Shruti Kaushik

Founder-led growth22 min read

LinkedIn Algorithm Changes in 2026: What Founders Need to Know

What LinkedIn actually changed in its feed in 2026, which reach claims hold up, where studies on links, comments, dwell time and AI content disagree, and what founders should do.

By

Chapters
  1. 011. What LinkedIn has said and published in 2026
  2. 022. The compiled dataset: 20 sources with sample sizes and dates
  3. 033. What the ranking model actually rewards
  4. 044. Reach in 2026: what the large-sample studies show
  5. 055. What the data disagrees on
  6. 066. Shruti's analysis: how I would read this as a founder
  7. 077. What to watch next
  8. 08Methodology
  9. 09Sources

LinkedIn's 2026 changes are a rebuilt feed, not a single penalty. In March 2026 it described LLM-based retrieval and a sequential transformer ranker serving 1.2 billion members, plus less distribution for engagement bait and generic AI posts. The measured effect on links, comments and frequency is disputed across studies.

Most advice about the LinkedIn algorithm in 2026 mixes three different things: what LinkedIn has published, what large samples of posts show, and what is folklore repeated until it looks like a fact. This piece separates them. It compiles 20 sources with sample sizes and dates, reconciles the places where studies contradict each other, and ends with a labelled analysis of what that means for a founder who posts to win customers, hires and investors, not to win a reach contest.

Short version of the findings:

  • LinkedIn has published real detail about its feed in 2026: a February research paper on the ranking model and a 12 March engineering post on retrieval and ranking. Neither contains a posting formula.
  • The widely cited "360Brew runs your feed" claim is not what LinkedIn's own paper says.
  • Reach has fallen for active creators. Richard van der Blom put the loss at about 60% over two years in June 2026, while engagement fell 20 to 25%. Metricool found overall engagement up in the same period. Both can be true.
  • Studies on external links disagree by more than 30 percentage points. The disagreement is explainable.
  • Dwell time is a real model input. The popular "15.6% engagement above 60 seconds" statistic is not from LinkedIn.

Disclosure: Wise Mango is run by the author of this page, and it works with founders on LinkedIn content. Nothing below depends on that.

#1. What LinkedIn has said and published in 2026

LinkedIn published a ranking paper on 12 February 2026 and an engineering post on 12 March 2026 describing LLM-embedding retrieval and a transformer ranker over up to 1,000 past interactions. It also said it will show less engagement bait, and in May said generic AI-looking posts are less likely to spread beyond a member's network.

The ranking paper (February 2026)

The most detailed primary source is "An Industrial-Scale Sequential Recommender for LinkedIn Feed Ranking" (arXiv:2602.12354), first submitted on 12 February 2026 by a 24-author LinkedIn team and later accepted to the 35th ACM CIKM conference. It describes Feed Sequential Recommender (Feed SR), a transformer that replaced a DCNv2-based ranker, deployed for a platform LinkedIn describes as having more than 1.2 billion members. Details that matter:

  • It reports a 2.10% increase in time spent in online A/B tests against the previous production model, and a 3.52% increase in likes, comments or reshares.
  • Daily active users rose 2.38%, weekly 1.84% and monthly 0.82% in the same tests. For new members the effect was not statistically significant.
  • Each member's history is capped at the most recent 1,000 impressions, drawn from one year of training data. Training samples are down-weighted with a 60-day half-life, so a sample two months old counts half as much.
  • Feed SR represents each post with two tokens, one for the post and one for the member's action. The paper says an LLM-based ranker needed hundreds of tokens per post.
  • The authors say Feed SR had served the majority of Feed traffic for more than three months at the time of the latest revision.

The paper also states that the LLM-based text ranker they built never beat the production model online. It did well on out-of-network posts and struggled with network-based recommendations, and it handled numeric features poorly when they were written as text. This matters for the 360Brew debate in section 5.

The engineering post (12 March 2026)

On 12 March 2026, Hristo Danchev published "Engineering the next generation of LinkedIn's Feed" on LinkedIn's engineering blog. As summarized by PPC Land and Search Engine Land:

  • LinkedIn replaced several separate retrieval sources (network activity, trending posts, collaborative filtering and topic systems) with one LLM-embedding retrieval model. A dual encoder embeds member and post prompts and compares them by cosine similarity.
  • Retrieval over an index of millions of posts runs in under 50 milliseconds, and content embeddings can update within minutes.
  • Converting raw engagement counts into percentile buckets raised the correlation between popularity and similarity from -0.004 to about 30 times that, and improved recall@10 by 15%. Adding two hard negatives per member improved recall@10 by 3.6%.
  • The ranker, a Generative Recommender, reads more than 1,000 past interactions and predicts passive actions (click, skip, long dwell) and active actions (like, comment, share).
  • For new members with little history, the model can infer interests from a headline and job title.

Engagement bait, pods and automation

Social Media Today reported on 12 March 2026 that LinkedIn plans, over the following months, to reduce engagement bait such as "comment Yes if you agree", videos unrelated to their text, and recycled thought leadership that adds little. Search Engine Land reported that LinkedIn is acting against comment automation tools, browser extensions and engagement pods, which it says create "inauthentic conversations".

AI-looking content (May to July 2026)

In May 2026, LinkedIn's Laura Lorenzetti wrote that content which appears AI-generated and lacks clear perspective is less likely to be widely distributed beyond a person's immediate network. LinkedIn said early tests identified generic content correctly 94% of the time and did not publish a false-flag rate. The wider evidence on AI-written posts is in the companion piece on AI-written LinkedIn posts.

Posting guidance (July 2026)

LinkedIn's July 2026 posting tips, reported by Social Media Today on 8 July 2026, suggested two to three posts a week, each with a question to prompt discussion.

#2. The compiled dataset: 20 sources with sample sizes and dates

The sources below range from LinkedIn's own papers to vendor studies of 358,000 to 1.8 million posts. Only the first three describe how the ranking system works. The rest measure outcomes, and each measures a different population, so their numbers cannot be stacked into one benchmark.

"Type" shows whether I read the original (P) or a named outlet quoting it (S). Sample sizes are as stated by the publisher.

2. The compiled dataset: 20 sources with sample sizes and dates
# Source and date Type Sample Finding relevant to founders
1 LinkedIn, Feed SR paper (arXiv:2602.12354), 12 Feb 2026, revised to 16 Sep 2026 P 1.2B-member platform, online A/B tests +2.10% time spent; long dwell is one of six predicted actions
2 LinkedIn Engineering, Danchev, 12 Mar 2026 S LLM retrieval and GR ranker One retrieval model; sub-50 ms; 1,000+ interaction history
3 LinkedIn, 360Brew paper (arXiv:2501.16450), Jan 2025 S 150B-parameter research model Foundation model research, not the 2026 feed ranker
4 LinkedIn News, Lorenzetti, May 2026 (updated 4 Jun 2026) S Initial internal tests Generic AI-looking content less distributed; 94% classifier accuracy
5 Social Media Today on LinkedIn feed update, 12 Mar 2026 P n/a Engagement bait and recycled thought leadership to be reduced
6 Search Engine Land on LinkedIn feed update, 2026 P n/a Action against pods, automation tools and extensions
7 Social Media Today on LinkedIn posting tips, 8 Jul 2026 S n/a 2 to 3 posts per week suggested
8 Van der Blom, Algorithm Insights 2026, via Creator Science ep. 307, 2 Jun 2026 S 1.3M posts, 50,000 creators Reach down about 60% over two years; engagement down 20 to 25%
9 Van der Blom, Algorithm Insights 2025, via Mercer-MacKay, 2 Jun 2025 S 1.8M+ posts Organic reach down nearly 50%; first 90 minutes critical
10 Ordinal link study, 2026 P Sample not stated on the page 26.5% lower reach for linked posts, concentrated in company pages
11 LinkPost, via Ordinal, 2026 S 358,000 posts Median 448 vs 705 impressions with and without a link
12 Metricool, via Influencer Marketing Hub, 25 May 2025 S 577,000+ posts, 47,735 pages Linked posts: +4.90% views, +13.57% interactions
13 Metricool LinkedIn study, 14 Apr 2026 S 673,658 posts, 63,108 accounts Comments -17% year on year; engagement up nearly 14%; profiles beat pages by 63%
14 Buffer, 28 Aug 2025 S 2M+ posts More posts per week gave more impressions per post
15 Buffer, 4 Dec 2025 S 72,000+ posts, ~25,000 accounts Authors who replied to comments saw about 30% higher engagement
16 Saywhat, State of the Algorithm, Q1 2026 P 397,605 posts claimed No method or data published on the page
17 Originality.ai, Jul 2026 S 5,000 posts 81.2% classified likely AI
18 Raj, Berg and Seamans, 2025 to 2026 S About 27,000 participants AI disclosure lowered evaluations of writing by 6.2% on average
19 Stanford HAI, 2023 S 7 detectors, TOEFL essays 61.22% average rate at which human TOEFL essays were misread as AI
20 Edelman and LinkedIn, B2B Thought Leadership Impact Report, 2024 S Survey of B2B decision makers 73% of executives say thought leadership is more trustworthy than marketing

Rows 8 to 16 are the outcome studies. None of them is peer reviewed, and only Metricool and Buffer describe their samples in enough detail to judge. Row 16 is the weakest: the Saywhat page gives a post count and a set of headline claims but no data, so I use it only to show the disagreement, not as evidence.

#3. What the ranking model actually rewards

The model predicts six actions: click, skip, long dwell, reaction, comment and share. It reads a member's last 1,000 interactions, weights recent data more heavily, and retrieves posts by meaning, not only by who follows whom. That makes topical consistency and substance the durable levers, and tricks the fragile ones.

Retrieval by meaning

The practical shift from the March post is that a post does not need to match a keyword or sit in someone's network to be retrieved. It needs to be close, in embedding space, to what the viewer's profile and history say they care about. LinkedIn's own example is a member interested in electrical engineering who engages with small modular reactor posts and is then shown grid and renewable energy content. For a founder, this means the profile, headline and post topics are all inputs to who sees you. A founder who posts about hiring one week, fundraising the next and a cricket match after that gives the retrieval model a blurred signal.

Van der Blom's June 2026 advice aligns with this: about 80% of content on one main topic and two to three related subtopics, per the Creator Science episode summary. That is a practitioner's recommendation from his data, not something LinkedIn has stated.

History and recency

Two details from the Feed SR paper are underused. The model sees at most 1,000 recent impressions per member, and older training examples count less on a 60-day half-life. The first means a viewer's feed is shaped by what they recently dwelled on or engaged with, which is why a post can reach people who never followed you if it lands near their current interests. The second means the system relearns quickly; a pattern that worked last quarter has no guaranteed shelf life.

Dwell time

Long dwell is one of the six predicted actions, and Feed SR improves its prediction using dwell-time features. The paper defines it as dwelling on a post longer than a threshold that depends on post type and does not publish the threshold. So the direction is confirmed and the number is not. Posts that people read to the end are more likely to be scored well on this action. I treat that as a reason to write posts with an argument that earns the read, and not as a reason to pad length.

Comments, shares and the "contribution" actions

Reaction, comment and share are the active actions the ranker predicts. Neither the paper nor the blog gives a weight for a comment relative to a like, and any claim such as "a comment is worth 2x a save" is not from LinkedIn. What is documented is the intent: reduce comments generated to farm comments (the "comment Yes" pattern, pods and automation) because they produce what LinkedIn calls inauthentic conversations.

#4. Reach in 2026: what the large-sample studies show

Reach per post has fallen for active creators while overall engagement has not. Van der Blom reported organic reach down nearly 50% in his 2025 data and about 60% over two years in June 2026. Metricool found engagement up nearly 14% between early 2025 and early 2026. Different populations explain the gap.

Four findings are worth stating with their limits.

Reach per post is down for active creators. Van der Blom's 2025 report, based on more than 1.8 million posts and summarized by Mercer-MacKay, put organic reach down nearly 50%. His 2026 report, based on 1.3 million posts from 50,000 creators, was discussed on Creator Science on 2 June 2026: reach down about 60% for active creators over two years, engagement down only 20 to 25%. These are samples of creators who post regularly and are tracked by his panel. They are not random samples of members.

Engagement overall was not falling. Metricool's April 2026 study of 673,658 posts from 63,108 accounts found overall engagement up nearly 14% from January and February 2025 to the same months in 2026, with comments down 17%. It also found personal profiles had a 63% higher average engagement rate than company pages. The detail is in the LinkedIn statistics compilation.

Personal profiles outperform pages. Ordinal's link study found company pages absorb the link penalty while personal profiles show almost none. Metricool's 63% gap points the same way. For a founder-led company, the person remains the stronger distribution channel. The pipeline side of that argument is in founder LinkedIn content and B2B pipeline data.

Frequency still pays, with a floor. Buffer's analysis of more than 2 million posts found accounts posting 2 to 5 times a week got 1,182 more impressions per post than once-a-week accounts, 6 to 10 times got 5,001 more, and 11 or more got 16,946 more. LinkedIn and van der Blom both point to two to three posts a week as sufficient for the average creator. Buffer's data is about impressions per post, not leads. The detail is in how often founders should post.

#5. What the data disagrees on

Studies disagree most on external links (from 36% lower to 4.90% higher reach), on how much reach has fallen, on dwell-time payoffs, and on whether 360Brew runs the feed. Most gaps trace to account type, medians versus means, sample year, or a vendor selling the fix.

Disagreement 1: do links cost reach?

5. What the data disagrees on
Study Sample Result
Ordinal, 2026 Not stated on page (a secondary summary says 900,000+ posts, Feb 2023 to Feb 2026) 26.5% lower reach with a link
LinkPost, 2026 358,000 posts Median 448 vs 705 impressions, a 36% gap
Van der Blom, Algorithm Insights Not published Single link penalty of 18.8% in April 2025 per Ordinal; Ordinal says 2026 is lower, while a separate audit attributes 18.8% to the 2026 report
Metricool via IMH, May 2025 577,000+ posts Linked posts +4.90% views, +13.57% interactions
Mercer-MacKay summary of van der Blom 2025 1.8M+ posts About 5% reach gain for linked posts, reversing a decline
Saywhat, Q1 2026 397,605 posts claimed Posts with several links outperform posts with none, per a secondary reading; no data on page

Five explanations reconcile most of the spread.

  1. Account type. Ordinal says the penalty falls on company pages and barely touches personal profiles. LinkPost mixed account types, so its 36% cannot be assigned to either. A founder posting from a personal profile is closer to the group where the penalty is smallest.
  2. Median versus mean. Ordinal and LinkPost both used medians. Metricool's headline uses averages across pages, and averages can be lifted by a few high performers. The two measure different things and can legitimately point in opposite directions.
  3. Sample year. Van der Blom's single-link figure is itself reported two ways: Ordinal says it fell from 18.8% in April 2025 to a lower number in 2026, while a separate audit attributes 18.8% to the 2026 report. Either way, a penalty that shrinks or shifts means a study from early 2025 and one from 2026 are not measuring the same system, especially after the March 2026 feed rebuild.
  4. Confounding. Posts with links are often promotional or news-driven, which can lower or raise engagement for reasons unrelated to the link. None of the studies randomizes who adds a link.
  5. Commercial interest. Ordinal sells tooling for scheduling and moving links into comments. That does not make its data wrong, but its page does not state a sample size, and I did not weight it above studies that do.

LinkedIn has not confirmed a link penalty. My reading: a link in the body probably costs some reach on company pages and a smaller, uncertain amount on personal profiles, and no study is strong enough to put a number on a founder's own account. Test it on your own posts rather than adopting a rule.

Disagreement 2: how far has reach fallen?

Van der Blom's two datasets give "nearly 50%" (2025 report) and "about 60% over two years" (June 2026). They are not the same measure on the same window, and 1.8M versus 1.3M posts mean the panels differ. Metricool's engagement-up result covers all public accounts in its dataset, not only active creators. The consistent reading is that creators who post a lot saw each post reach fewer people, while total activity on the platform did not shrink. If the March 2026 rebuild moves reach toward recent, topic-matched content, per-post reach for generalist accounts would fall and for focused accounts hold. That is an inference; no published study isolates it.

Disagreement 3: dwell time

A widely repeated statistic says posts held for 61 or more seconds reach 15.6% engagement versus 1.2% for posts skimmed in under 3 seconds. I could not find it in any LinkedIn publication, and a secondary audit of 2026 algorithm claims reached the same conclusion: such engagement-by-dwell-bucket figures appear nowhere in LinkedIn's output. What LinkedIn does publish is that long dwell is a predicted action and that dwell-time buckets improved the prediction of it by about 2.5 points in the paper's ablation, as summarized in that audit. Both can be true: dwell is a real signal, and the viral table is a number nobody can trace.

Disagreement 4: is 360Brew your feed?

Many 2026 explainers say LinkedIn's 150-billion-parameter 360Brew model decides what you see. The January 2025 paper describes 360Brew as a decoder-only foundation model for personalized ranking and recommendation. The February 2026 Feed SR paper, which documents the production ranker, says an LLM-style text ranker never beat the production model online, and the production ranker is a sequential transformer using two tokens per post. The Trust Insights Q1 2026 guide, which cites both papers, reaches the same reading. My conclusion: LLMs are used in 2026 retrieval, and the ranker the paper describes is smaller and purpose-built. "360Brew is the algorithm" is not supported by LinkedIn's own documents.

Disagreement 5: more posts, or fewer?

Buffer says more posts per week produced more impressions per post, up to 11 or more a week. LinkedIn suggests two to three. Van der Blom says reducing frequency lifted reach about 10 to 15% for a large share of members, though he stated this loosely. The reconciling point is the unit. Buffer measured impressions per post for accounts that post often, a population that is likely to be committed creators with an audience. Van der Blom speaks to the average creator. For a founder with limited time, the conservative reading is that two to three well-built posts a week is enough to stay in the system and more only pays if quality holds.

Disagreement 6: are comments still the lever?

LinkedIn is reducing comment-bait. Metricool shows comments down 17% year on year while overall engagement rose. Buffer shows authors who reply to comments get about 30% higher engagement. Van der Blom said about 80% of the early comments on his own posts are AI-written. Taken together: raw comment count has become a noisier signal, and replying to real comments is still associated with better outcomes. Both are consistent with a platform trying to separate real conversation from manufactured conversation.

#6. Shruti's analysis: how I would read this as a founder

This section is my synthesis. It uses only the figures above, and anything calculated is labelled as arithmetic.

Three tiers of evidence

I sort every claim about the algorithm into one of three tiers:

  1. Confirmed by LinkedIn: retrieval by meaning, a ranker that predicts six actions from up to 1,000 recent interactions, less distribution for engagement bait and generic AI-looking content, and action against pods and automation.
  2. Measured, not confirmed: link effects, the size of the reach decline, the value of replying to comments, the payoff from frequency.
  3. Repeated, not traceable: dwell-time engagement tables, comment-to-like weight ratios, "Depth Score", and a single "external link penalty" percentage with no named study.

A founder should build a posting system on tier 1, adjust it with tier 2 evidence from their own account, and ignore tier 3.

What the arithmetic implies

Arithmetic, not a study result. If an active creator's reach per post fell 60%, the same absolute reach needs about 2.5 times the posts, or posts that perform 2.5 times better (1 divided by 0.4). That is why I do not recommend chasing the old reach with volume. If a link really cost 26.5% of reach, a linked post would need to be about 36% stronger to break even (1 divided by 0.735). Few founders have a link-worthy post that is a third stronger by default, which is a reason to link when the link is the point, such as a hiring page or a report, and not by habit.

Where the 2026 changes help founders

The shift toward topical matching and substance favours a founder with a specific, repeated point of view on a narrow subject. Retrieval by embedding is indifferent to your follower count, so a post on a niche problem can reach the right buyers outside your network. That is consistent with the personal-profile advantage seen in Metricool and Ordinal. The founder-led case is made with numbers in founder-led marketing statistics for 2026.

Where they hurt

Three founder habits are exposed:

  • Generic AI-assisted drafts. LinkedIn says they travel less. The trust research I compiled elsewhere finds disclosure lowers evaluation of writing by 6.2% on average. The guardrails are in the AI-written posts piece.
  • Comment farming. Pods, automation and "comment Yes" prompts are named targets.
  • Topic drift. If your last 20 posts cover eight subjects, the model has a weak signal about who you are for.

A weekly operating rule set

These rules follow from the evidence above. They are my judgment, not LinkedIn requirements.

  1. Pick one main topic and two or three subtopics. Hold roughly four of every five posts to them (the 80% figure is van der Blom's recommendation).
  2. Post two to three times a week and move above that only if quality holds. This is where LinkedIn's guidance, van der Blom and the lower end of Buffer's data overlap.
  3. Put the specific thing only you know in each post: a number from your own business, a decision, a mistake. That is the substance test LinkedIn describes in its May 2026 note.
  4. Link when the link is the point, and otherwise keep the post self-contained. Check your own impressions on linked and unlinked posts over a month, since the studies do not agree.
  5. Reply to real comments quickly, especially in the first hours, and skip automation. Replying is the one comment behaviour with consistent positive evidence (Buffer) and no platform opposition.
  6. Edit AI drafts for specificity and delete stock phrasing before publishing. Use the free AI slop checker as a screen, not as proof.
  7. Review the profile as a retrieval document. Headline, about section and featured items all feed the same topic signal as your posts.

If you outsource drafts, the questions to ask a writer are in the ghostwriting guide for founders.

What I would not conclude

I would not conclude that LinkedIn punishes links, that frequency beyond three posts a week hurts, or that any particular dwell-time target exists. Each of those is a claim at least one published source contradicts.

#7. What to watch next

  • Revisions to the Feed SR paper. It has been revised twice since February 2026 (29 May and 16 September 2026). Later versions may change the reported numbers.
  • Engagement-bait enforcement. LinkedIn said the reduction would roll out over the following months. A shift in comment counts at a platform level would show up in Metricool-type panels first.
  • A van der Blom report on the post-March period. The 2026 report analyzed posts in a window that straddles the rebuild.
  • Any LinkedIn statement on links. None exists today, and one would settle Disagreement 1.
  • Independent studies with disclosed samples. The weakest part of this field is that most data comes from vendors.

#Methodology

Scope and dates. I compiled sources published between January 2025 and October 2026, with emphasis on the period after LinkedIn's 12 March 2026 engineering post. I included 2025 studies where they are the only published measurement of a topic (links, frequency) and flagged their age.

Source selection. Priority order: LinkedIn's own papers, engineering posts and news releases; then studies that state a sample size and method; then practitioner datasets (van der Blom) discussed by named outlets; then vendor studies with commercial interests. I excluded explainers that supply statistics with no named origin. I did not use patent-based speculation.

Reading primary versus secondary. I opened the arXiv abstract page for the Feed SR paper and its full text, the Search Engine Land, Social Media Today and PPC Land coverage, and the Ordinal and Creator Science pages. I could not open LinkedIn's engineering blog or news pages directly during this work, so the Danchev post is cited through named outlets, and the Lorenzetti note through this site's earlier compilation of LinkedIn's statements. Where a figure comes from a secondary summary (for example van der Blom 2025 via Mercer-MacKay, or the dwell-time audit), the text says so. I did not open the van der Blom reports themselves, which are sold products.

Handling conflicts. Where studies disagreed I recorded each result with its sample, year and type of statistic (median or average), and looked for the cause before judging. I do not average conflicting results, because their populations differ.

Limits. Sample sizes are as stated by publishers and I could not verify them. Several outcome studies come from companies that sell related tools. No source here measures founders specifically. The arithmetic in section 6 is mine and is labelled. Figures were checked on 30 October 2026 and will date quickly given LinkedIn's rollout schedule.

#Sources

By Shruti Kaushik

Quick questions

What changed in the LinkedIn algorithm in 2026?

LinkedIn's engineering team described a rebuilt feed in March 2026: one LLM-embedding retrieval model in place of several separate candidate sources, and a transformer ranker that reads up to 1,000 of a member's past interactions. LinkedIn also said it would show less engagement bait and recycled thought leadership, and in May 2026 said generic, AI-looking posts are less likely to travel beyond a person's network.

Does LinkedIn penalize posts with links in 2026?

LinkedIn has not confirmed a link penalty. Independent studies disagree. Ordinal reported 26.5% lower reach for linked posts and LinkPost a 36% gap in median impressions, while Metricool's 2025 study of 577,000 posts found linked posts got 4.90% more views. Account type, medians versus means and the year of the sample explain much of the spread.

How much does dwell time matter on LinkedIn?

LinkedIn's Feed SR paper lists long dwell as one of six predicted actions, alongside click, skip, reaction, comment and share. It defines long dwell as staying on a post longer than a threshold that depends on the post type and publishes no threshold in seconds. Widely repeated engagement rates by dwell bucket do not appear in LinkedIn's publications.

Is LinkedIn's feed run by the 360Brew model?

Not according to LinkedIn's own February 2026 paper. The 150-billion-parameter 360Brew model was described in January 2025 as a research foundation model, and the Feed SR paper says an LLM-based text ranker never beat the production model online. The production feed ranker described in 2026 is a smaller sequential transformer.

How often should a founder post on LinkedIn after the 2026 changes?

LinkedIn's July 2026 posting tips suggested two to three posts a week, and Richard van der Blom said two to three is enough for the average creator. Buffer's analysis of more than 2 million posts found higher frequency still produced more impressions per post, so the evidence supports a floor of two to three, not a ceiling.

About the author

runs Wise Mango, a LinkedIn ghostwriting studio for SaaS and AI founders, and Klype (klype.ai), an AI platform that writes in a founder’s voice. She wrote The LinkedIn Blackbook and was named one of Forbes Game-Changing Leaders in 2025. Every figure on this page links to its source.