LinkedIn Algorithm 2026: How the Feed Decides What Gets Reach

The LinkedIn algorithm in 2026 ranks every post with one large AI model that predicts how valuable that post will be to each specific viewer. It shows new posts to a small seed audience first, measures dwell time and comment quality, then decides whether to expand distribution across an interest-based feed. That early test, far more than your follower count, decides your reach.

We refresh this post quarterly because the mechanics keep moving, and because most of what gets written about the algorithm is confident guesswork. LinkedIn publishes very little. So below we separate three things carefully: what LinkedIn has confirmed, what independent research consistently infers, and what is folklore.

One model now decides what the feed shows

For years the feed was a stack of hand-tuned models, one predicting clicks, another predicting reactions, another filtering spam. That era is ending. LinkedIn researchers have described 360Brew, a 150 billion parameter decoder-only foundation model built to handle more than 30 predictive tasks across the platform, ranking and recommendation included. One big model, reading your post, your history, and each viewer's behavior, replaces thousands of narrow ones.

The visible consequence is the shift everyone felt in their feed: distribution now follows an interest graph rather than a relationship graph. The old feed showed you people you knew. The 2026 feed shows you topics you engage with, from anyone. Richard van der Blom's Algorithm Insights research, built on data from roughly 1.3 million posts, tracks the fallout: public summaries of his 2026 report describe average post views down roughly 50 percent year over year and follower growth down nearly 60 percent. Reach did not vanish. It got redistributed toward creators the model can confidently match to an interested audience.

The initial test wave, explained

The initial test wave is the small slice of viewers LinkedIn shows a new post to before deciding whether it deserves wider distribution.

Here is the sequence as best independent research can reconstruct it. Your post goes live and enters the feeds of a seed group: some followers, some connections, weighted toward people who are currently active and have engaged with you before. LinkedIn has never said how large this group is; third-party estimates put it at a low single-digit percentage of your audience. The model then watches what the seed does. Long reads, comments, sends, and saves push the post outward to the next, larger circle. Fast scrolls, "show fewer posts like this" clicks, and silence bury it.

Two practical notes follow directly from this design. First, the test only works if your seed audience is awake, which is why the best time to post on LinkedIn is really a question about when your specific readers are online, not a universal magic hour. Second, the test is per-post. A weak post does not doom your account, but a long run of skipped posts teaches the model that your seed audience does not care, and later tests start from a colder position.

Dwell time is the quality score you never see

Dwell time is the time a viewer keeps your post on screen, whether or not they ever click, react, or comment.

This one is not inference. LinkedIn's engineering team wrote publicly about using feed dwell time in ranking back in 2020, including the detail that they model a skip threshold: posts viewed for less than a certain duration are treated as skipped. Everything since suggests dwell has only grown in importance, because it is the engagement signal hardest to fake. A pod can manufacture likes. Nobody can manufacture ten thousand strangers reading to the end.

What earns dwell in practice: a first line that makes the "see more" click feel necessary, formatting a reader can move through without effort, and documents or carousels that hold attention across swipes. What kills it: a post that reveals its entire point in the preview, and openings that read like every other AI-drafted post in the feed, which get pattern-matched and scrolled past in under a second.

Why comments outweigh every reaction

A reaction is one tap. A comment is a decision, and the feed treats it accordingly. Every credible external study ranks a comment well above a like or a repost; the exact multipliers quoted online vary so wildly between studies that we treat the numbers as noise and the direction as settled. The reasoning is easy to reconstruct: a comment costs real effort, it creates new content the feed can rank again, and it keeps two people on the platform instead of one.

Depth compounds the effect. A comment that pulls the author into a reply creates a thread, and threads generate fresh ranking events every time someone answers. Substantive comments also appear to carry more weight than drive-by praise, which is one reason we care so much about how long a LinkedIn comment should be. For small accounts this is the most underused fact about the 2026 algorithm: a deliberate LinkedIn commenting strategy borrows distribution from audiences that already exist, while your own posts are still fighting cold seed tests.

Interest signals and topic authority

Interest signals are the topics the model believes you and each viewer care about, inferred from everything you post, read, comment on, and linger over.

In an interest-graph feed, these signals are the routing table. Post consistently about one subject and the model builds a confident profile: this person is a pricing strategy account, show their posts to pricing people. Post about pricing on Monday, parenting on Wednesday, and crypto on Friday, and the model hedges, testing each post on a mismatched seed and getting mediocre results back. Topic focus is not a branding nicety anymore; it is a distribution mechanism.

This also explains why engagement pods aged badly. Pod members engage out of obligation, not interest, so their audiences rarely match your topic, and the model reads the mismatch. We cover the mechanics in our piece on whether LinkedIn engagement pods work, but the short version is that manufactured engagement from the wrong people now sends a weaker, sometimes negative, signal.

LinkedIn algorithm 2026: what is confirmed, inferred, and myth

Status	Claim
Confirmed by LinkedIn	Dwell time is a ranking input; a skip threshold exists; a large unified AI model increasingly drives ranking and recommendation.
Consistently inferred	The initial test wave; comments outweighing reactions; interest-based distribution; reduced reach for posts whose only goal is an external click.
Folklore	Exact multipliers ("comments are worth 15 likes"); a universal golden hour cutoff; format hacks that work regardless of content; the idea that editing a post resets its reach.

Our rule of thumb: when two studies agree on direction but disagree on magnitude, believe the direction. When a claim comes with a suspiciously precise number and no dataset, assume it was invented for a hook.

How we adapt our own routine

Everything above collapses into a short list. Pick one lane and stay in it. Publish when your actual readers are online. Write openings for the see-more click and bodies for the full read. Then spend at least as much time in other people's comment sections as in your own drafts, because comments are how you increase LinkedIn engagement on an account the feed does not yet trust.

That last habit is the one people abandon first, which is why we built Commenti (our product) around it: it writes comments in your tone on posts from the creators and keywords you choose, with an approve-or-autopost mode for control. We should be equally plain about the trade-off. LinkedIn's User Agreement prohibits third-party automation, Commenti included, and enforcement risk is real, rising with volume, speed, and generic output. Our daily caps and randomized pacing exist precisely because velocity and repetition are the clearest patterns enforcement systems watch for; they reduce that risk, they do not remove it. If your account is brand new or already flagged, do not automate it yet. We lay out the full picture in is LinkedIn automation safe.

The algorithm will shift again before our next refresh. The signals underneath it, held attention, real conversation, and a clear topic identity, have pointed the same direction for five years. Optimize for those and the quarterly changes mostly happen to other people.