ktg-plugin-marketplace/plugins/linkedin-studio/docs/remediation/research/01-linkedin-algorithm-signals.md

type	created	question	confidence	dimensions	mcp_servers_used	local_agents_used	external_agents_used
trekresearch-brief	2026-05-29	What does the 2026 LinkedIn feed-ranking system actually reward — comment-vs-reaction weighting, document/carousel engagement rate, external-link reach effect and first-comment status, the early-engagement window incl. delayed reinjection, and the deployed ranking model's verifiable name and date — with a source and confidence per claim?	0.82	8	tavily gemini-deep-research		docs-researcher community-researcher security-researcher contrarian-researcher gemini-bridge

2026 LinkedIn Feed-Ranking — Canonical Signal Statement

Generated by trekresearch (high-effort swarm: 4 external + Gemini) on 2026-05-29. Topic 1 of 3 for the linkedin-studio remediation. This is the substrate: the Phase-0 fixes that reconcile the plugin's contradictory algorithm stats consume it.

Research Question

What does the 2026 LinkedIn feed-ranking system actually reward — comment-vs-reaction weighting, document/carousel engagement rate, external-link reach effect and the current first-comment-workaround status, the early-engagement ("golden hour") window incl. delayed/evergreen reinjection, and the deployed ranking model's verifiable name and deployment date — with a primary or credible source and a confidence level per claim?

Executive Summary

The plugin's algorithm "facts" are directionally right but numerically indefensible: every specific magnitude it states (comment "15x", carousel "6.6%"/"1.92%", link "40-50%"/"25-40%", a clean "−40-60% before distribution", "360Brew, January 2026") is either third-party-only, self-contradictory, conflated across denominators, or — for the model name/date — not establishable from any primary source. What IS defensible and high-confidence: an LLM-based relevance-ranking system is live in 2026; the engagement hierarchy is saves > shares > quality comments > reactions with dwell-time a top-tier signal (the only two signals LinkedIn officially confirms by name are dwell time and topic/interest relevance); documents/carousels are the #1 format; body links reduce reach (magnitude contested, ~19–60% across studies, LinkedIn denies it is intentional); the early window is 60–90 min (90 is the 2026 consensus); and — the single best-supported actionable finding — LinkedIn now officially suppresses generic AI "slop" (named executive, May 2026), which directly justifies a short-form de-AI gate. Key caveat: treat every number as directional and per-account-testable; encode ordering + sourced direction, never hard coefficients. (Overall confidence 0.82 — high on direction, medium on magnitude.)

Dimensions

D1. Deployed ranking model — name & date — Confidence: high (on the negative claim)

External findings:

• The arXiv paper "360Brew: A Decoder-only Foundation Model…" (2501.16450) is dated 2025-01-27, self-labels as a "research pre-production model" (V1.0, 150B params) claiming offline parity only, and was withdrawn 2025-08-23 (submitter lacked license rights). It is neither a deployment announcement nor a clean citable artifact. [arXiv 2501.16450]
• LinkedIn's own 2026 communications describe a live LLM-based feed system but the production name is not reliably establishable: the docs + contrarian agents both read a LinkedIn Engineering post ("Generative Recommender / GR", attributed to Hristo Danchev, 2026-03-12); the independent Gemini pass flagged a third-party citation of that same post as possibly fabricated (Danchev's verifiable authorship is on AWS OpenSearch work). So even the "GR" name carries a provenance question.
• "January 2026" as a deployment date appears in no primary source; it is third-party extrapolation from the paper's Jan-2025 date.

Contradictions: docs/contrarian treat the GR engineering blog as primary; Gemini casts doubt on its provenance. Conservative resolution: assert neither name nor date. An LLM relevance-ranking system is live (high confidence); its deployed name and go-live date are not publishable as fact.

D2. Comment vs reaction weighting + saves/dwell hierarchy — Confidence: high (ordering) / medium (magnitude)

External findings:

• "Comment = 15x a like" is unverified folklore — no primary source; meet-lea labels it "industry estimate, original source unclear." Sources span 2x–15x with no anchor. AuthoredUp's NLP-quality-scored analysis puts the real comment-vs-like effect ~2x. [authoredup.com/blog/linkedin-algorithm; meet-lea]
• Convergent across AuthoredUp + Vertebrae + van der Blom (1.8M): a save ≈ 5x a like, ≈ 2x a comment — saves are the top signal (and a follow-graph signal: saving a post gives the author's next post ~80% feed-appearance odds). The plugin's stray "5x" is the saves number mis-assigned to comments.
• Officially confirmed (the only two named): dwell time is a ranking signal (LinkedIn Eng "Understanding feed dwell time" 2020; "Leveraging Dwell Time" / Auto-Normalized-Long-Dwell model 2024); LinkedIn describes active (like/comment/share) vs passive (click/skip/long-dwell) tasks but assigns no weights. [linkedin.com/blog/engineering/feed/leveraging-dwell-time-to-improve-member-experiences-on-the-linkedin-feed]

Resolution (for the canonical statement): order is saves > shares > quality comments > reactions/likes, with dwell-time top-tier; comment ≈ 2x like (quality-weighted, single-vendor). Drop "15x" and the comment-"5x" entirely.

D3. Document/carousel engagement rate — Confidence: high (format rank) / medium (number)

External findings:

• Three independent large-N studies agree documents/carousels are #1: Socialinsider (1.3M) native document 7.00% (multi-image 6.80%), Buffer (2M) carousel 21.77% median, Metricool (673K) 49.52%. The 7 vs 21.77 vs 49.52 spread is a denominator/methodology artifact, not disagreement about the winner. [socialinsider.io/social-media-benchmarks/linkedin; buffer.com/resources/data-best-content-format-social-media/; metricool.com/linkedin-trends/]
• The "6.6%" is a stale 2024 multi-image figure (now ~6.45% multi-image / ~7.00% document) — and LinkedIn removed native carousels Dec 2023, so "carousel" = PDF document post; the multi-image↔document conflation is real.
• The plugin's "1.92%" is NOT a carousel rate — it matches the personal-profile per-post baseline (Metricool personal 2.60% / company 1.74%; AuthoredUp 2.10–2.67%). The plugin mixed a format benchmark with a personal-profile baseline.

Resolution: documents/carousels = top format (high confidence). For a number use ~7% (Socialinsider, conservative, company-page per-impression); never present 1.92% as a carousel figure; state the format-vs-account-type distinction.

D4. External-link reach effect + first-comment status — Confidence: medium (effect) / low (intent, first-comment)

External findings:

• A body-link reach reduction is real and observational. The most rigorous source (Ordinal, 900K posts, Mann-Whitney p<0.001) shows it changed over time: 5% (2023) → 35% (2024) → 42% (2025) → ~38% (2026 YTD), 37-month avg 26.5%. van der Blom reports a milder ~18.8% median; DigitalApplied/Gemini cite ~60%. So the plugin's "40-50%" ≈ the 2024-25 peak and "25-40%" ≈ the long-run average — both partial views of one moving number. [tryordinal.com/blog/linkedin-link-penalty-study]
• LinkedIn denies an intentional penalty (Sr. Director Product, reported Aug 2025): no penalty "if the post leads with value"; the effect is engagement-driven, not a flat tax. The observed reach gap is real regardless of intent. [threads.com/@mattnavarra/post/DOWa_61Cown/]
• First-comment workaround is genuinely contested: Ordinal data leans "still net-positive but reduced (~−5 to −10%)"; multiple 2026 blogs claim it's now detected as "bridge behavior" and throttled — but that claim is practitioner-only, no large-N backing. The one officially-confirmed principle: what gets limited is off-platform-funnel intent + thin standalone value, regardless of link location.

Resolution: state it as a correlational reach reduction (~38% in 2026, contested band ~19–60%, LinkedIn disputes intent), not a hard penalty. Reframe first-comment as neither a magic fix nor a confirmed penalty — lead with standalone value; native formats are the durable answer. Drop the precise % from the enforcing hook.

D5. Early-engagement window + evergreen reinjection — Confidence: high (60-90 min) / low (24-72h timing)

External findings:

• 2026 consensus has widened from "strict 60 min" to 60–90 min (90 is van der Blom's current figure), with the first 15–30 min the highest-leverage sub-window and ~70% of reach decided in it. [buffer.com/resources/linkedin-algorithm/; expandi.io/blog/best-time-to-post-on-linkedin/]
• Evergreen resurfacing is real in direction (the 2026 relevance model resurfaces strong-save / high-dwell posts days-to-weeks later on viewer intent; AuthoredUp: posts now live 2–3 weeks vs days) — but no large-N source confirms a specific "24–72h reinjection" rule; it is intent-driven and irregular.

Resolution: "60–90 min golden window; first 15–30 min highest-leverage"; describe evergreen as "can resurface days-to-weeks later on intent-match", not a fixed 24–72h second wave. The plugin both over-indexes the strict first hour AND omits evergreen — fix both.

D6. Profile/topic relevance as a ranking input — Confidence: high (signal) / none (the −40-60% figure)

External findings:

• Officially confirmed (qualitatively): topic/interest relevance drives distribution, including beyond your network — Tim Jurka (Head of Feed AI, 2025-08-11): "Exceptional content may even be distributed broadly … to members interested in the type of content you post, even if they don't follow you." 2026 comms add an Interest Picker + "relevant to your interests, not a popularity contest." [linkedin.com/pulse/how-does-linkedin-feed-work-tim-jurka-oxraf]
• No primary source states any −40-60% reach reduction for off-topic content, nor a discrete "validation-before-distribution gate" with a number. That figure is third-party.

Resolution: keep "profile/topic alignment is a real ranking input" (sourced direction); drop the "−40-60% before anyone sees it" figure entirely.

D7. Buzzword penalty — Confidence: high (that it is NOT a measured ranking mechanic)

External findings:

• No primary source ties specific words to a measured reach penalty. Evidence is either editorial/clarity advice (Inc.) or unmeasured vendor assertion (linkboost "LLMs throttle corporate speak"). A semantic-relevance ranker may indirectly favor specific over generic phrasing — inferred, not confirmed. [inc.com/...buzzwords; linkboost.co/blog]

Resolution: keep buzzword-avoidance as editorial guidance, not a "reduces reach" ranking claim. (The plugin already enforces a buzzword list via a hook — keep the list, fix the justification.)

D8. AI-content down-rank — Confidence: high (officially confirmed) — the build-justifying finding

External findings:

• Officially confirmed, named executive: LinkedIn VP & Executive Editor Laura Lorenzetti (2026-05-19) confirmed an active program targeting (1) generic AI-written posts/comments, (2) automation tools, (3) attention-bait video. Mechanism: ML models trained on thousands of human-annotated posts distinguish "original thinking" from "posts lacking substance"; low-quality-flagged posts are reach-suppressed (reportedly down to first-degree connections), not deleted. [entrepreneur.com/business-news/linkedin-is-fighting-back-against-ai-slop-and-ai-comments]
• Corroborated: Jobanputra (Feed) — "we actively detect and limit the reach of spammy or low-quality content, including bot-generated posts." Originality.ai (8,795 posts): likely-AI posts saw 45% less engagement (correlational). [prdaily.com/...guardians-of-the-feed; originality.ai/blog/ai-content-published-linkedin]
• Also officially confirmed and relevant: engagement-pod crackdown (VP Product Gyanda Sachdeva, 2026-02-16 — auto-comments demoted out of "Most Relevant", scoped to own network, repeat offenders restricted). [socialmediatoday.com/news/linkedin-outlines-more-measures-to-combat-engagement-pods/812290/]

Resolution: build the short-form de-AI / differentiation gate — it targets an officially-confirmed suppression surface. Enforce the signals LinkedIn named (personal substance, original thinking, concrete specifics, genuine voice), not an unverified SEO "tell-list."

External Knowledge

Best Practice (official / primary)

Only two ranking signals are officially named: dwell time and topic/interest relevance. LinkedIn officially denies an intentional link penalty and officially confirms an AI-slop down-rank + engagement-pod enforcement. Everything else (coefficients, multipliers, windows) is third-party.

Alternatives / contrarian

The contrarian pass refuted 6 of 7 plugin claims on magnitude/naming, not direction: the strategic advice (favor native formats, prompt quality comments, write with substance, expect link posts to underperform, post when the audience is active) survives; the specific numbers and the "360Brew, Jan 2026" branding do not. Two need outright correction: the model name/date, and the "no analytics API → CSV only" premise (see D9 in Topic 2 — Member Post Analytics API launched 2025-07-08).

Known issues

Numbers rot: every magnitude is observational and moves year-to-year (link penalty 5%→42%→38%; carousel 6.6%→6.45%). A fabricated citation ("Hristo Danchev / Mar-12-2026") is actively circulating — do not propagate any single named-source deployment claim without first-hand re-verification.

Gemini Second Opinion

Independent ~22-min deep-research pass (27 grounding sources). Agreements with the swarm: 360Brew is a Jan-2025 pre-production paper, not a confirmed 2026 production system; saves/dwell primacy; carousel #1 with methodology-driven rate spread; 90-min window; per-post Saves ARE visible in the native UI for your own posts; a Member Post Analytics API exists but is gated behind Community Management API approval (not self-serve). Unique contribution: independently flagged the "Hristo Danchev / March 2026 engineering post" citation as likely fabricated, which is why this brief refuses to publish any deployed-model name even though two of the swarm agents cited "GR."

Synthesis

Three insights emerge only from triangulation:

1. The plugin's contradictions are mostly denominator/era artifacts, not errors of fact. "40-50% vs 25-40%" = the same link number at peak vs average; "6.6% vs 1.92%" = a format benchmark vs a personal-profile baseline; "15x vs 5x" = a folklore comment figure vs the real saves figure mis-assigned. The fix is therefore one canonical statement that names the era, the denominator, and the account type — not a hunt for "the right number." This is the single most important design instruction for Phase 0.2.

2. Encode ordering + officially-named signals, not coefficients. The only durable, defensible spine is: dwell + topic-relevance are the two officially-named signals; saves > shares > quality-comments > reactions is the engagement order; documents are the top format. Every coefficient must carry a source + confidence + "directional, test per account" caveat. A references/algorithm-signals-reference.md rebuilt around named signals + ordering + per-claim source column makes the contradictions structurally impossible to reintroduce.

3. The two highest-confidence findings each map to a Phase-2 build decision. The officially-confirmed AI-slop down-rank justifies the short-form de-AI gate (D8); the officially-confirmed link-intent principle (value-first, location- secondary) rewrites the link advice (D4). Both are now grounded in named-executive sources, not vendor blogs — the strongest evidence in the whole pass.

Open Questions

• Deployed model name/date — unresolvable from open sources and partly contaminated by a fabricated citation. Carry as: do not assert; state "an LLM relevance model is live in 2026" only. No further research will likely fix this before publication.
• Link-penalty exact magnitude & first-comment status — genuinely contested (~19–60%; first-comment net-positive vs detected). Carry as a range + "test per account"; do not hard-code.
• Member Post Analytics API self-serve depth — answered enough here to act, but is the primary subject of Topic 2 (verify gating + saves-UI before writing boundary prose).

Recommendation

For the Phase-0 "reconcile to one sourced statement" step, adopt this canonical spine and make every command/agent cite it:

1. Model: "An LLM-based relevance-ranking system is live on LinkedIn in 2026." No name, no date. Remove "360Brew" and "January 2026" from CLAUDE.md/README/profile.
2. Signals (officially named): dwell time; topic/interest relevance. Engagement order: saves > shares > quality comments > reactions; likes ≈ 1x baseline. No coefficients without a source column; comment ≈ 2x like is the most defensible single figure (medium).
3. Format: documents/carousels are the top organic format (~7%, Socialinsider, company-page per-impression). Delete the 1.92% carousel claim (it's a personal-profile baseline). Native video #2 and declining.
4. Links: correlational reach reduction (~38% in 2026; contested ~19–60%); LinkedIn denies intentional penalty; value-first matters more than link location; first-comment is a hedge, not a fix. Soften the enforcing hook from a hard % mechanic.
5. Timing: 60–90 min early window (first 15–30 min highest-leverage); add evergreen resurfacing (days-to-weeks, intent-driven); drop the strict-60-min fixation and the "24–72h reinjection" precision.
6. Profile/topic: real ranking input (keep); drop the −40-60% figure.
7. Buzzwords: editorial guidance only (keep the list, fix the "reduces reach" claim).
8. Build the de-AI gate (D8, officially-confirmed surface) and reframe link advice around intent (D4). Both are Phase-2 builds with named-executive backing.

Sources

#	Source	Type	Quality	Used in
1	arXiv 2501.16450 — 360Brew (withdrawn 2025-08-23)	official	high	D1
2	LinkedIn Eng — Engineering the next-gen Feed (provenance contested)	official(?)	low	D1
3	LinkedIn Eng — Leveraging Dwell Time (2024-10-01)	official	high	D2
4	Tim Jurka — How Does the LinkedIn Feed Work? (2025-08-11)	official	high	D6
5	AuthoredUp — LinkedIn Algorithm (621K posts)	community	medium	D2, D3, D5
6	Socialinsider — LinkedIn benchmarks (1.3M)	community	medium	D3
7	Buffer — Best Content Format (2M+)	community	medium	D3
8	Metricool — 2026 LinkedIn study (673K)	community	medium	D3
9	Ordinal — Link Penalty Study (900K, p<0.001)	community	medium-high	D4
10	Threads/Matt Navarra — LinkedIn denies intentional link penalty	official (relayed)	medium	D4
11	Entrepreneur — LinkedIn fights AI slop (Lorenzetti, 2026-05-19)	official (reported)	high	D8
12	PR Daily — Guardians of the Feed (Jobanputra)	official (reported)	medium-high	D4, D8
13	Social Media Today — engagement-pod crackdown (Sachdeva, 2026-02-16)	official (reported)	high	D8
14	Originality.ai — AI content on LinkedIn (45% gap)	community	medium	D8
15	van der Blom — Algorithm Insights 2025 (1.8M)	community	medium	D2, D4, D5
16	meet-lea — LinkedIn Algorithm Explained 2026	community	low-medium	D2
17	Microsoft Learn — Member Post Statistics API	official	high	D2/Topic-2
18	Inc. — buzzwords to scrub	community	low	D7