LinkedIn algorithm breakdown: how impressions are distributed

An impression is a display. LinkedIn's combined post analytics defines it as the number of times your post was displayed on screen, and it says the number is an estimate. Members reached is the count of distinct members and Pages who viewed it, so one person can produce several impressions. The same screen splits those impressions into two piles you can actually read: in network, people who follow you or are connected to you, and out of network, people who do neither.

That split is the distribution report. It is not a follower count, and it is not a secret percentage of your network that LinkedIn has published as a first-hour test. On 12 March 2026, Hristo Danchev's engineering post described the system that decides which displays happen. Retrieval picks candidate posts for a member. Ranking decides which of those candidates land on that member's screen. Wider distribution means more members' feeds retrieve the post and rank it high enough to display. Boykli does not calculate that for you. The Algorithm page is a plain-language read of these public notes. Your impression number still lives on LinkedIn.

Two stages, then a display

Danchev's account replaces the older picture of several separate retrieval pipes (a chronological network index, geographic trending, collaborative filtering, industry trending, and embedding search) with one retrieval system, then a sequential ranker.

The feed still mixes people you know with people you have never connected with. Danchev says the point of the rebuild is to surface useful posts from the immediate network and from professionals outside it. The help-center split is how that mix shows up on a single post.

What retrieval actually puts in the prompt

The post is not matched on a bag of hashtags. Danchev describes a prompt library. For a post, the text includes format, author name, headline, company, industry, engagement counts, article metadata, and the post text. For a member, it includes profile, skills, work history, education, and a chronological list of posts that member engaged with positively.

Three engineering choices change how impressions spread.

Meaning travels further than the keyword. The example in the post is an electrical engineer who also reads about small modular reactors. A keyword index can miss that pair. The language model is expected to connect them because the topics sit near each other in its training, not because the posts share a token. A founder who writes "on-call" and "incident review" can be retrieved for a reader who has been engaging with "production reliability," even when the strings do not match. That is the out-of-network path: the reader does not follow you, and the post is still a neighbor of what they already chose.

Raw view counts were a bad feature. Passing views:12345 as ordinary text produced almost no correlation (−0.004) between popularity and how close the member and post embeddings sat. LinkedIn switched to percentile buckets wrapped so the model can tell a count from prose. Their example turns a raw view total into a view percentile such as 71. They report the correlation between popularity and embedding similarity jumped about 30 times, and recall@10 improved by 15%. Popularity still matters. It enters as a bucket, and the bucket is refreshed as the post collects displays and actions.

The member vector learns from what they chose, not from everything they were shown. Training on every impressed post, including scrolls-past, hurt retrieval quality and cost more GPU memory. The production setup keeps positive engagements in the member history. Hard negatives, posts that were shown and not engaged with, still teach the contrast: two of those per member improved recall@10 by 3.6% over easy negatives alone. A skip is useful to the model. It is a poor thing to optimize your draft toward.

Profile fields travel with the post because they are inside the item prompt. Headline, company, and industry are not decorative. If those three describe a different profession from the post, retrieval is matching two stories at once. Geography is also a profile signal Danchev names, alongside industry, experience, and skills. Demographic attributes are not ranking inputs, and the post says the models are audited so creators compete on equal footing.

How a display is scored once you are a candidate

Retrieval does not publish the post to a fixed slice of followers. Ranking scores it for one member at a time. The Generative Recommender treats that member's feed history as an ordered story. Danchev's example is machine learning on Monday, distributed systems on Tuesday, and a Wednesday session that treats those as one trajectory. Actions are stored next to the posts: long dwells, likes, comments, and shares. After the sequence model, extra context (device, profile embeddings, counts, affinity) is fused in. Passive tasks and active tasks get separate gating. Skip is a prediction, not a vibe.

The dwell definition is older than the March 2026 rebuild, and the rebuild still names it. Fengyu Zhang, Mohit Kothari, and Birjodh Singh Tiwana (1 October 2024) define dwell time as how long a member spends on a post. They separate it from likes, comments, and shares because many weekly active members read and do not react, and because out-of-network posts often have few explicit actions. A short view below a threshold they call Tskip is a skip: engagement probability near zero, used as a negative term. A universal "long dwell" cutoff failed their tests. They use a 30-second line as the illustration of that failure, because 30 seconds is a different event on an image than on a video, and because the median dwell on a format moves over months. The model they shipped predicts whether a member will dwell longer than a percentile of similar posts, with the percentile recomputed daily inside groups such as content type, creator type, and distribution method. The March 2026 ranker still lists long dwell next to click and skip.

So the practical reading of "the algorithm wants dwell" is narrower than the roundups. LinkedIn has published that short dwell is a skip signal, that long dwell is relative to the format, and that it does not want dwell bait. It has not published a public second-count that unlocks a wider impression tier. Treat any "61 seconds" or "first 90 minutes to 8% of followers" line as unpublished.

What gets fewer impressions on purpose

Distribution is also a quality filter, documented outside the ranking paper.

On 12 March 2026, the same day as the engineering post, Tim Jurka's product note said the feed would show less repetitive, click-driven material and less engagement bait, including prompts in the shape of "comment yes if you agree."

On 4 June 2026, LinkedIn's newsroom said content that looks generated and lacks a clear perspective is less likely to be distributed beyond the author's immediate network. The same note says early tests identified that generic content correctly 94% of the time, and that automated comments and replies that only restate the post are in scope.

On 15 September 2026, a second newsroom post said those efforts had reduced views of content classified as AI slop by 40%, that more than one million members had used a "Seems like AI slop" control since August, and that the company was blocking hundreds of thousands of automated comment attempts a day. "AI slop," in that post, is generic, repetitive, recycled, or built to game attention. It is LinkedIn's classifier and LinkedIn's 40% figure. It is not an independent audit, and it is not a claim that every AI-assisted draft is suppressed. A specific post with a specific experience can be drafted with a model. A recycled list with a bait comment is what the notes describe limiting beyond the people who already follow you. A line-by-line split between a token closer and a post that already contains the point is in engagement bait versus deep value.

The named actions that shrink that impression count — bait closers, pods, automated comments, recycled generic posts, and website automation LinkedIn treats as abuse — are the teardown in LinkedIn algorithm penalties.

That is why a hook pack and a comment pod are weak levers in this system. The first lines still matter, because a skip is predicted before a like exists. The line has to be about the post you actually wrote. How Boykli builds that first line from your own note, and refuses a borrowed stat, is in the LinkedIn hook generator guide. The reason the draft has to start from material you brought is the constraint in building Boykli.

Claims worth retiring

Sponsored impressions are a different meter. Campaign Manager has its own viewability rules for ads. Organic post analytics are the displays described above. Do not add the two totals and call the sum reach.

Read one post's distribution, then hold the variables still

Do this on a personal profile. Boykli publishes only to the personal profile you connect with OAuth. A company Page uses LinkedIn's Page analytics, and this product does not queue to it.

1. Open the post on LinkedIn and read the discovery block. Write down impressions, members reached, in-network percent, and out-of-network percent. Note the format: text, image, document, or video. How those three feed objects (text, PDF carousel, native video) sit against format-relative dwell is video vs carousel vs text impressions. Note the topic in one sentence, using your words.
2. Divide nothing by your follower count and call it a grade. Followers who never open the app cannot produce an impression. Members reached is the closer headcount, and LinkedIn already warns that both figures are estimates.
3. Read the first two lines against the skip problem. If those lines could sit on a different author's post about a different week, the predicted skip has little to contradict it. Rewrite the opening from the specific fact in the body. The hook generator path does that swap inside a draft you already have. It does not fetch a viral opener from someone else's analytics.
4. Read your headline, company, and industry next to the post. Those fields are in the retrieval prompt. A headline that names the work makes the post easier to neighbor with the right readers. A headline that names a different market splits the match.
5. Publish the next three pieces in the same format, on the same topic, at one clock. Schedule on Boykli is that experiment: posting-day toggles, one time of day, one timezone. The hour itself is chosen in best time to post on LinkedIn. Changing the hour and the topic in the same week makes the impression delta unreadable.
6. Compare the discovery split across those three, at least a day after each one, because the combined analytics graph is described as complete through yesterday. Look at whether out-of-network displays moved, and whether members reached moved with impressions. A post that only recirculates among people who already follow you can still be the right post. It is a different outcome from one that retrieval is handing to strangers.
7. If several on-topic posts stay almost entirely in network, change one input that the prompt actually contains: the specificity of the text, or the headline that rides with the author. Do not add a comment-for-comment arrangement. Jurka's March note and the June newsroom note both describe coordinated, generic, and bait engagement as material the feed is being tuned to show less of.

The Algorithm page on boykli.com walks the same public posts in shorter sections, with the diagrams that belong next to Danchev and Zhang. It will not import this post's impression row. After you publish, LinkedIn's own analytics remains the meter.

FAQ

How are LinkedIn impressions distributed in 2026?

A member opens the feed. Retrieval embeds that member against millions of posts and returns candidates in under 50 milliseconds. A Generative Recommender scores those candidates from more than a thousand of the member's past interactions, predicting skip, click, long dwell, like, comment, and share. A display on screen is an impression. As engagement changes, the post's embedding refreshes within minutes, so later sessions can retrieve it for a wider or a narrower set of members. The share you can see is in-network versus out-of-network on the post's analytics.

Does the first hour decide the rest of the impressions?

LinkedIn has not published a first-hour cohort size. It has published that a new post is embedded in near real time and that popularity percentiles update as the post collects activity, on a pipeline that stays fresh within minutes. Early reading and early actions can move the bucket. They are not a published switch at minute 60. Put the post out when the readers are likely to be in the app, then judge the discovery split the next day. The creator name for that opening, Buffer's unconfirmed 60- and 90-minute window, and what to do in the hour you can sit are in how early engagement affects reach.

Do comments distribute a post further than likes?

Both are active predictions, along with shares. Long dwell and skip are separate passive predictions, and LinkedIn's 2024 write-up says a large share of feed use is reading without a reaction. No public weight in these sources says a comment is worth a fixed multiple of a like. A comment that only repeats the post is called out in the June 2026 authenticity note. A comment that adds a fact is a different object, and this guide will not invent a multiplier for it.

Will LinkedIn suppress a post because a model helped draft it?

The September 2026 newsroom post defines the reduced class as generic, repetitive, recycled, or built to game attention, and reports a 40% drop in views of what its classifier calls AI slop. The June note says perspective-free generic content is less likely to travel beyond the immediate network. Using a model to turn a real incident, a real number, or a real transcript into a draft is a different act from publishing a template. Boykli's writer is built around source material you supply. It does not promise a distribution outcome, and it does not show impression analytics that could support that promise.

Where do I see the distribution if Boykli has no analytics?

On the post, in LinkedIn's discovery block: impressions, members reached, in-network, out-of-network. Boykli's Algorithm page explains the public engineering notes. It does not rank your archive. The Library stores what you wrote in Boykli. The queue sends the next approved item to the personal profile at the slot you saved.

Does a link off LinkedIn cut impressions?

These engineering and newsroom posts do not publish an off-site link penalty with a number. What they do publish is a lower chance of wide distribution for repetitive and bait posts, a skip prediction for short dwell, and a retrieval prompt that includes the post text and the author's professional identity. If the useful part of the post is entirely behind a click, the on-feed dwell the ranker can observe is the preview. Put the point in the post. Use the link for the evidence.