LinkedIn Organic Reach in 2025-2026

What changed, what did not, and the evidence-based founder playbook

Research date: August 3, 2026
Scope: US/English-language B2B relevance; global studies are labeled as such
Evidence standard: LinkedIn engineering/policy disclosures first; transparent multi-account datasets second; vendor studies with missing methods third; anonymous cases and uncited blogs are not accepted as proof

Executive summary

LinkedIn has changed its feed materially, but the defensible story is not “organic reach was cut by two-thirds to force companies to buy ads.” The production system now combines network content with semantic, interest-based retrieval outside the member’s network. In 2025–2026 LinkedIn added an LLM-based out-of-network retriever and deployed a sequential ranker that reads more than 1,000 historical interactions. It also began explicitly limiting distribution associated with engagement pods, automated comments, repetitive generic AI content, and engagement bait. Network affinity, professional identity, recency, content semantics, dwell, clicks, skips, likes, comments, and shares all remain inputs. LinkedIn does not publish their weights. (LinkedIn Engineering, March 12, 2026; Feed-SR paper, submitted February 12 and revised May 29, 2026)

The headline numbers in the fast-scan brief do not survive verification as platform facts:

• No transparent study supports a universal 60–66% company-page reach collapse from 2024 to early 2026. Metricool’s global study of 673,658 posts from 63,108 accounts found company-page impressions per post down 10% from January–February 2025 to the same period in 2026, while clicks rose 5% and its calculated engagement rate rose 14%. Its exclusion of zero-interaction posts, use of averages, global sample, and changing account composition make even that result directional. (Metricool, April 14, 2026)
• “Relationship Graph to Interest Graph” is a useful shorthand, not LinkedIn’s disclosed architecture. Interest-based, out-of-network retrieval is confirmed; abolition of the relationship graph and the claim that only 31% of the feed is first-degree are not. LinkedIn still calls the network a key signal. (Tim Jurka, August 11, 2025; Gyanda Sachdeva, December 18, 2025)
• The claimed 18.8% external-link penalty is not a universal rule. The exact figure traces to Richard van der Blom’s paid 2026 report, whose relevant comparison method is not publicly auditable. Metricool found personal-profile link posts had 27% fewer impressions but company-page link posts had 51% more, while MagicPost found in-body URLs had 9% more median impressions than no-link posts but attached link-preview cards had 47% fewer. All are observational. (Metricool, April 14, 2026; MagicPost, June 2026)
• Dwell time is important, not proven dominant. LinkedIn rejects one universal dwell cutoff and normalizes long-dwell thresholds daily by content type, creator type, and distribution context. The viral 0–3 seconds = 1.2% engagement and 61+ seconds = 15.6% table has no credible published sample or method. (LinkedIn Engineering, October 1, 2024)
• There is no disclosed 60- or 90-minute “golden hour” gate. LinkedIn refreshes representations within minutes, and freshness and early response can matter, but it has not disclosed a first-degree test batch, a pass/fail window, or a 30-minute author-reply multiplier. Buffer found a 30% association between author replies and engagement in a cross-platform, within-account analysis; it explicitly warns of reverse causality. (Buffer, March 5, 2026)
• Artificial engagement can cause reach limits, but the 8,500 → 340 impressions and 60–90 day recovery story is unsupported. LinkedIn confirms reach limitation, comment de-ranking, warnings, program removal, and account restrictions. It publishes no standard impression loss or organic-reach recovery duration. (LinkedIn, March 13, 2026)
• People generally create more conversation than pages, but 4–5× reach is not a general benchmark. Metricool found nearly equal average impressions per post for pages and profiles, with profiles earning 63% higher engagement and 238% more comments; AuthoredUp found profiles at 1.72× page median impressions in more than 3 million posts. Small employee-advocacy studies can produce much larger ratios, but they do not establish a platform-wide rule. (Metricool, April 14, 2026; AuthoredUp, 2026)

The practical conclusion: founders should treat LinkedIn as a relevance-and-trust market, not a reach hack. Use the founder profile as the main conversation surface, the company page as durable proof and paid/retargeting infrastructure, and posts as the warm-up layer for targeted conversations. Optimize for the right readers, qualified replies, profile visits, DMs, and pipeline—not raw impressions or a mythical first-hour score.

Claim-by-claim verification

Provenance flags

• The 31% first-degree, 18.8% link loss, and several first-30-minute reply claims resolve to Richard van der Blom’s paid 2026 Algorithm Insights report, promoted as analyzing 1.3 million posts and 80,000+ profiles. The public material does not expose the relevant cohort definitions, controls, tables, or uncertainty. These figures may be useful hypotheses; they are not independently verifiable platform facts. (Report launch, 2026)
• The exact 1.2% versus 15.6% dwell table first appears in an uncited July 21, 2025 marketing article with no named dataset, sample, dates, or collection method. That page also says direct dwell data is unavailable and later implies it appears in native analytics. Flag it as single-source and internally inconsistent. (Blue Gift Digital, July 21, 2025)
• The 8,500→340 / 60–90 days story is source laundering, not a documented recovery case. Repeating pages do not identify the creator or show native analytics, and one citation chain points to a page dated after the citing article and unavailable as of this research date. Do not publish the numbers.

What LinkedIn actually says the feed does

Retrieval: network plus interest, not one graph replacing the other

LinkedIn’s feed first retrieves candidate posts and then ranks them. The 2025 LLM-retrieval paper describes an out-of-network system that selects 2,000 candidates from hundreds of millions of posts in milliseconds. Training used 5 million member–item pairs. In LinkedIn’s online A/B test, replacing the prior suggested-content retriever while holding the final ranker constant produced relative lifts of 0.8% in revenue and 0.2% in daily unique professional interactors overall. Among lower-liquidity members, it reported +0.23% daily active users, +1.17% professional interactions, and +3.29% revenue. These are product-level lifts against a prior retrieval system, not creator reach multipliers. (LinkedIn-authored retrieval paper, October 16, 2025)

The March 2026 engineering disclosure describes a hybrid inventory of network posts, followed sources, and suggestions from the broader Economic Graph. Semantic retrieval uses post text and format; author headline, company, and industry; article metadata; engagement; recency; affinity; and the member’s profile, skills, work history, education, and prior activity. New-post embeddings arrive near real time, refresh as activity changes, and the stated candidate-retrieval latency is under 50 milliseconds. (LinkedIn Engineering, March 12, 2026)

Interpretation: more posts can now earn out-of-network distribution when their professional meaning matches a member’s evolving interests. That is a real interest-graph shift. It does not mean first-degree distribution disappeared, that 69% of every feed is strangers, or that follower quality no longer matters.

Ranking: sequential and multi-objective

The 2026 production ranker treats activity as an ordered sequence and processes more than 1,000 historical interactions. It includes long dwell, clicks, skips, likes, comments, shares, content representations, profile embeddings, device context, and affinity/count features. Passive actions and active actions route through specialized tasks. The revised Feed-SR paper says it served most feed traffic across more than 1.2 billion members for over three months and produced +2.10% time spent and +3.52% likes, comments, or reshares against the prior DCNv2 ranker. (Feed-SR paper, submitted February 12 and revised May 29, 2026)

This is why simple public weight tables are implausible. A comment is not universally “two likes”: the utility of any action depends on the member, post, author, session, downstream outcomes, and model calibration.

Dwell time: important, contextual, and private

LinkedIn defined on-feed dwell in 2020 as time when at least half of a post is visible, plus separate after-click dwell. It introduced a private skip threshold and used predicted skipping as a negative signal. The threshold was not published. (LinkedIn Engineering, May 12, 2020)

In 2024 LinkedIn added a normalized long-dwell classifier. A single universal threshold was rejected because it would favor inherently longer formats such as video. Instead, daily percentiles are calculated for combinations that include content type, creator type, and distribution method; the resulting long-dwell score enters ranking at a private adjustable weight alongside other actions. A/B tests improved sessions and time measures, but LinkedIn did not publish n or effect sizes. (LinkedIn Engineering, October 1, 2024)

Publishable wording: “Dwell is an important ranking signal, normalized by context.”
Do not publish: “61 seconds unlocks distribution” or “dwell is the dominant signal.”

The “golden hour”: a heuristic, not a disclosed gate

What is confirmed:

• New post and member representations update within minutes, while recency, activity, and affinity contribute to retrieval and ranking. (LinkedIn Engineering, March 12, 2026)
• LinkedIn deliberately relaxed recency in a mid-June 2025 test so older relevant posts could resurface. That weakens the claim that a post’s fate is fixed in its first hour. (Gyanda Sachdeva on record, reported July 15, 2025)
• Metricool estimates that 50% of lifetime impressions arrive within 48 hours, not one hour, in its 673,658-post January–February 2025/2026 sample. (Metricool, April 14, 2026)

What is not disclosed: a 60- versus 90-minute window; the size or connection degree of an initial test audience; a pass/fail expansion score; a percentage of lifetime reach decided in hour one; or a 30-minute author-reply bonus.

The practitioner advice to be available after posting is still sensible because authentic replies sustain conversation. Buffer’s March 2026 analysis compared replying and non-replying posts within the same accounts and found 30% higher LinkedIn engagement when authors replied; 83% of LinkedIn profiles in that analysis performed better with replies. The reply study covered nearly 2 million posts and 220,000+ accounts across six networks, did not disclose LinkedIn-only n, and was observational. Buffer explicitly notes that successful posts may simply attract more replies from their authors. (Buffer, March 5, 2026)

Operational rule: reply because there is a useful conversation, preferably while it is live; do not build a team SLA around a fictitious 30-, 60-, or 90-minute cliff.

Enforcement, pods, automation, and “AI slop”

LinkedIn’s March 2026 policy disclosure confirms it looks for coordinated groups that manufacture likes, comments, and shares; browser extensions, scripts, bots, and third-party tools that post engagement; suspicious patterns; and artificially boosted posts. It may limit feed reach, keep automated comments out of “Most Relevant,” prevent them spreading outside the commenter’s network, remove engagement-pod groups, warn members, remove program badges, or restrict accounts. (LinkedIn, March 13, 2026; automated activity Help Center, current in 2026)

LinkedIn does not disclose timing regularity, graph density, text-similarity thresholds, rate limits, classifier precision, or exact penalty duration. Therefore claims such as “five comments inside three minutes triggers suppression,” “97% pod detection,” or “stop posting for seven days to reset the account” are unsupported.

In June 2026 LinkedIn also said it was reducing distribution of repetitive, generic, low-substance AI content and generic comments, while allowing AI-assisted content that contains the member’s own perspective. LinkedIn says its initial classifier was correct at identifying generic content 94% of the time, but supplied no sample size, class balance, precision/recall, or external audit. Publish this only as LinkedIn’s own initial-test claim, not “94% AI detection accuracy.” (LinkedIn, June 4, 2026)

What a credible recovery case would require

No public 2025–2026 case located in this review contains all four of the following:

1. Dated before-and-after native analytics across multiple comparable posts.
2. A confirmed enforcement cause or LinkedIn notice.
3. A dated intervention and recovery point.
4. A comparison that holds format, topic, cadence, audience, and broader feed changes reasonably constant.

An account-access restriction that expires is not proof of organic-distribution recovery. Nor is one post bouncing back after a week. LinkedIn says impressions and members reached are estimates, may include the author’s own views, and differ because impressions include repeat displays. (LinkedIn Help, updated July 2026)

Recovery protocol founders can defend

• Stop pods, automated comments, scraping, mass messaging, and unauthorized extensions immediately.
• Preserve native analytics; do not delete evidence in a panic.
• Record medians across at least 10 comparable posts before diagnosing an account-level break. Segment profile versus page, format, topic, in-network/out-of-network reach where available, and 48-hour versus seven-day performance.
• Check for explicit account, feature, or content notices. Appeal explicit restrictions; do not claim a shadow-ban cause that LinkedIn has not confirmed.
• Resume a normal human cadence with original expertise and natural comments. Do not promise a 7-, 30-, 60-, or 90-day reset.

Format performance in 2026

There is no format that wins for every account and objective. Current studies disagree because they sample different populations and use different denominators. Use these as priors, then test on your own audience.

Socialinsider’s 2026 benchmark covers 1.3 million posts from 16,645 active business pages. Its website says January 2024–December 2025 while the PDF method states January–December 2025 even though charts span two years; that internal inconsistency and undisclosed geography/outlier policy lower confidence. (Socialinsider, 2026)

Founder-led distribution playbook

1. Assign each surface a job

Founder profile: point of view, lived evidence, replies, relationship building, and demand capture. Current data supports higher conversation and modestly higher reach—not a guaranteed 5× multiplier.

Company page: product proof, customer evidence, jobs, governance, employee association, newsletters, paid amplification, and retargeting. Pages are not “dead”; they are a less personal surface with a different job.

Team profiles: expertise from practitioners who genuinely own the insight. Do not make employees publish identical copy or coordinate timed reactions; that resembles the behavior LinkedIn says it is suppressing.

2. Publish around a stable professional promise

Choose three durable lanes:

1. Market truth: a counterintuitive observation backed by customer data or firsthand evidence.
2. Operating proof: decisions, experiments, failures, metrics, and how the company works.
3. Buyer utility: frameworks, teardowns, and decision aids the ideal customer can apply.

This is not an attempt to game a rumored “topic-authority score.” LinkedIn has not disclosed such a score. It is a practical response to confirmed semantic matching of content, professional identity, and member interests. (Tim Jurka, August 11, 2025)

3. Use a sustainable weekly portfolio

A founder can start with three substantial posts per week: one text or image insight, one document/multi-image framework, and one case, customer lesson, or native video when the medium adds value. The exact number is a workflow recommendation, not an algorithm threshold.

There is no credible evidence that posting more automatically throttles reach. Buffer analyzed more than 2 million posts from 94,000+ LinkedIn accounts and used within-account normalization plus fixed-effects regression. Relative to one post per week, 2–5 posts were associated with +1,182 impressions per post and +0.23 percentage points of engagement; 6–10 with +5,001 impressions and +0.76 points; 11+ with +16,946 and +1.4 points. These are observational associations in Buffer’s user base, not a command to maximize volume. (Buffer, August 28, 2025)

4. Treat links as a conversion choice

Do not automatically hide every link in the first comment. Write a post that stands alone, then:

• Put the link in the body when the reader needs the source, tool, event, or next step.
• Test a clean body URL separately from an attached preview card because current data suggests they behave differently.
• Use UTMs and compare qualified clicks and conversions, not impressions alone.
• Analyze profiles and company pages separately; Metricool found opposite link correlations by author type in April 2026.

5. Engineer conversation, not engagement bait

• Ask a question only when the answer changes your view or helps the audience.
• Reply with new information, a useful follow-up, or a real question.
• Do not coordinate early reactions, use AI auto-comments, or ask for empty “agree?” comments.
• Be present after publishing when practical, but measure at 48 hours and seven days rather than declaring a post dead at minute 60.

Metricool found questions associated with 77.39% more comments and comment CTAs with 80.07% more in its 673,658-post global sample, but this is observational and LinkedIn now explicitly targets engagement bait. The safe lesson is “invite substantive participation,” not “put ‘comment YES’ on every post.” (Metricool, April 14, 2026; LinkedIn, March 12, 2026)

6. Measure a 90-day founder funnel

Maintain separate profile and page dashboards. For every post capture:

• format, lane, link rendering, publish time, and 48-hour/seven-day impressions;
• members reached where available, profile views, follower quality, comments from the ideal customer, saves/sends if available, DMs, and connection requests;
• UTM sessions, conversions, meetings, sourced pipeline, and assisted pipeline.

Use median, not only average, across at least 10 comparable posts. Impressions are displays, not unique people; LinkedIn warns both impressions and members reached are estimates and can include self-views or repeat display. (LinkedIn Help, updated July 2026)

7. Turn reach into founder-led outbound

Use the feed to create recognition, then follow up selectively:

1. Publish a strong point of view for the target market.
2. Leave thoughtful comments where buyers already participate.
3. Connect with relevant people who engaged or share the problem context.
4. Send a short contextual DM; use InMail when access or urgency justifies the cost.
5. Add email for proof, scheduling, and multi-threading rather than treating channels as substitutes.

This is the economic advantage: not “free reach,” but lower-friction context before the ask.

Channel economics: cold email versus LinkedIn messages

There is no clean 2026 apples-to-apples US B2B sales benchmark. Definitions and funnel stages differ too much.

Verdict on “the last cheap channel”: overstated. LinkedIn messages are higher-context and often higher-response, but they are relationship-gated, capacity-constrained, and sometimes paid. Cold email scales further but produces lower and far more variable replies. The supported strategy is founder-led, low-volume LinkedIn plus targeted email—not abandoning one for the other.

As of July 1, 2026, LinkedIn listed Sales Navigator Core from $119.99 monthly or $1,079.88 annually; Sales Navigator includes 50 InMail credits per month with a maximum bank of 150. Those limits make “cheap” dependent on deal size and founder time. (LinkedIn pricing, July 1, 2026; LinkedIn Help, accessed August 3, 2026)

Primary versus inferred: what can go behind your name

LinkedIn-disclosed or directly measured in LinkedIn papers

• LinkedIn combines network, followed-source, and out-of-network suggested content. (March 12, 2026)
• Out-of-network retrieval uses semantic and professional-interest signals; a 2025 production A/B test reported small positive platform lifts. (October 16, 2025)
• The sequential ranker reads more than 1,000 historical interactions and separately models passive and active behavior. (February/May 2026)
• Dwell thresholds are dynamically normalized rather than a fixed public number. (October 1, 2024)
• Pods, automated engagement, engagement bait, and high-volume inauthentic activity can cause reduced distribution or account action. (March 12–13, 2026)
• Generic AI content may be less likely to spread beyond the immediate network; AI assistance itself is not forbidden. (June 4, 2026)
• LinkedIn says ordinary organic distribution is not influenced by third-party payments to LinkedIn except clearly labeled promoted content. (LinkedIn Help, current in 2026)
• 360Brew is not the 2026 production feed algorithm; its small early-2025 test was stopped. (Tim Jurka, April 2026; research paper, January 27, 2025)

Third-party observations worth publishing with method attached

• Metricool’s April 2026 global sample found company-page impressions per post down 10% year over year, not 60–66%, while profiles generated more comments and engagement. State the 673,658-post/63,108-account sample, January–February comparison, exclusions, and global scope. (Metricool)
• Buffer found a 30% association between author replies and engagement; say “associated,” not “caused,” and note the study covered nearly 2 million posts/220,000+ accounts across six networks with no LinkedIn-only n. (Buffer, March 5, 2026)
• AuthoredUp’s profile sample found documents at 1.39× account-median reach and profiles at 1.72× page median impressions; identify it as a commercial-tool dataset of 3M+ posts with account count and geography undisclosed. (AuthoredUp, 2026)

Inference or myth: do not present as LinkedIn-confirmed

• “The relationship graph was replaced,” or first-degree posts are exactly 31% of feeds.
• A universal 18.8% link penalty.
• A universal three-second skip or 61-second long-dwell threshold.
• Dwell as the single dominant signal.
• Comments worth exactly 2× likes; saves or DMs worth a fixed multiple.
• A 60/90-minute golden-hour audition or first-degree test batch.
• Replies within 30 minutes cause 35% more visibility.
• A standard 60–90-day silent-penalty recovery.
• A public “topic-authority score.”
• 4–5× profile reach as a universal benchmark.
• A deliberate organic-paywall mechanism designed to force ad purchases.

Dated 2025–2026 timeline

Steelman: why “deliberate throttling to sell ads” may be wrong

The strongest contrary case has four parts:

1. LinkedIn explicitly denies payment influence on ordinary feed distribution. Its relevance policy says third-party payments do not affect distribution except promoted content that is clearly labeled. This is a first-party claim, not an independent audit, but it directly contradicts the asserted mechanism. (LinkedIn Help, current in 2026)
2. Content supply rose. LinkedIn said content sharing increased 15% and comments 24% during 2025. More posts competing for finite attention can reduce median distribution even when total activity grows. (Gyanda Sachdeva, December 18, 2025)
3. Per-post reach and value can move in opposite directions. Metricool found company-page impressions down 10%, yet clicks up 5% and calculated engagement up 14% from Jan–Feb 2025 to Jan–Feb 2026. That looks more like allocation/competition and metric mix than a simple kill switch. (Metricool, April 14, 2026)
4. Measurement is unstable. Impressions are estimated displays, not people. Studies alternate between averages and medians, profiles and pages, follower-normalized and impression-normalized rates, link cards and body URLs, fixed cohorts and changing vendor users. Metricool excluded zero-impression or zero-interaction posts and retained outliers; Socialinsider has an internally inconsistent study window. Apparent “collapse” can be partly compositional.

The best synthesis is therefore:

LinkedIn is reallocating distribution toward predicted professional relevance and authentic human interaction while explicitly limiting suspected artificial engagement. Some cohorts—especially small company pages—show lower per-post impressions, but public evidence does not establish a platform-wide 60–66% collapse engineered to force ad spend.

The 12 most authoritative sources

1. LinkedIn Engineering: next-generation Feed — March 12, 2026; best official production architecture overview.
2. Feed-SR paper — February/May 2026; LinkedIn-authored production ranker with disclosed online A/B lifts.
3. LLM retrieval paper — October 16, 2025; LinkedIn-authored out-of-network retrieval design, training scale, candidate scale, and A/B results.
4. LinkedIn authenticity enforcement — March 13, 2026; primary policy source for pods, automated comments, reach limits, and restrictions.
5. LinkedIn dwell-time engineering — October 1, 2024; primary explanation of contextual, daily-normalized dwell thresholds.
6. Tim Jurka: how the Feed works — August 11, 2025; on-record product explanation of semantic relevance, professional context, network, and out-of-network reach.
7. Metricool LinkedIn Study 2026 — April 14, 2026; largest transparent mixed profile/page study reviewed, with 673,658 posts and explicit limitations.
8. Socialinsider LinkedIn Benchmarks 2026 — 2026; 1.3 million company-page posts and useful format/follower-tier comparisons, with method caveats.
9. Buffer State of Social Media Engagement 2026 — March 5, 2026; large cross-platform dataset and within-account reply analysis with reverse-causality caveat.
10. AuthoredUp format analysis — updated June 25, 2026; 3M+ personal-profile posts normalized to each profile’s median.
11. Belkins cold-email benchmark — June 26, 2026; 7.53 million 2025 emails, transparent sends denominator, and explicit warning about historical denominator changes.
12. Pin recruiting outreach benchmark — July 8, 2026; 4M+ messages and the most current cross-channel reply comparison, clearly limited to recruiting.

Open questions LinkedIn does not disclose

• Feed inventory share by first-degree, follower, page, second-degree, and unconnected source.
• Whether “31% first-degree” is true for any population, geography, or session type.
• Network-retrieval candidate counts and how they combine with the disclosed 2,000 out-of-network candidates.
• Signal weights, utility coefficients, and whether a comment’s marginal value changes by author or commenter affinity.
• The size and composition of any early audience cohort and whether a fixed evaluation window exists.
• Any direct causal rank penalty for a body URL, preview card, multiple links, or destination type.
• Pod/automation timing, graph-density, similarity, volume, and device thresholds.
• Duration, decay, and recovery criteria for post-level or account-level distribution limits.
• Whether and how creator reputation, subject consistency, and prior negative feedback persist across posts.
• AI-generic-content classifier sample, class balance, false-positive rate, language coverage, and appeal path.
• Rollout coverage by country, language, account type, member tenure, and activity level.
• Independent evidence separating algorithm changes from growing content supply and analytics-definition changes.

Research notes and limitations

This report prioritizes 2025–2026 material but uses LinkedIn’s 2020 and 2024 dwell disclosures because they remain the platform’s most precise public explanation of that signal. It reviewed primary engineering posts, papers, policy/help pages, staff statements, and multiple 2025–2026 vendor datasets. The largest studies are global and vendor-selected, not probability samples of US B2B accounts. None provides a randomized creator-side experiment for links, formats, replies, or founder-versus-page distribution. Consequently, third-party results are associations and starting priors, not algorithm weights.

The absence of a public source is not proof that an internal mechanism does not exist. It is proof that the number cannot responsibly be presented as verified.