What LinkedIn is testing
LinkedIn's feed has always mixed two kinds of post: those from people a member has chosen to connect with or follow, and those LinkedIn recommends from beyond that circle. For most of its history both arrived in a single stream. On Friday 18 September 2026, LinkedIn's chief product officer, Hari Srinivasan, posted that the company was testing something different.
“We're testing a new design on the top of the feed,” he wrote. “The Network tab helps you stay connected to your network, ie the people you have explicitly connected or followed. We all want more real connection with real people, and we hope this plays a small part in rebuilding that world.” The second tab, For You, “helps you stay informed” with posts “from real practitioners inside and outside your network”. He added: “There's more to come here too.”
That is the whole of LinkedIn's public statement. It does not say how many members see the test, in which markets, whether the Network tab is ordered by time or by a ranking model, or which tab opens by default. Trade coverage has filled some of the gaps, and should be read as such. Social Media Today, reporting the post on 20 September, noted that LinkedIn ran other alternative-feed tests in 2025 and in June 2026 that have not been expanded. Le Blog du Modérateur reported on 28 September that the test was on mobile and placed it in a series of experiments going back to February 2025, none yet made general. An Inc. column on 27 September described the split as rolling out; we could not read the column itself, and LinkedIn has not used that word.
For an executive who posts on LinkedIn, the test matters for a simple reason. If members can choose a tab that contains only their own network, then being in that network, rather than being recommended into someone's feed, becomes a distinct way of being seen. To know how much that matters, it helps to know how the feed already works. LinkedIn has said more about that this year than it usually does.
What LinkedIn says decides distribution
LinkedIn published two accounts of its feed on 12 March 2026: a long post on its engineering blog, “Engineering the next generation of LinkedIn's Feed”, and a shorter member-facing article by Tim Jurka, “Updates to The LinkedIn Feed Focusing on Authentic, Relevant Conversations”, to which the engineering post links. Both describe the same system. In LinkedIn's words, “Our systems evaluate a number of signals, like information you've chosen to share on your own profile such as industry, experience, skills and geography. Then, the system looks at how you engage with content over time, including what you've read, liked, commented on, returned back to, or simply scrolled past.”
Three details in the engineering post are useful to anyone trying to be seen.
Posts are read as text, and the author's profile is part of that text. LinkedIn replaced several separate ways of finding candidate posts, including “a chronological index of your network's activities, trending posts in your geography, collaborative filtering” and others, with one retrieval system built on a large language model. To feed it, each post is converted into a text prompt that includes “format, author information (name, headline, company, industry), engagement counts, article metadata, and post text”. Engagement counts enter as percentile bands rather than raw numbers. LinkedIn's own example is that a member interested in electrical engineering may be shown posts on small modular reactors, because the model knows the subjects are related.
Ranking follows a member's recent trajectory. The ranking model, which LinkedIn calls a Generative Recommender, “processes more than a thousand of your historical interactions” in order, so that what a member engaged with this week shapes what they see next. The actions it predicts are long dwells, likes, comments and shares. A paper by 24 LinkedIn authors describing the model says it had been serving “the majority of LinkedIn's Feed traffic for over three months” and lifted time spent by 2.10 per cent and likes, comments or reshares by 3.52 per cent in A/B tests.
Some behaviour is being pushed down. Jurka's article says LinkedIn is working to make engagement pods “ineffective” and to “curb comment automation, including the use of third party software or browser extensions, which are not allowed on LinkedIn”. It promises less engagement bait (“Comment ‘Yes’ if you agree”), fewer posts whose video “has nothing to do with the associated text post”, and fewer “‘recycled’ thought leadership posts that don't say much in terms of substance or insight”. Its stated aim is a feed that is “not a popularity contest”.
How much the network matters
LinkedIn's blog posts describe the system in general terms. The clearest number comes from a research paper. “Connected Content Retriever: Dense Graph Edge Features Powering Pre-Ranking at LinkedIn”, by nine engineers at LinkedIn in Mountain View, was posted to arXiv on 18 September 2026, the same day as Srinivasan's post. It describes the stage of the feed that narrows tens of thousands of candidate posts from a member's network to a few hundred for final ranking, within 120 milliseconds.
Its opening sentence is the figure that has since circulated: “content generated by a member's network (connections and follows) makes up over 70% of impressions and engagement.” That network has two layers. The first degree is posts by people a member is connected to or follows. The second is “posts that a 1st-degree connection reacted to, commented on, or reshared but did not author (a.k.a. stranger viral)”. The introduction is blunter about what that means: distribution “is fundamentally graph-mediated rather than globally popularity-ranked”, and the strength of the connection between viewer and author is treated “as a first-class ranking feature, co-equal with topical interest”.
The paper measures how much the model depends on that connection. Its main example of an “edge feature” is viewer–author action affinity, which records a viewer's past engagement with each author “across action types (like, comment, share, click, long-dwell, follow, etc.) and time buckets (last hour, last day, last week, etc.)”, capturing “both the intensity and the recency of engagement”. When the engineers removed each kind of input in turn and measured how well the model picked the posts the final ranker would choose, losing the viewer–author features hurt most.
Fall in recall@10, per cent, relative to the full model
Source: Gupta, Srinivasa Ramanujam et al., Connected Content Retriever, LinkedIn, arXiv 2609.22441, 18 September 2026, Table 2.
An offline test on a fixed dataset of the connected-content pre-ranker only, not the whole feed. Recall@10 measures how many of the final ranker's top ten posts the pre-ranker also put in its top ten.
The paper adds that when LinkedIn tried running this stage on language-model content embeddings alone, without the connection features, it saw “a statistically significant degradation in online engagement metrics”. In other words, for posts from a member's own network, LinkedIn's engineers found that knowing what a post is about was not enough; they needed to know how much this viewer has engaged with this author, and how recently.
Two cautions. The paper is about one stage of one part of the feed, and its offline numbers are not a measure of anyone's reach. And the 70 per cent figure is an aggregate across the feed; it says nothing about the share for a particular member, market or type of post.
What the split feed changes for an executive
LinkedIn has not said what the Network and For You tabs mean for authors. What follows is our reading of its published descriptions, and should be taken as inference.
The Network tab rewards a network that wants to hear from you. Its contents are, by Srinivasan's definition, posts from people a member has “explicitly connected or followed”. Inside the existing feed, LinkedIn's engineers describe connection strength, built from past likes, comments, clicks, long reads and follows, as co-equal with topic. An executive with many connections who rarely interact with them is in many networks but close to few. The paper's time buckets, measured in hours, days and weeks, suggest that recent interaction counts; it does not say how quickly older interaction stops counting.
Comments travel. The second-degree route described in the paper means that when a first-degree connection comments on or reshares an executive's post, it can enter that person's network's feeds. The same mechanism works in reverse: a substantive comment by an executive on someone else's post is itself an activity their own connections may be shown. LinkedIn's crackdown on automated comments and pods is aimed at faking exactly this route.
The For You tab is where the profile text works. For recommended posts from outside a member's network, LinkedIn's retrieval model reads each post as text that includes the author's name, headline, company and industry, and matches it to the viewer's interests. A headline that says plainly what someone is expert in, and posts that stay on that subject, give the system something to match. That is consistent with what we wrote about profile basics in executive LinkedIn presence and with the idea of being known for one thing set out in topical authority; LinkedIn's descriptions give it a mechanism.
Popularity alone matters less than it looks. LinkedIn says it uses engagement counts as a signal, in percentile bands, but its engineers describe network distribution as not “globally popularity-ranked”, and its product team says it is deliberately suppressing engagement bait. For an executive, a post that a few hundred relevant people read to the end is likely to travel further through the network than one that collects reactions from strangers through a pod.
What we cannot tell from LinkedIn's statements is how many members will use the Network tab if it launches, or whether For You will remain the default. If most members stay on For You, the split may change little; if they switch, posts that depend on recommendation to strangers would lose a share of their audience.
What LinkedIn has said, and what others say it said
LinkedIn's feed attracts a large advice industry, and the week's coverage shows how easily claims drift from their source.
An old explainer presented as news. A WeRSM article updated on 28 September, headlined “LinkedIn Explains How Its Feed Algorithm Works”, summarises LinkedIn's “Mythbusting the Feed” post. That post is dated 18 August 2022, and its author's profile on the page now gives her role as being at TikTok. WeRSM's text does call it a “2022 product explainer”; the headline does not. The 2022 post's advice, to follow the people you want to hear from and engage with their posts, is consistent with what LinkedIn says now, but it is not a description of the 2026 system.
The 70 per cent figure, accurately reported. Leaders.social reported the Connected Content Retriever paper correctly, with a section on what the paper does not prove, including that it gives no breakdown for company pages. That is a fair reading: the paper is about members' networks.
Terms LinkedIn has not used. SocialPilot's “September 2026 update” guide says a “Depth Score” replaced engagement rate and that “external links” are “actively deprioritized”. Neither phrase appears in LinkedIn's March 2026 engineering post, Jurka's article, the 2022 mythbusting post or the Connected Content Retriever paper. The same guide attributes the 2026 feed to a model called 360Brew; LinkedIn's March posts describe LLM-based retrieval and a Generative Recommender without using that name. LinkedIn does not publish everything about its feed, and such claims may rest on testing we have not seen; but they are not LinkedIn's statements, and an executive should weigh them accordingly.
This matters beyond LinkedIn. How AI answer engines and search systems choose what to show is subject to the same cycle of vendor claims and folklore, discussed in what signals answer engines use to judge a source.
Methodology and limitations
We read Hari Srinivasan's LinkedIn post of 18 September 2026 on its public page, dating it from the post's activity identifier and from Social Media Today's report. We read LinkedIn's engineering blog post “Engineering the next generation of LinkedIn's Feed” and Tim Jurka's LinkedIn article of the same date (12 March 2026), LinkedIn's “Mythbusting the Feed: How the Algorithm Works” (18 August 2022), and two arXiv papers by LinkedIn authors: “Connected Content Retriever” (2609.22441, 18 September 2026) in full, and “An Industrial-Scale Sequential Recommender for LinkedIn Feed Ranking” (2602.12354, version 3) in its abstract and system overview. All were read on 1 October 2026. Coverage by Social Media Today, Le Blog du Modérateur, WeRSM, Leaders.social and SocialPilot was read in full on the same day and is labelled as secondhand where we rely on it; the Inc. column of 27 September did not load for us and is cited only from its headline and search summary. Four limitations apply. LinkedIn's blog posts are descriptions written for the public, not specifications. The research papers describe systems at the time of writing and report offline and A/B results we cannot reproduce. We did not see the Network tab test ourselves, measure any account's reach, or contact LinkedIn. And the section on what the changes mean for executives is our inference from LinkedIn's descriptions, marked as such.
Conclusion
LinkedIn's own engineers have now put a number on what experienced users suspected: most of what members see, over 70 per cent of impressions and engagement, comes from their own network, and for those posts the viewer's history with the author outweighs what the post is about. The Network tab test would make that network a separate destination. For recommended posts, LinkedIn says its system reads each post together with its author's headline, company and industry and matches it to what members have engaged with recently, while working to suppress pods, automation and engagement bait. For an executive, that points away from tricks and towards three plain things: a network of people who actually read you, steady recent activity with them, and a profile and posts that say clearly what you know. Much of the advice in circulation goes further than LinkedIn has; the safe course is to rely on what LinkedIn and its engineers have said, and to treat the rest as unconfirmed.
Frequently Asked Questions
What decides who sees a LinkedIn post?
By LinkedIn's own account, two things: what the viewer has shared on their profile and how they have engaged with posts over time, including what they read, liked, commented on, returned to or scrolled past. For posts from people the viewer is connected to or follows, a paper by LinkedIn engineers says the viewer's past interactions with that author are the strongest signal its pre-ranking model uses.
What is LinkedIn's Network tab?
A test, announced on 18 September 2026 by LinkedIn's chief product officer, Hari Srinivasan, that splits the top of the feed in two. The Network tab shows posts from people a member has connected with or follows; the For You tab is for recommended posts from inside and outside that network. LinkedIn has not said how widely the test is running or whether it will launch.
How much of the LinkedIn feed comes from a member's own network?
Over 70 per cent of impressions and engagement, according to a paper by nine LinkedIn engineers published on arXiv on 18 September 2026. That counts posts by connections and follows and posts those people reacted to, commented on or reshared. It is an aggregate figure, not a share any individual should expect.
Does the LinkedIn algorithm read an author's headline?
LinkedIn's engineering blog says that when it turns a post into text for its language-model retrieval system, it includes author information: name, headline, company and industry, alongside the post's format, engagement counts and text. So the author's profile is part of how the system represents what a post is about.
Does LinkedIn penalise posts with external links?
We found no statement saying so in the LinkedIn sources we read, which include its March 2026 engineering post, its March 2026 feed update and two 2026 research papers. Claims of a link penalty, and of a ranking measure called a depth score, come from third-party guides; LinkedIn may not have published everything, but those claims are not LinkedIn's words.
What should an executive do differently because of the split feed?
LinkedIn's own sources point to three things: build the network of people who actually want to hear from you, since connection strength is a first-class signal; keep posting so that recent interaction with your audience exists; and make the headline and post text say plainly what you are expert in, since that is what the recommendation side reads. These are our inferences from LinkedIn's descriptions, not LinkedIn advice.
Sources
- official Hari Srinivasan (LinkedIn chief product officer), We're testing a new design on the top of the feed, LinkedIn post, 18 September 2026. Network and For You tabs. Read 1 October 2026.
- official LinkedIn Engineering, Engineering the next generation of LinkedIn's Feed, 12 March 2026. Profile and engagement signals; post prompts including author name, headline, company and industry; Generative Recommender over more than a thousand interactions. Read 1 October 2026.
- official Tim Jurka, LinkedIn, Updates to The LinkedIn Feed Focusing on Authentic, Relevant Conversations, 12 March 2026. Engagement pods, comment automation, engagement bait, recycled posts. Read 1 October 2026.
- research Akhilesh Gupta, Sudarshan Srinivasa Ramanujam et al. (LinkedIn), Connected Content Retriever: Dense Graph Edge Features Powering Pre-Ranking at LinkedIn, arXiv 2609.22441, 18 September 2026. Over 70 per cent of impressions and engagement from the network; viewer–author affinity features; feature ablation (Table 2). Preprint, not peer reviewed. Read 1 October 2026.
- research Lars Hertel et al. (LinkedIn), An Industrial-Scale Sequential Recommender for LinkedIn Feed Ranking, arXiv 2602.12354, version 3. Serving the majority of feed traffic; +2.10 per cent time spent and +3.52 per cent contributions in A/B tests; long dwell and contribution objectives. Preprint. Read 1 October 2026.
- official LinkedIn, Mythbusting the Feed: How the Algorithm Works, 18 August 2022. Cited for its date and content, as the source of the WeRSM article. Read 1 October 2026.
- trade press Andrew Hutchinson, Social Media Today, LinkedIn tests Network tab to encourage engagement among connections, 20 September 2026; José Billon, Le Blog du Modérateur, LinkedIn teste un fil d'actualité limité aux relations et aux abonnements, 28 September 2026. Earlier feed tests; mobile test. Secondhand. Read 1 October 2026.
- trade press WeRSM, LinkedIn Explains How Its Feed Algorithm Works, updated 28 September 2026; Leaders.social, LinkedIn Says Networks Drive More Than 70% of Feed Impressions and Engagement, September 2026. Read 1 October 2026.
- vendor SocialPilot, How the LinkedIn Algorithm Works: September 2026 Update. Depth score, external-link and 360Brew claims, cited as third-party claims. Read 1 October 2026.
- blocked source Alicia Teltz, Inc., LinkedIn Is Rolling Out a Split Feed, 27 September 2026. Page did not load for us; cited from its headline and search summary only.