged posts, and without that error rate an accuracy figure alone says little.
The consequence described is reduced reach. According to the interview, visibility of flagged posts often stays limited to first-degree connections. The source does not describe complete removal from all recommendations, and the rollout was still unfinished in May 2026. Anyone with few followers of their own loses proportionally more distribution under such a limit.
Gets throttled
- Repetitive posts with little substance
- Engagement bait such as "Comment 'Yes' if you agree"
- Videos that do not match the post text
- Recycled thought leadership
- Template phrasing, with "it's not X, it's Y" named specifically
- Automated comments from tools and extensions
- Coordinated engagement pods
Stays permitted
- AI as support for research and structure
- AI for editing your own text
- Your own observations and your own numbers
- Technical terms and evidenced claims
- Recurring topics approached from a new angle
Source: LinkedIn News 12 and 13 March 2026, Entrepreneur and The Next Web, May 2026
What Article 50 of the EU AI Act means for your posts
The transparency obligations in Article 50 of the EU AI Act apply from 2 August 2026. Marketing commentary often turns this into a labelling duty for every AI-assisted post. The European Commission's FAQ does not support that reading.
The regulation separates two roles. Providers of generative systems must mark their outputs in a machine-readable format so they are detectable as artificially generated. That duty falls on the model providers, not on you as the author of a post. For systems already on the market before 2 August 2026, the Commission says this marking duty applies only from 2 December 2026.
Deployers face several separate disclosure duties, drawn from different subsections with criteria of their own. Under paragraph 4, deepfakes must be clearly labelled, as must AI-generated or manipulated text published to inform the public on matters of public interest. Under paragraph 3, people must be informed when emotion recognition or biometric categorisation is in use. These are distinct obligations rather than three equivalent cases.
The exemption for editorial control is what matters most in practice. The Commission puts it plainly: published text that has undergone human review or editorial control does not need to be labelled. Purely formal or procedural checks such as spell-checking or grammatical correction expressly do not count as editorial control. A professionally reviewed B2B post published under someone's responsibility therefore falls outside the deployer disclosure duty in the normal case.
The penalty ceiling of up to 15 million euros or 3 percent of worldwide annual turnover applies to infringements of the regulation and depends on role and breach, with explicit regard for the size of small and medium-sized enterprises. It is not a price tag on an unlabelled LinkedIn post. Content generated before 2 August 2026 does not require retroactive labelling, according to the Commission.
None of this is legal advice. If you deploy deepfakes or publish on matters of public interest, have the specific case reviewed.
What external links actually cost
No topic produces more contradictory analyses. The resolution lies in separating two questions that usually get treated as one.
The first concerns account type. Metricool analysed 673,658 posts from 63,108 accounts across two windows from January to February in 2025 and 2026, and finds an opposing relationship for links depending on the sender: company pages sit 51 percent above on impressions and 41 percent above on interactions with a link, while personal profiles sit 27 and 20 percent below.
This comparison carries a limitation that is almost always missing from the way it gets repeated. For personal profiles, Metricool captures by its own account only posts published through Metricool, because the LinkedIn API allows nothing else. No such restriction applies to company pages. The two groups are therefore not collected on equal terms, and part of the difference may come from the sample. As a correlation within this dataset the finding is usable. As a rule about account types it does not hold.
The second question concerns the number of links. The widespread observation that reach recovers from three links onward appears in two reports: 140 percent across 223,996 posts at Saywhat, and 236 percent across 318,842 posts in the third quarter of 2025 from Will McTighe and Chris Donnelly. These two figures are not independent confirmation. Will McTighe is Saywhat's chief executive, so both reports come from the same source with different collections.
| Category | Value |
|---|---|
| Three or more links | +236 % |
| Three or more links | +140 % |
| Company page, impressions | +51 % |
| Company page, interactions | +41 % |
| One link in the text, median reach | −18.8 % |
| Personal profile, interactions | −20 % |
| Personal profile, impressions | −27 % |
Source: Metricool 2026, Saywhat and McTighe / Donnelly, Richard van der Blom Algorithm Insights 2026 Separate studies with different samples and metric definitions. These values cannot be netted against one another.
As an instruction this is useless. Posts with three or more links are usually resource lists, a different kind of content from the post carrying one link to a landing page. The most plausible reading is a selection effect: what gets measured is the difference between content types. That remains a hypothesis consistent with the data.
For the widespread claim that a link in the first comment costs up to 80 percent of visibility, no verifiable primary source exists. More notable by comparison is what the LinkPost analysis of 438,413 posts writes about external links: "not measured here". That kind of candour is rare in the advice literature.
The screenshot trick and where it ends
One tactic spread visibly in 2026: screenshots of text messages, direct messages or chat threads posted as an image. There is neither a platform statement nor a solid study on its effect. What follows is a reasoned hypothesis and a practical warning.
The hypothesis: a screenshot has to be read, and reading produces dwell time. Since long dwell time is a training signal in its own right, the format could benefit. The AuthoredUp finding fits, with image posts reaching the highest interaction multiplier of any format at 1.33. That connection is untested, and image posts are a far broader category than screenshots.
The widespread claim that screenshot posts outperform video, text and carousels by 85 percent comes from a Reddit thread. It has no methodological basis.
Two limits belong with this. First, LinkedIn has explicitly throttled repetitive low-substance posts since March 2026. A tactic that hardens into a formula works against that criterion. Second, publishing private messages engages the GDPR. A screenshot containing identifiable details processes personal data, and blacking out names and profile pictures does not produce reliable anonymisation, because context and phrasing can keep a person identifiable. The dependable routes are consent from the person concerned or a complete rewrite without personal reference. LinkedIn notifies nobody about a screenshot, which says nothing about whether publishing it is lawful.
Dos and don'ts for practice
Every line here hangs on evidence from the sections above.
Do | Why, with evidence |
|---|---|
Plan for reading attention before interaction | Long dwell is a training signal of the ranking model in its own right (LinkedIn Engineering, 12 Mar 2026) |
Use documents and images for technical content | Document 1.39x reach, image 1.33x interaction against a profile's own median (AuthoredUp, 3m posts, Mar 2025 to Feb 2026) |
Measure reach per account type yourself rather than taking it from studies | Metricool measures 2.60 against 1.60 percent engagement rate, but collects personal profiles selectively (see the section on links) |
Measure your own link effect per account type | The study picture is contradictory and personal-profile data is selectively collected (Metricool, see the section on links) |
Stay with a post for the first 48 hours | 50 percent of all impressions fall in that window (Metricool Study 2026, 673,658 posts) |
Build in your own observations and your own numbers | What gets throttled is interchangeable content (LinkedIn, May 2026) |
Document professional review before publishing | Editorial control is an exemption from the disclosure duty (Article 50 EU AI Act, Commission FAQ) |
And the other side, likewise evidenced rather than assumed:
Avoid | Why, with evidence |
|---|---|
Engagement bait and comment prompts as a reach trick | Named explicitly as grounds for reduction (LinkedIn, 12 Mar 2026) |
Engagement pods and comment automation | Being removed, accounts can be restricted (LinkedIn, 13 Mar 2026) |
Judging polls on impressions | 1.78x reach at 0.37x interaction beyond the vote itself (AuthoredUp, 3m posts) |
Expecting high view counts from video | Average views down 36 percent year on year, while engagement rate rose (Socialinsider, 1.3m posts) |
Templates such as "it's not X, it's Y" | Named by LinkedIn as a target pattern (The Next Web, May 2026) |
Videos that do not match the text | Named explicitly as grounds for reduction (LinkedIn, 12 Mar 2026) |
Chasing a target dwell time in seconds | The threshold is format-relative and unpublished (LinkedIn Engineering, Oct 2024) |
How to build a dependable rhythm out of this rather than a collection of single tactics is what we describe in the 4R system . For building topical authority over longer periods, our guide to thought leadership and category design is the better entry point, and for the individual post in a sales context, the walkthrough of the social selling post .
Eight figures that are unevidenced or not comparable
These claims circulate widely. None of them is disproven; all of them lack a verifiable basis or the comparability they are given. We name where they appear, because a claim without an address cannot be checked.
"360Brew is the 2026 algorithm." The corresponding arXiv paper 2501.16450 describes the model itself as a "research pre-production model". It therefore describes no production state. The paper was also withdrawn by the arXiv administrators, because the submitter did not have the right to agree to the licence at the time of submission. That is a licensing matter and not a retraction on substance. The systems evidenced in production are Feed SR and causal-LM retrieval. Circulates at dataslayer.ai, linkboost.co, expandi.io and falia.co.
Fixed dwell-time thresholds in seconds. In circulation: "30 seconds equals strong interest", "61+ seconds gives a 15.6 percent engagement rate", "15+ seconds gives 3.2x the reach" and "carousel equals 42 seconds". No primary source exists for any of these, and LinkedIn describes the threshold as format-relative. Circulates at viralbrain.ai, socialpilot.co, growwithghost.io, contentdrips.com, meet-lea.com and prospectory.ai.
"Comments count 15 times as much as a like." There is no source for this factor. LinkedIn publishes no signal weights, which makes every such number an estimate. Competing estimates range from roughly a factor of 2 to a factor of 15. None of them cites a traceable measurement. Circulates at meet-lea.com and viralbrain.ai.
The "Depth Score". This metric appears in no LinkedIn publication. Circulates at digitalapplied.com and socialboostdigital.com.
"The authenticity update cuts AI content reach by up to 47 percent." LinkedIn has named no figure for reach effects. The claim has no evidenced basis. Circulates at futurefactors.ai.
"Poll engagement rate fell to 0.07 percent." No primary source, and Socialinsider measures 4.20 percent over the overlapping period. Circulates at dataslayer.ai.
Carousel engagement rates of 21.77, 40.5 or 49.52 percent. These values rest on different metric definitions and reference bases. Socialinsider normalises on impressions, other analyses on followers or posts. Set against each other they produce a contradiction that comes purely from the arithmetic. Found at Buffer, influencia.net, dataslayer.ai.
"A link in the first comment costs up to 80 percent of visibility." No verifiable primary source. Circulates at contentdrips.com and at Melanie Goodman on Substack.
A note about our own work belongs here. An earlier version of this article circulated unevidenced claims of its own, among them a quotation attributed to then LinkedIn CEO Ryan Roslansky, along with terms such as "Sustained Relevance Scoring" and "Topic Authority Score" that appear in no LinkedIn source. Those passages are gone. We document that because an article about the throttling of interchangeable content would otherwise be poorly placed to make the argument.
Frequently asked questions about the LinkedIn algorithm
How does the LinkedIn algorithm work in 2026?
LinkedIn ranks the feed with language models and transformer models. One paper on retrieval describes around 2,000 candidates for suggestions from outside a member's network. A second describes the ranker Feed SR, which serves the majority of feed traffic and trains on passive signals such as dwell time alongside active ones such as likes and comments. How the two stages fit together is not published.
How important is dwell time for reach?
Long dwell time is a training signal of the ranking model in its own right and sits alongside likes, comments and shares. That appears in LinkedIn's engineering blog of 12 March 2026. LinkedIn has published no weighting between the signal groups.
Is there a number of seconds at which dwell time starts to count?
No universal target has been published. LinkedIn describes the threshold as relative to format and writes explicitly that 30 seconds may count as long for an image but not for a video. The absolute values in circulation have no primary source.
Does LinkedIn penalise content created with AI?
No blanket penalty for AI assistance is documented. What gets throttled is generic content, and automation is sanctioned separately from that. LinkedIn cites 94 percent detection accuracy in early tests, with no figure for the error rate. Visibility of flagged posts often stays limited to first-degree connections, per the interview. AI for research, structuring and editing remains permitted.
Do I have to label AI content on LinkedIn?
Under Article 50 of the EU AI Act from 2 August 2026, providers of generative systems must mark their outputs in machine-readable form. Deployers must disclose for deepfakes and for text on matters of public interest. Text under editorial control is exempt. For a professionally reviewed B2B post there is normally no disclosure duty. This is not legal advice.
Do external links in posts cost reach?
The data points both ways. Metricool observes 51 percent more impressions with a link for company pages and 27 percent fewer for personal profiles, but captures only posts published through Metricool on personal profiles. A single link in the text goes together with 18.8 percent lower median reach at van der Blom. All observational data without causal control, which is why your own measurement takes precedence.
Why does reach recover with three links?
Two reports observe the effect, at 140 and 236 percent. Both come from the same source, since Will McTighe is Saywhat's chief executive. The most plausible explanation lies in the type of content: posts with several links are usually curated resource lists with value of their own. That explanation is unverified.
Which format has the best reach on LinkedIn in 2026?
In the available datasets, documents deliver the best combination of reach and interaction, at a factor of 1.39 and 1.30 against a profile's own median and at 7.00 percent engagement per impression on company pages. The collections run to February 2026, and to December 2025 for Socialinsider. Treat them as a starting point for your own tests rather than a ranking for 2026.
If you want to work out which of these levers hold for your company, we will look through your recent posts together and put the numbers in context. A strategy conversation takes 30 minutes and ends with an assessment.