Kjell Tore Guttormsen
5c6393f2f7
Del 1 (RE-verifisert mot ground truth): README maturity-note (herding 29/29 + kald-review 29/29; gjenstår = GUI), CLAUDE.md maturity-linje (B-F10), hardening-plan-køen t.o.m. S31a/b/c, second-brain-header (SB-S3a-e landet, kun S4 gjenstår). Del 2: D-1 BLOCKER — algorithm-signals GR-seksjonen omskrevet mot primærkilde (LinkedIn engineering-blogg 2026-03-12, Hristo Danchev: Generative Recommender (GR) offisielt navn + LLM-retrieval + utrulling annonsert); fabrikasjonsflagget avviste en ekte primærkilde og er trukket med korreksjonsnote; 360Brew-skepsis beholdt. D-2 — saves-begrunnelse: Marketing API v202604 har POST_SAVE på /memberCreatorPostAnalytics (partner-gated; manuell inntasting forblir riktig UX). B-F11 — README refs-badge 26->28 + 25-document->28-document (ls references/*.md = 28). CLAUDE.md Architecture faar specifics-bank + contract-gate-linjer. CHANGELOG-catchup kommer i release-committen (0.6.0) for aa holde versjonsdeklarasjonene konsistente per commit. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> Claude-Session: df0a1ca3-78dd-455e-99a2-e7c133fcb5f6
202 lines
13 KiB
Markdown
202 lines
13 KiB
Markdown
# LinkedIn Algorithm Signals Reference (2026)
|
||
|
||
**Single source of truth** for what the 2026 LinkedIn feed-ranking system rewards.
|
||
Every other file in this plugin cites this one — do not restate magnitudes elsewhere,
|
||
link here instead.
|
||
|
||
## How to read this file
|
||
|
||
The 2026 feed is ranked by an **LLM-based relevance system** (live in 2026; LinkedIn has
|
||
**no publicly verifiable production name or go-live date** — see the model note below).
|
||
Almost every "coefficient" circulating in the creator community is **third-party,
|
||
observational, and moves year-to-year.** So this reference encodes **ordering + the two
|
||
officially-named signals + directional magnitudes with a source and a confidence per
|
||
claim** — never hard coefficients to optimize against.
|
||
|
||
- **Confidence: high** = officially confirmed by LinkedIn, or convergent across multiple
|
||
large-N studies.
|
||
- **Confidence: medium** = single credible large-N source, or convergent direction with a
|
||
contested magnitude.
|
||
- **Confidence: low / directional** = practitioner heuristic, no primary source. Treat as
|
||
a hypothesis to test on your own account, not a fact.
|
||
|
||
**Rule of thumb: trust the *ordering*, test the *number*.**
|
||
|
||
## Officially-named ranking signals (the only two LinkedIn confirms by name)
|
||
|
||
| Signal | Direction | Source | Confidence |
|
||
|--------|-----------|--------|------------|
|
||
| Dwell time | Time spent on a post is a ranking input (active vs passive tasks; long-dwell modeled). No public weight. | LinkedIn Eng — "Leveraging Dwell Time" (2024) | high |
|
||
| Topic / interest relevance | Content matched to a viewer's interests is distributed — *including beyond your network* for strong content. | Tim Jurka, Head of Feed AI (2025-08-11) | high |
|
||
|
||
Everything below this line is direction + sourced estimate, not officially-weighted.
|
||
|
||
## Engagement order (not coefficients)
|
||
|
||
The defensible spine is the **order**, not the multiplier:
|
||
|
||
> **saves > shares > quality comments > reactions/likes**
|
||
|
||
| Signal | Direction / estimate | Source | Confidence |
|
||
|--------|----------------------|--------|------------|
|
||
| Saves | Top engagement signal; also a follow-graph signal (saving a post raises the author's next-post feed odds). ≈ 5x a like / ≈ 2x a comment in single-vendor data. | AuthoredUp, Vertebrae, van der Blom (1.8M) | medium |
|
||
| Shares (feed + DM) | Strong distribution signal; public endorsement. | van der Blom (1.8M) | medium |
|
||
| Quality comments (15+ words) | Substantive comments outweigh short ones; **comment ≈ 2x a like** (quality-scored, single vendor). The popular "comment = many-x a like" claim is **unverified folklore** — dropped. | AuthoredUp (NLP-quality-scored) | medium |
|
||
| Reactions / likes | Baseline engagement unit (≈ 1x). | van der Blom (1.8M) | medium |
|
||
|
||
> **Note on the old "comment = 15x" / "= 5x" framing:** there is no primary source for it.
|
||
> The "5x" was the **saves** figure mis-assigned to comments. Encode the order above; do
|
||
> not quote a comment multiplier.
|
||
|
||
## Content format
|
||
|
||
| Format | Direction / estimate | Source | Confidence |
|
||
|--------|----------------------|--------|------------|
|
||
| Documents / carousels | **Top organic format (~7%, Socialinsider, company-page per-impression).** "Carousel" = PDF document post (LinkedIn removed native carousels Dec 2023). The 7% / 21.8% / 49.5% spread across studies is a **denominator artifact**, not disagreement about the winner. | Socialinsider (1.3M), Buffer (2M), Metricool (673K) | high (rank) / medium (number) |
|
||
| Native video | #2 format and **declining**; add captions (most watch muted). No hard aspect-ratio gate — 4:5 / 1:1 preferred, captions are the enforceable spec. | Socialinsider; van der Blom | medium |
|
||
| Text-only | Most resilient format; generates the best comment quality. | Buffer, van der Blom | medium |
|
||
| Multi-image | Strong, slightly below documents. | Socialinsider | medium |
|
||
| Polls | Declining effectiveness; audience research only. | van der Blom | low / directional |
|
||
| Link posts (link in body) | Underperform — see external links below. | Ordinal (900K) | medium |
|
||
|
||
> The personal-profile **per-post baseline** (~2.0–2.6%) is a *different denominator* from
|
||
> a format benchmark — never present an account baseline as a carousel rate.
|
||
|
||
## External links (in post body)
|
||
|
||
| Claim | Statement | Source | Confidence |
|
||
|-------|-----------|--------|------------|
|
||
| Reach effect | **Correlational reach reduction (~38% in 2026; contested band ~19–60% across studies).** It is a moving number, not a flat tax. | Ordinal (900K, p<0.001); van der Blom; DigitalApplied | medium |
|
||
| Intent | **LinkedIn denies an *intentional* penalty** (Sr. Director Product, Aug 2025): no penalty "if the post leads with value" — the effect is engagement-driven. | Matt Navarra (relaying LinkedIn) | medium |
|
||
| First comment | **Neither a magic fix nor a confirmed penalty** — contested. Lead with standalone value; native formats are the durable answer. | Ordinal; practitioner blogs (no large-N) | low |
|
||
|
||
> **Design rule:** value-first matters more than link location. Soften any enforcing hook
|
||
> from a hard "−X% penalty" mechanic to "body links correlate with lower reach — prefer a
|
||
> first comment, but lead with value either way."
|
||
|
||
## Early-engagement window + evergreen resurfacing
|
||
|
||
| Claim | Statement | Source | Confidence |
|
||
|-------|-----------|--------|------------|
|
||
| Golden window | **60–90 min** (90 is the 2026 consensus); the **first 15–30 min** is the highest-leverage sub-window (~70% of reach decided there). | Buffer; Expandi; van der Blom | high |
|
||
| First-hour velocity | Strong early engagement unlocks broader distribution. Directional, not a fixed threshold. | van der Blom | medium |
|
||
| Evergreen resurfacing | The relevance model **can resurface strong-save / high-dwell posts days-to-weeks later** on viewer intent (posts now live 2–3 weeks vs days). **No** confirmed fixed "24–72h reinjection" rule — it is intent-driven and irregular. | AuthoredUp | medium (direction) / low (timing) |
|
||
|
||
## Profile / topic alignment
|
||
|
||
| Claim | Statement | Source | Confidence |
|
||
|-------|-----------|--------|------------|
|
||
| Topic alignment is a ranking input | Real and officially confirmed (qualitatively): topic/interest relevance drives distribution, including beyond your network. | Tim Jurka (2025-08-11) | high |
|
||
| Off-topic reach reduction magnitude | **No primary source** states a discrete off-topic reach-reduction figure. Treat profile/topic alignment as a real input; **do not quote a percentage.** | — | n/a (figure removed) |
|
||
|
||
## AI-content down-rank (officially confirmed — justifies the de-AI gate)
|
||
|
||
| Claim | Statement | Source | Confidence |
|
||
|-------|-----------|--------|------------|
|
||
| AI-slop suppression | LinkedIn confirmed an **active program** suppressing (1) generic AI-written posts/comments, (2) automation tools, (3) attention-bait video. ML models distinguish "original thinking" from "posts lacking substance"; flagged posts are **reach-suppressed (reportedly to first-degree), not deleted.** | VP & Exec Editor Laura Lorenzetti (2026-05-19) | high |
|
||
| Correlational engagement gap | Likely-AI posts saw ~45% less engagement (correlational). | Originality.ai (8,795 posts) | medium |
|
||
| Engagement-pod crackdown | Auto-comments demoted out of "Most Relevant", scoped to own network; repeat offenders restricted. | VP Product Gyanda Sachdeva (2026-02-16) | high |
|
||
|
||
> **Enforce what LinkedIn *named*** — personal substance, original thinking, concrete
|
||
> specifics, genuine voice — not an unverified SEO "tell-list."
|
||
|
||
## Buzzwords
|
||
|
||
Buzzword avoidance is **editorial guidance for clarity, not a measured reach mechanic.**
|
||
No primary source ties specific words to a reach penalty. Keep the buzzword list (it
|
||
improves writing); do not justify it as "reduces reach."
|
||
|
||
| Claim | Source | Confidence |
|
||
|-------|--------|------------|
|
||
| Specific phrasing reads better than corporate generic | Inc. (editorial) | low / directional |
|
||
| A semantic ranker *may* indirectly favor specific over generic phrasing | inferred | low (not confirmed) |
|
||
|
||
## The deployed ranking model — what we can and cannot say
|
||
|
||
> **LinkedIn's feed ranking model has an official name: the Generative Recommender (GR).**
|
||
> Announced 2026-03-12 on LinkedIn's engineering blog (Hristo Danchev,
|
||
> [Engineering the next generation of LinkedIn's feed](https://www.linkedin.com/blog/engineering/feed/engineering-the-next-generation-of-linkedins-feed)):
|
||
> a sequential transformer-based ranker that treats member interaction history as a
|
||
> timeline, paired with a unified LLM-embedding retrieval system. Rollout announced in
|
||
> the same post.
|
||
|
||
| Claim | Statement | Source | Confidence |
|
||
|-------|-----------|--------|------------|
|
||
| Production name | **Generative Recommender (GR)** — official, primary-source. | LinkedIn engineering blog, 2026-03-12 | high |
|
||
| LLM-based retrieval | Confirmed: "a unified retrieval system leveraging advances in LLMs to generate a high-quality representation of our members and content." | Same post | high |
|
||
| Deployment | Rollout announced 2026-03-12 ("rolling out a new advanced ranking system"). Full-coverage completion date not stated — do not assert one. | Same post | high (announcement), n/a (completion) |
|
||
| "360Brew" as the production name | **Still not publishable.** The arXiv paper (2501.16450) is a Jan-**2025** *pre-production research* model (V1.0, 150B params, offline parity only), **withdrawn 2025-08-23**; the "360Brew" label is third-party and has no official confirmation. GR is the official name. | arXiv 2501.16450 | high (on the negative claim) |
|
||
|
||
*Correction note (2026-07-17): this section previously said "No public name. No
|
||
deployment date." and flagged the Generative Recommender / Hristo Danchev
|
||
engineering-post citation as likely fabricated. That flag was wrong — and was already
|
||
wrong at "Last updated 2026-05": the official post had been live since 2026-03-12,
|
||
two months earlier. The fabrication flag rejected a genuine primary source.*
|
||
|
||
## Operational heuristics (directional — test per account)
|
||
|
||
These are creator-community heuristics with no primary-source weights. Use as starting
|
||
hypotheses, not targets. Confidence: **low / directional** for every row.
|
||
|
||
### Engagement velocity (first 90 min)
|
||
|
||
| Time | Rough target | If well below |
|
||
|------|--------------|---------------|
|
||
| 15 min | a few | check timing / hook |
|
||
| 30 min | building | engage in comments |
|
||
| 60–90 min | momentum | golden window closing |
|
||
|
||
### Posting time windows (CET / European audience)
|
||
|
||
| Day | Commonly-cited peak |
|
||
|-----|---------------------|
|
||
| Tue | 8–11 AM (often best overall) |
|
||
| Wed | 8 AM, 12 PM |
|
||
| Thu | 9 AM–1 PM (extended) |
|
||
| Fri | before 3 PM |
|
||
| Mon | 7–9 AM |
|
||
| Weekend | weaker |
|
||
|
||
*For global audiences: post 8–11 AM local to catch multiple zones.*
|
||
|
||
### Quick decision rules
|
||
|
||
| Situation | Decision |
|
||
|-----------|----------|
|
||
| Linking? | First comment, lead with value either way |
|
||
| Multiple ideas? | Split into separate posts |
|
||
| Off your usual topic? | Topic alignment is a real input — stay on-domain or accept lower reach |
|
||
| Video or text? | Text for authority, video (captioned, 4:5/1:1) for connection |
|
||
| Carousel or text? | Documents for frameworks/guides, text for stories/opinions |
|
||
| Comment or like first? | Comment (higher in the engagement order) |
|
||
|
||
### Comment strategy (CEA)
|
||
|
||
1. **Compliment** — a specific point you appreciated
|
||
2. **Expand** — your insight or related experience
|
||
3. **Ask** — a question to continue dialogue
|
||
|
||
Minimum quality: 15+ words, genuine perspective. AI-generated / "Great post!" comments are
|
||
actively suppressed (see AI-slop down-rank).
|
||
|
||
## 2026 reach context
|
||
|
||
Organic reach declined platform-wide in 2026 — focus on **relative performance** (your
|
||
posts vs your own baseline), not absolute numbers. Smaller engaged audiences outperform
|
||
large passive ones. (Direction: high confidence; exact YoY %: directional, varies by
|
||
source.)
|
||
|
||
---
|
||
|
||
*Last updated: 2026-07-17 (GR-model correction). Maintained as the single canonical
|
||
algorithm statement; cite, do not restate.*
|
||
|
||
*Sources (per-claim quality/confidence noted inline): LinkedIn Engineering — "Engineering
|
||
the next generation of LinkedIn's feed" (Hristo Danchev, 2026-03-12); arXiv 2501.16450 (pre-production
|
||
research paper, withdrawn 2025-08-23); LinkedIn Engineering — "Leveraging Dwell Time" (2024); Tim Jurka, Head of Feed
|
||
AI (2025-08-11); Laura Lorenzetti, VP & Exec Editor (2026-05-19); Gyanda Sachdeva, VP
|
||
Product (2026-02-16); Matt Navarra relaying LinkedIn Sr. Director Product (Aug 2025);
|
||
Ordinal link-penalty study (900K, p<0.001); Socialinsider (1.3M); Buffer (2M+); Metricool
|
||
(673K); AuthoredUp (621K, NLP-quality-scored); van der Blom Algorithm Insights 2025 (1.8M);
|
||
Originality.ai (8,795 posts); Inc. (buzzword editorial). Full provenance: research brief
|
||
`docs/remediation/research/01-linkedin-algorithm-signals.md`.*
|