LinkedIn dwell time: what LinkedIn's own engineers published

Dwell time is how long a member spends on your post before scrolling past, and LinkedIn has been ranking the feed on it since 2020. This is not a theory. LinkedIn's engineering team published the model: a classifier that predicts P(skip), the probability that a member's dwell time on an update falls below a threshold called Tskip. Later work replaced that with a “long dwell” classifier, and in LinkedIn's most recent published feed model, long dwell is one of only two headline objectives. The part almost every article gets wrong: LinkedIn has never published the threshold in seconds. If a page tells you the number is three seconds, or seven, it is not quoting LinkedIn.

What is dwell time on LinkedIn?

LinkedIn measures two different things and calls them both dwell time. The distinction matters if you want to reason about your own posts.

The first is feed dwell time. In LinkedIn's own description, it begins “when at least half of a feed update is visible as a member scrolls through their feed”. That is a rendering event, not an interaction. You do not have to tap anything for the clock to start. Half of your post enters the viewport and the measurement begins.

The second is post-click dwell time: the time a member spends on the content after clicking through. For a shared article that means the time on the article. For a document post it means time inside the document viewer.

LinkedIn gave three reasons for preferring dwell time over the alternatives. It is “always measurable”, it provides a “real-valued measure of engagement” rather than a yes or no, and it is a “more reliable indicator of engagement” than a binary click. Those three phrases are worth holding onto, because they explain the design of everything that came after.

Why did LinkedIn start measuring dwell time instead of clicks?

Because clicks lie in two different directions at once, and LinkedIn said so plainly.

“Clicks are noisy indicators of engagement. For example, a member may click on an article, but quickly close out, realizing it's not relevant, and return to the feed within a few seconds. We call these ‘click bounces.’”

That is the first failure. A click can register on content the member immediately regretted opening. Optimising for clicks rewards the headline that oversells.

The second failure is the opposite problem. LinkedIn noted that “click and viral actions can be rare, especially for passive consumers of the feed”. Most members read far more than they react to. If the only signal you collect is a like or a comment, the majority of the audience is invisible to your ranking model. Every post they read carefully and every post they scrolled past look identical.

Dwell time fixes both. It is present on every impression, and it separates a post that was read from a post that was skipped even when nobody touched a button. This is the single most useful thing to understand about how the LinkedIn algorithm decides what to show: the feed is not scored only on the reactions you can see.

How does LinkedIn's skip model actually work?

The 2020 objective was stated as one line:

P(skip) = P(member's dwell time on this update < Tskip secs)

Read it backwards and it is simple. Pick a threshold in seconds. Any update a member looked at for less time than that counts as skipped. Train a model to predict, before the update is ever shown, how likely that member is to skip it. Then rank so that fewer likely-skips reach the top of the feed.

The interesting question is how they chose the threshold, and LinkedIn answered it. They plotted the probability that a member takes any action on an update against how long they dwelled on it. The threshold was set at “the value T where the blue curve, P(action on update | dwell time = T), starts to become non-zero”. Below that point, essentially nobody ever acts. Above it, some people start to. The threshold is the moment attention becomes capable of producing a response.

LinkedIn also reported something counterintuitive that survives to this day in watered-down form: “a single choice of the threshold Tskip works well for all of the heterogenous update types”. Text, image and video updates shared one number. The model was a logistic regression, and the features that helped most were, in their words, “a combination of member-update features...together with update-side features”. In plain language: who is looking, and what is being looked at, in combination.

Results were reported as an offline area under the ROC curve improvement of “as much as 10% over multiple trainings”, and online, “a large decrease in the number of skipped updates” alongside members who “interacted much more with their feed updates through clicks and viral actions”.

How long is “long dwell” on LinkedIn?

This is the section every other page on this topic gets wrong, so here is the honest answer: LinkedIn has never published the number.

The 2020 post names Tskip, describes how it was chosen, and shows a chart. It does not print the value. And by 2024 the fixed threshold was gone. The LiRank paper describes the successor as “a ‘long dwell’ binary classifier predicting whether there is more time spent on a post than a specific percentile (e.g., 90th percentile)”.

A percentile, not a constant. And not one percentile either. LiRank says: “Specific percentiles are determined based on contextual features such as ranking position, content type, and platform, forming clusters for long-dwell threshold setting.” The bar for a video near the top of the feed on mobile is a different bar from a text post in position forty on desktop. And the bars move: “By daily measuring cluster distributions, we capture evolving member consumption patterns and reduce bias and noise in the dwell time signal.”

LinkedIn's February 2026 sequential recommender paper says the same thing in one clause: the long dwell threshold “varies by post type”.

So the correct answer to “how many seconds do I need?” is that there is no fixed number, that whatever number applies to your post today is relative to what comparable posts in comparable slots are currently getting, and that the target moves daily. A page quoting a constant is quoting itself.

What does LinkedIn optimise for now, and where does dwell time sit?

Four published sources describe the objectives directly. Laid side by side they show a clean progression rather than a revolution.

Two things fall out of that table. First, dwell has been a first-class objective for six years and has grown in status, not shrunk. In the 2026 recommender paper it is one of two named targets, sitting next to “Contribution”, which the paper defines as likes, comments or shares.

Second, and more useful for anyone writing posts: LinkedIn consistently splits “passive tasks (click, skip, long-dwell) and active tasks (like, comment, share)” into separate parts of the model. They are not one score with dwell weighted in. They are different predictions that get combined. A post can be strong on one and weak on the other, and the common advice that dwell time has “replaced” engagement is not supported by anything LinkedIn has published.

How much does dwell time actually move the feed?

Less than the excitement suggests, and LinkedIn is the one telling you so. LiRank's production win from its whole ranking overhaul, dwell modelling included, was relative improvements of 0.8% in overall time spent, 1% in time spent per post and 0.2% in member sessions. The 2026 sequential recommender reports +2.10% overall time spent against the model it replaced, rising to +2.38% among daily active members and falling to a statistically insignificant result for new members.

Those are platform-level numbers across a billion-member product, where a 2% shift in time spent is an enormous engineering outcome. They are not a promise that one post optimised for attention gets 2% more reach. There is no published mapping from a single post's dwell performance to a single post's distribution, and anybody offering you one has made it up.

LinkedIn added a line at the end of the 2020 post that is worth quoting because it explains why dwell time is not a lever you can pull indefinitely: “As we explore these areas, we remain cognizant that we don't want to blindly increase members' time spent on the feed.” The objective is attention that predicts action, not attention for its own sake. That is precisely why Tskip was set at the point where the probability of an action stops being zero.

Which post formats hold attention longest in 12,988 posts?

LinkedIn does not publish dwell time by format, and we cannot measure dwell time at all. What we can do is look at which formats travel, in a dataset where the underlying ranking system was already optimising for dwell. Our cohort is 12,988 English posts from 65 creators with 1,000-plus followers, scored by (reactions + 4 x comments) divided by followers.

Native video over-indexes by roughly 2.6 times in the top decile. Shared article and link posts run the opposite way, making up a much larger share of the bottom half than the top. Image posts, the most common format in the whole cohort, are also over-represented at the bottom.

Set that next to what LinkedIn said about click bounces and the shape makes sense. A shared link is an invitation to leave. If the member clicks and closes the tab within a few seconds, LinkedIn has both a click and a very short post-click dwell, which is the exact pattern the 2020 model was built to stop rewarding. Video and long text hold the member on the update itself, where feed dwell time accumulates. We went through the contradictory published research on this in what actually happens when you put a link in a LinkedIn post, and the honest summary there is that studies disagree while LinkedIn denies any explicit penalty exists.

The rate table tells the same story with a different denominator. These are median engagement rates per 1,000 followers with post and author counts attached.

Document posts lead the table, and we would not build a plan on that row: 60 posts from 13 authors is thin, and carousels were novel during the period the data covers. The rows worth trusting are the big ones, and the pattern in them is that formats which keep the reader on the post beat formats which send them elsewhere. Text-only posts carry no media at all and still sit level with video. The format is not doing the work.

Does writing a longer post increase dwell time?

It is the most common piece of dwell time advice and it does not survive contact with our data. Median visible hook length in the top decile is 206 characters. In the bottom half it is 205. In words, 36 against 34. Length is not what separates the groups.

A caveat we have to state: the source dataset truncates post text at LinkedIn's “see more” fold, so we are measuring the visible hook rather than the full body. That is a limitation for questions about total post length. It is not a limitation for this one, because the fold is where the skip decision happens. A member who does not open your post has already produced a dwell time, and it is a short one.

Mechanically, more words below the fold cannot increase dwell time on a post nobody expanded. The expansion is the event. That is why our work on how long a LinkedIn post should actually be keeps landing on the same conclusion: optimise the decision at the fold, then write however much the idea needs.

What raises dwell time in the first 200 characters?

Nothing in the published research answers this, so treat what follows as correlation in our cohort rather than a rule. Three patterns separate the top decile from the bottom half by a wide margin.

• First-person openers. 19.6% of top-decile hooks open with I, my or we, against 10.1% of bottom-half hooks. Roughly double.
• Second person anywhere in the hook. 47.5% against 32.7%. A hook that names the reader gives them a reason to keep reading past the fold.
• Hashtag restraint. Four or more hashtags appear in 25.9% of top-decile hooks and 41.9% of bottom-half hooks. Hashtags in the visible portion consume the exact characters that could have been the argument.

And two patterns that do nothing, which is more useful than it sounds. Numbers in the hook appear in 32.1% of top-decile posts and 32.8% of bottom-half posts. Question marks separate the groups only slightly, 19.0% against 14.2%. If your dwell time plan is “add a statistic and a question mark”, our data says you have not changed anything. There is more of this in the breakdown of what top-decile hooks do differently and in how to open a LinkedIn post.

How do you test dwell time on your own account?

You cannot measure it directly. LinkedIn's post analytics report impressions, defined as “the number of times your post was shown on LinkedIn”, and members reached, “the number of distinct members and Pages that saw your post”. Neither is dwell time. There is no dwell metric in any public LinkedIn report.

Here is the closest honest procedure.

1. Pull your last 30 posts. Record date, format, impressions, members reached, reactions and comments.
2. Compute impressions divided by members reached. A higher ratio means the same people saw the post more than once, which is what happens when a post keeps getting re-surfaced rather than shown once and dropped.
3. Split by format. Compare your median impressions for native video and text-only posts against your median for shared links. If the link posts sit below your own baseline, you have reproduced the pattern in the table above on your own account.
4. Run a fold test. Take a post that underperformed and repost the same idea weeks later with the argument moved above the fold and the hashtags moved to the end. Compare impressions. One test proves nothing; ten start to mean something.
5. Watch comments, not reactions. Comments are the scarcer signal in our data by a long way: a median of 72 for top-decile posts against 5 for bottom-half posts. They are also the clearest sign that someone read to the end.

One thing not to do: chase time-on-post by padding. LinkedIn's threshold was set at the point where the probability of an action becomes non-zero, which means the signal they want is attention that leads somewhere. Attention that leads nowhere is what they said they do not want to blindly increase.

Does dwell time explain a sudden drop in reach?

Usually not on its own. Dwell time is one input into a ranking function that also predicts clicks, skips, likes, comments and shares, and the whole thing sits behind a retrieval stage that has itself changed shape. LinkedIn's March 2026 engineering post describes a retrieval system built on “LLM-generated embeddings” feeding a generative recommender that processes “more than a thousand of your historical interactions”. A post can fail at retrieval and never get the chance to be dwelled on.

If your numbers fell off a cliff, the diagnostic order we would use is in why LinkedIn reach drops, and the platform-wide version of the question is in whether LinkedIn organic reach is genuinely declining. Dwell time is rarely the first thing to check.

What are the limits of everything on this page?

• LinkedIn publishes architecture, not weights. Every paper cited here describes what is modelled. None of them publishes the coefficients, the threshold values, or how the predictions combine into a final score.
• The 2020 post is six years old.Its concepts survive in the 2026 work, but the logistic regression it describes does not. Treat it as the origin of the idea rather than a description of today's system.
• Our data is correlational. Nothing here shows that switching format causes a rate change. It shows which posts sit where.
• Our cohort skews established. 65 creators, all above 1,000 followers, many far above. Smaller accounts behave differently.
• Engagement counts are lifetime totals at scrape time, and post ages vary, which flatters older posts.
• We measure the visible hook only, because the source truncates at the fold.
• We cannot measure dwell time. No public dataset can. Anyone claiming a dwell time benchmark for your niche is estimating.

LinkedIn dwell time, restated

Dwell time is a real ranking signal with a documented history. Feed dwell time starts when half your post is visible. LinkedIn built it because clicks bounce and because most members never react to anything. The 2020 model predicted the probability of a skip against a threshold called Tskip. The current model predicts whether a post is dwelled on past a percentile that varies by position, content type and platform, and that percentile is recalculated every day.

What you should take from that is narrower than most articles suggest. There is no number of seconds to hit. There is no published route from one post's dwell performance to one post's reach. The reported platform gains are in the single digits of a percent. The practical version is unglamorous and matches what the rest of our study of 12,988 posts keeps finding: earn the expansion at the fold, keep the reader on the post, and give them something specific enough to answer.