How the LinkedIn algorithm works

LinkedIn's engineering team published its new feed architecture in March 2026. Here is what it documents, what independent data shows, and what is still inference.

8 minute read

LinkedIn rebuilt the feed, and wrote it up

On 12 March 2026, LinkedIn's engineering blog published a post by Hristo Danchev describing a full rebuild of the feed recommendation system. The old architecture stitched together several separate retrieval sources, including trending content, collaborative filtering and embedding-based systems, each with its own infrastructure, and scored every candidate post independently. The replacement has two parts: a single retrieval model built on a fine-tuned large language model, which represents member profiles and posts as vectors in a shared space, and a transformer-based ranker LinkedIn calls a Generative Recommender, which reads more than a thousand of your past interactions as a sequence rather than as a pile of counts. Retrieval runs in under 50 milliseconds across millions of posts. That is unusually specific documentation for a feed, and it is worth reading before trusting anyone's tips.

The signals it actually consumes

In that same post, LinkedIn names the engagement signals the ranking model is trained on: long dwells, likes, comments and shares, alongside information about the post and about the viewer. LinkedIn's own help page on relevance groups its inputs into three families. Identity covers your profile, employer, skills and location. Content covers how much a post was viewed and reacted to, what it is about, how recent it is, whether it came from a connection, and the quality of the conversation under it. Activity covers what you react to, how often you engage, and how long you spend viewing. Note the order in that first list. Dwell is named ahead of likes, and LinkedIn has been consistent about that for years.

Dwell time, and why LinkedIn trusts it more than likes

LinkedIn's engineers explained the reasoning back in 2020, in a post by Siddharth Dangi on using dwell time in feed ranking. Their case was about signal quality rather than philosophy. Clicks and viral actions are binary, sparse, and especially rare among people who read without engaging, while dwell is always measurable, continuous, and abundant. They defined a threshold below which a post counts as skipped, and reported that the dwell-aware model produced a large drop in skipped updates plus more clicks, more reactions and more time spent on the feed, which is why it shipped to everyone. The practical translation is simple: a post that holds someone reading for half a minute is a stronger signal than a like collected in half a second.

Native video: the two numbers that disagree

Here the published data pulls in opposite directions, and the contrast is the useful part. In February 2025, LinkedIn's director of product Lakshman Somasundaram said video watch time on the platform was up 36% year on year, and that video creation was growing twice as fast as other post formats. Socialinsider's 2026 benchmarks, built from 1.3 million posts across 16,645 business pages between January 2024 and December 2025, found video views down 36% year on year across every page size. Both can be true at once. Platform-wide watch time grew while views per page fell, because the amount of video being published grew faster than the audience watching it. Rising demand for video does not mean your individual video gets more of it.

Documents (native)7.00%Multi-image6.45%Video (native)6.00%Single image5.30%Text only4.50%Polls4.20%Link posts3.25%
LinkedIn engagement rate by post format, from Socialinsider's analysis of 1.3 million posts across 16,645 pages (January 2024 to December 2025). Link posts sit last, but LinkedIn has never published a rule demoting them, so the gap is measured and the cause is inference.

What LinkedIn said it is trying to suppress

Alongside the rebuild, LinkedIn said it would spend the following months reducing repetitive, click-driven posts and filtering out engagement bait, naming the comment-if-you-agree pattern and recycled thought leadership as targets. The move to LLM-based retrieval also changes what topical relevance means. LinkedIn's own example is that someone interested in electrical engineering who engages heavily with posts about small modular reactors would previously have been missed by keyword matching. Two consequences follow for anyone posting video. Formulaic hooks that worked by extracting comments are now a liability rather than a lever. And a genuinely specific post about a narrow subject is more findable than it used to be, because the system is matching meaning rather than tags.

What is inference rather than documentation

A lot of confident LinkedIn advice has no published basis, and it is worth separating out. The golden hour, the claim that engagement in the first sixty minutes decides a post's fate, is a plausible inference from how ranking systems test content, but LinkedIn has not published a window or a threshold. The specific dwell-time targets circulating in creator posts are not LinkedIn's numbers. The exact weighting of comments against likes is not published. Neither are hashtag effects, an optimal post length, or a penalty for editing a post after publishing. None of that makes the advice wrong. It means you should hold it loosely and test it on your own account instead of treating it as documented fact.

The practical version for video

Strip it back and the documented parts all point the same way. Upload video natively rather than linking out, because dwell accrues in the feed and a link ends it. Burn in captions, since a large share of professional viewing happens with sound off, a claim usually attached to an 85% figure that has no primary source but a direction nobody seriously disputes. Write the post text so it stands on its own and front-load it, because the feed truncates after a couple of lines. Post specifically enough that a meaning-matching retrieval model has something to grab. And publish often enough to learn, which for most people means cutting clips out of talks, panels and recordings they already have rather than scripting new ones, which is the workflow FrameOS is built around.

FAQ

How does the LinkedIn algorithm work in 2026?

LinkedIn's engineering blog described the current system in March 2026. A single retrieval model built on a fine-tuned language model narrows millions of posts to candidates in under 50 milliseconds, then a transformer-based ranker scores them against more than a thousand of your past interactions. The engagement signals LinkedIn names are long dwells, likes, comments and shares.

What is dwell time on LinkedIn?

It is how long you spend with a post before scrolling on, counting both time viewing it in the feed and time on the content after a click. LinkedIn's engineers adopted it because clicks and reactions are binary and sparse while dwell is continuous and always measurable. Views below a skip threshold are treated as not indicating genuine interest.

Does LinkedIn punish posts with external links?

LinkedIn has never published such a rule. Socialinsider's analysis of 1.3 million posts did find link posts last on engagement at 3.25%, well behind native video at 6.00% and documents at 7.00%. The gap is measured, the explanation is inference. Putting the link in the first comment is a workaround with no documented basis either.

Is video still worth posting on LinkedIn?

The evidence is mixed and worth seeing in full. LinkedIn reported video watch time up 36% year on year in early 2025, with video creation growing twice as fast as other formats. Socialinsider found views per page down 36% over roughly the same window, while video engagement rate rose 7% to 6.00%. More competition, but an audience that engages.

Sources

Related pages