LinkedIn is no longer a virality machine, and treating it as one is what costs most posters their reach. The ranking system reads what a post means against your profile and the reader's history, and it scores quiet signals like time spent above loud ones like reaction counts. Specific, consistent posting inside one recognisable subject beats every format trick. The surface with the best odds of being seen is not the feed at all.
LinkedIn documents its feed thoroughly, but it documents it for machine learning engineers. The two best sources on how ranking works in 2026 are an arXiv paper about transformer architecture and an engineering blog post about GPU inference, and neither says a word about what a poster should write. Meanwhile every source that does address what to write measures reach using data from a third-party scheduling or analytics tool, which means the sample is people who pay for social media software. Mechanism is well documented and comes straight from LinkedIn. Outcomes are measured only by commercially interested outsiders with self-selected samples. Nobody has audited the gap, and LinkedIn's own analytics give each member their own numbers with no comparison set, so a diligent poster cannot check most of the circulating claims against anything but themselves.
There is one number in the public record that calibrates the whole category. When LinkedIn replaced its entire feed ranking model in February 2026, with full access to the training data, the labels, and the serving infrastructure, it bought a 2.10 percent move in time spent. That is the ceiling on what the best-resourced team in the world extracted from rebuilding the ranker. Any advice claiming a formatting change will multiply your reach is describing an effect larger than that, and it never carries evidence proportional to its size. This report separates what LinkedIn published from what the advice market asserts, prices the difference, and names the tactics that have been dead since 2024.
Nine findings on how LinkedIn ranks what you publish for free, drawn from LinkedIn's February 2026 Feed SR paper, its March 2026 feed engineering post, its 2020 and 2024 dwell-time work, and the three third-party samples that measure outcomes. Each is labeled by evidence type. Vendor primary sources are graded separately from practitioner aggregates, and where the honest answer is that nobody has measured it, that is the finding.
LinkedIn's February 2026 paper describes Feed SR, a decoder-only transformer with causal attention that reads up to 1,000 of a member's past feed impressions as an ordered sequence. The same paper states that an LLM-ranker the team built never achieved superior online performance over the existing production model. Advice built on the premise that a generative model now reads and judges your prose is describing retrieval, not ranking.
That is the headline online result for Feed SR against the DCNv2 model it replaced: plus 2.10 percent time spent overall, plus 2.38 percent among daily actives, and no significant effect on new members. It is the best calibration device available for the whole category, and a useful ceiling to keep in mind when a post promises to triple your reach with a formatting change.
Feed SR trains on two groups of binary outcomes. The active group is like, comment, and share. The passive group is click, skip, and long-dwell. Half the objective function is made of things a reader does silently, and one of them, skip, is a penalty you incur simply by appearing in a feed and being scrolled past.
In October 2024 LinkedIn replaced its fixed skip threshold with an auto-normalised long-dwell model that scores a post against a percentile of its own counterparts, recalculated daily, controlling for content type, creator type, and distribution path. A long video no longer beats a short text post simply by consuming more seconds. You compete against posts like yours, not against the feed at large.
Socialinsider's benchmark of 1.3 million posts across 16,645 business pages puts native documents at a 7.00 percent engagement rate for 2025, with text at 4.50 percent. Buffer's cross-platform study puts LinkedIn carousels at a 21.77 percent median engagement rate against 3.18 percent for text. Same direction, wildly different magnitude, because the denominators and the samples differ. Neither number should be quoted without the other.
Microsoft reported video watch time up 36 percent year over year and video posts shared twenty times more than any other content type. Socialinsider measured video views on business pages down 36 percent year over year across every follower band, with the 50,000 to 100,000 band down 53 percent. Both can be true if video supply grew faster than video demand. Aggregate watch time is a platform statistic; views per post is your statistic.
LinkedIn's Senior Director of Product Management stated publicly in September 2025 that posts are not intentionally demoted for containing a link. A 1.3 million post practitioner sample reports one external link in the body cutting median reach by 18.8 percent. A July 2026 Forbes column put the figure at 60 percent; following the citation leads to a marketing blog with no sample size, no date, and no methodology. Treat 18.8 percent as the defensible working estimate and the 60 percent figure as unusable.
Originality.ai classified 81.2 percent of 5,000 sampled long-form July 2026 posts as likely AI. LinkedIn's own classifiers, per its chief product officer, put roughly 40 percent of long-form and 30 percent of short-form posts in the fully-AI bucket. On 30 July 2026 LinkedIn shipped a "Seems like AI slop" report button, retired its "enhance your post" writing tool in favour of a proofreader, and said it blocks hundreds of thousands of automated comment attempts a day.
A newsletter edition triggers a notification and an email to every subscriber, which is delivery rather than ranking. A personal post is ranked but starts from the strongest position in the feed. Native video is ranked and currently oversupplied. A company page is ranked from the weakest position on the platform. The combination that compounds is post to build subscribers, newsletter to keep them.
Not ranked by engagement rate, and not by how hard LinkedIn is promoting the surface. Ranked by how much of the outcome is decided by a ranking model you do not control. Each entry carries what it takes and where it fails, because a surface you cannot sustain is not an advantage.
Publishing an edition notifies every subscriber and emails them, which sidesteps retrieval, the sequence model, the six predicted actions, and normalised dwell entirely. Editions also get static URLs that search engines index, so they accumulate readers on a timescale the feed does not operate on. Of the four surfaces it is the only one that behaves like an owned asset rather than a rented one.
Fails on cadence. A newsletter that misses its interval trains subscribers to ignore the notification, and there is no algorithm to blame.
Semantic retrieval matches your post embedding against your profile, then the sequence model ranks it against reader history. Documents or text, three to five times a week, is where the large samples converge. This is also the surface that builds the subscriber base the newsletter needs, so a poster with no list should treat it as item one.
Fails on subject drift. Retrieval matches your post against your profile, so posting outside your stated expertise weakens the match before ranking begins.
Dwell is normalised within content type, so you compete against other video, and video upload volume grew for three straight quarters. Production effort sits well above a text post for a per-post view count that fell in every follower band in the most recent benchmark. If some arguments only work on camera, make it anyway and judge it on completion rather than views.
Fails on economics. It is the only surface where the platform's promotion and the measured per-post outcome point in opposite directions.
Practitioner measurement consistently puts page organic reach far below personal profiles for equivalent content, and the gap widened through 2024 and 2025. The same effort as a personal post for a fraction of the organic result. Employee amplification is the one thing that reliably helps: ten personal profiles enter retrieval where one page did.
Fails when treated as a reach channel at all. It is an advertising, hiring, and credibility asset, and it is good at those.
The advice market for LinkedIn is larger and older than the evidence base, and a poster following 2021 guidance in 2026 is actively harming their own distribution. Each ruling names what it rests on. Where the answer is that nobody knows, that is the ruling.
The most confidently asserted claim in this category with the least behind it. No LinkedIn engineering document describes a fixed early evaluation window. One practitioner analyst working from a 300,000 post sample states flatly that there has never been one. What the published architecture supports is weaker and different: skip is a trained negative, so impressions that go nowhere cost you regardless of when they happen, and no documented clock is attached to that.
Rests on: the Feed SR objective function, and a 300,000 post practitioner sample
LinkedIn confirmed in February 2024 that it stopped using "read more" clicks as a value signal once it understood the behaviour. Posters had learned that one-line paragraphs forced the truncation, that truncation forced a click, and that the click read as value. LinkedIn watched, called it an artefact, and deleted the signal. The formatting persists as a convention; the mechanical advantage that created it is gone.
Rests on: LinkedIn statements via Social Media Today, February 2024
The usual argument is that you will get caught, and the 2026 enforcement record supports it. The stronger argument needs none of it. Pods manufacture likes and comments, three of six trained objectives, from people who did not read the post, which damages the other three. A pod is a machine for producing high reaction counts alongside short dwell, and short dwell is the thing the ranker has been getting better at pricing since 2020.
Rests on: the published objective function, plus second-hand enforcement reporting
Never established as effective by any controlled measurement, and widely reported as patched. LinkedIn's stated position is that there is no intentional link penalty in the first place. If you are moving the link to avoid a penalty the platform denies exists, using a workaround nobody has measured, you are two unverified steps from the evidence. Include the link when the click is the objective, accept a cost in the mid-to-high teens, and skip the rituals.
Rests on: a first-party denial of intent, and one opaque outside estimate near 19 percent
LinkedIn's own editorial staff describe hashtags as a nice to have, not a need to have. Practitioner samples find no reach benefit. Bait tags such as #follow and #like are associated with low-quality classification. The rational move is zero to three, chosen for topical accuracy rather than reach.
Rests on: LinkedIn editorial statements, and practitioner samples finding no effect
Buffer's 4.8 million post analysis finds midweek afternoons ahead of other slots, and its authors attach the caveat themselves: posting at four on a Wednesday will not rescue a post that is not relevant. LinkedIn's editor in chief has told members not to chase timing at all. Treat it as a tiebreaker between two equally ready posts, never as a reason to delay a good one.
Rests on: a 4.8 million post scheduling-tool sample, with its own stated caveat
Polls score well on impressions in the Socialinsider sample, particularly at large follower counts, while sitting at the bottom on engagement rate growth. Practitioner reporting describes them as strong for reach and weak for follower growth. A poll buys impressions from people who will not remember you.
Rests on: the 1.3 million post business-page benchmark
Independent detection puts likely-AI at 81.2 percent of long-form posts; LinkedIn's own classifiers put fully-AI at roughly 40 percent of long-form. LinkedIn shipped a member-facing report button in July 2026 and says the reports tune its models, which makes member flags training data rather than only complaints. Whatever the distribution effect is today, the platform has publicly committed to making it negative.
Rests on: two undocumented classifiers that disagree by a factor of two, and a stated platform intent
The full report walks the feed end to end in four steps, from LLM dual-encoder retrieval at sub-50 millisecond latency through the mixture-of-experts prediction head; carries a dated timeline of every change since May 2020 with a locatable source attached to each row; presents three independent format samples side by side with their denominators exposed rather than picking the flattering one; ranks seven failure modes by how much reach each destroys, with the fix for each; runs four scenarios to mid-2028 with subjective probabilities and earliest visible signs; lists five leading indicators any reader can watch without private information; states what the document cannot answer, including whether format effects are causal at all; and prints a half-life on every section, from six months on the format numbers to architectural claims unlikely to decay. The call is falsifiable and the condition is printed: I would move off this position if LinkedIn published per-member reach data that contradicted the third-party samples everything here rests on.
This isn't a vendor summary. Every sentence is labeled by what stands behind it: verified fact, vendor claim, third-party estimate, my assessment, hypothesis, or scenario. Sources are numbered and clickable. Forward-looking sections use scenarios with observable tripwires, not forecasts. It's the same method behind every market assessment I write.
Twenty-eight pages, built from public sources with no client brief and no interviews. Read it in the browser or take the PDF.
Each report here answers a real question, directed and researched against public sources and evaluated against a stated assumption, then delivered as Word and PDF. If you're weighing a platform, sizing a category, or defending a number to a board, tell me the decision behind it and I'll tell you honestly whether a report is the right tool.
Commission an assessment