360Brew Research
✓ v2 — Primary Sources
LinkedIn 360Brew
Optimisation Checklist
Rebuilt from LinkedIn’s own engineering blog (March 2026), the arXiv retrieval paper (Oct 2025), and the 360Brew research paper — not secondary marketing blogs.
Overall progress0 / 0 done
How the system actually works — two separate models
Stage 1 — Retrieval (Causal LLM)
A fine-tuned LLaMA-3 dual encoder narrows hundreds of millions of posts down to ~2,000 candidates by matching your semantic profile embedding to post embeddings. Your profile headline, summary, skills, job history, and activity history are all encoded here.
Stage 2 — Ranking (360Brew)
360Brew (a ~150B-parameter decoder-only transformer) re-ranks those ~2,000 candidates using your full engagement sequence as a narrative — not just isolated clicks. It predicts what comes next in your professional interest arc.
Most LinkedIn advice focuses only on 360Brew (ranking). But if your content doesn’t pass Stage 1 retrieval, 360Brew never sees it. Both stages need optimising. Source: arXiv:2510.14223 + LinkedIn Engineering Blog, March 2026.
Signal source key (dot on each item)
Confirmed by LinkedIn engineering source
Strongly inferred from research papers
Avoid — penalised behaviour
Tactical tip / best practice
Profile as a Semantic Signal
Your profile metadata travels with every post you publish – confirmed by LinkedIn engineering
Confirmed by LinkedIn Engineering (March 2026): Your profile metadata – headline, listed skills, work history, summary – is bundled into the semantic embedding of every post you publish. The retrieval LLM uses it to determine relevance and audience targeting. A vague headline weakens every post’s distribution automatically.
✓
Write a specific headline that clearly names your professional niche in plain language. Because your headline is encoded into your post embeddings at retrieval stage, a generic headline makes your posts harder to match to the right audience.
✓
Align your headline, About section, and recent posts to the same topic cluster. The retrieval model encodes all of these together – conflicting signals reduce your matching accuracy across both retrieval and ranking stages.
✓
Complete all profile fields including skills, job history, education, certifications, and languages. The retrieval model’s member prompt uses all of these fields — gaps reduce embedding quality.
✓
Write a substantive About section that reads as coherent professional prose. The retrieval LLM’s system prompt reads it as text — a well-crafted summary produces a richer semantic embedding than a list of keywords.
✓
Keep your profile current. When members update their profiles, LinkedIn regenerates their embedding within minutes. Stale profiles create a mismatch between your actual expertise and what the system surfaces your content for.
✓
Invest in profile quality for spokespeople and staff who post under their own names on behalf of your business. The profile-to-post embedding connection means their individual profiles affect your brand’s content distribution.
✓
Use a custom banner that communicates your value proposition. It is the first thing a profile visitor sees and drives the profile dwell time that signals relevance — though this is a user experience signal, not directly confirmed in the engineering papers.
Retrieval Stage — Getting Past Gate 1
The Causal LLM retrieval stage is separate from 360Brew — most advice ignores this entirely
From the arXiv paper you uploaded: The retrieval model uses your positive engagement history (posts you have actually interacted with) plus your full profile to build your member embedding. Posts are then matched by cosine similarity. Optimising for retrieval means making both your profile text and your content semantically coherent and consistent.
✓
Build a consistent positive engagement history on LinkedIn by genuinely interacting with content in your niche. The retrieval model trains your member embedding using only your positive PI (Professional Interaction) actions — likes, comments, reposts, long dwells.
✓
Write posts whose text semantically matches your profile. Cosine similarity between your member embedding and a post embedding determines retrieval ranking. The closer your post text is to your established topical identity, the higher it scores.
✓
The retrieval system is designed to surface your content to non-connections whose interest profiles match yours. This is LinkedIn’s “suggested content” mechanic — optimising your topical niche directly expands your potential audience beyond your network.
✓
New and smaller-network members benefit most from retrieval optimisation. The arXiv paper shows the LLM retrieval system produced +1.17% daily interactions and +3.29% revenue uplift specifically for members with fewer connections. If you are building your network, topical alignment matters more, not less.
✓
Keep post popularity features in mind. The retrieval model encodes post popularity counts (likes, views above a threshold) as quantised features alongside text. Early engagement on a post therefore improves its retrievability for subsequent audience waves.
✓
Author information is encoded at retrieval stage. Your author name, profile headline, company, industry, and title are features in the post embedding — not just the post text. This reinforces why profile quality directly affects post distribution.
Ranking Stage — How 360Brew Scores You
360Brew reads your audience’s engagement history as a narrative arc, not isolated clicks
From LinkedIn Engineering Blog (March 2026): The Generative Recommender (360Brew at ranking) processes over 1,000 of each viewer’s past interactions as a sequence — a professional interest narrative. It asks “given where this person has been going, what comes next?” — not just “has this person liked this topic before?”
✓
Create content that fits the professional interest arc of your target audience. 360Brew is predicting continuations of a journey, not pattern-matching to static preferences. Content that fits naturally into a progression (e.g. procurement strategy → tendering → supplier management) gets ranked higher for the right audience.
✓
Dwell time is a top-tier ranking signal — even passive reading counts. LinkedIn’s confirmed engineering work shows the ranking model weights passive signals (reading slowly, pausing) separately from active ones (liking, commenting). Someone reading carefully with no click still sends a positive signal.
✓
Posts scrolled past without any engagement are “hard negatives.” LinkedIn’s engineers use this exact term — posts that receive zero engagement from a shown audience actively train the system to show you less to similar users. Every low-performing post has a cost beyond just low reach.
✓
Topic consistency builds a legible trajectory for 360Brew. Because the model reads engagement history as a sequence, people who post coherently within a subject area build a clearer signal of who their content should reach. Jumping between unrelated topics creates noise in the system’s audience-matching.
✓
Saves, thoughtful comments, and shares carry more weight than likes. This is inferred from how the system defines “Professional Interactions” (PIs) — which include long dwell, react, comment, repost. Actions signalling genuine value (saves, substantive replies) score higher than quick reactions.
✓
360Brew is capable of zero-shot reasoning — it can assess relevance for content or people it hasn’t encountered before by understanding semantic patterns. This means new voices with strong topical clarity can gain reach without a long engagement history, as long as their profile and content are consistent.
Content Quality & Format
What the engineering sources actually confirm about content signals
Note on format statistics: Figures like “carousels get 6.6% engagement” or “vertical videos retain 87% of viewers” come from third-party analytics tools, not LinkedIn’s engineering sources. They are plausible but treat them as directional, not precise.
✓
Write posts that earn genuine reading time. The ranking model weighs dwell time as a primary positive signal (confirmed). Design your posts so that a professional in your target audience would read past the first line — not just stop-scroll.
✓
Open with a hook that earns the “See More” click. This is the first active engagement signal — it matters for early-stage velocity, which affects how far the post is distributed. Confirmed by multiple sources as a make-or-break moment in the first 1–3 hours.
✓
Upload video natively. External video links reduce dwell time on LinkedIn (users leave the platform). Native video keeps attention on-platform, which is what the system measures. This is confirmed directionally by LinkedIn’s own platform behaviour incentives.
✓
Use plain, semantically clear language. The retrieval LLM encodes meaning, not keywords. Clear, direct prose with specific professional vocabulary produces a stronger, more accurate embedding than jargon-heavy or buzzword-laden text.
✓
Rewrite AI-generated drafts in your own voice before publishing. The system was trained on authentic professional engagement data. Formulaic AI-written text produces weaker embeddings and likely generates more “scroll past” behaviour from readers, generating hard negatives.
✓
Keep external links out of the post body. LinkedIn’s incentive model penalises content that moves users off-platform. Place links in the first comment. This is widely confirmed by practitioners and is consistent with LinkedIn’s platform behaviour design.
✓
Avoid excessive tagging and hashtag stuffing. These are spam signals identified at the pre-distribution filter stage. Limit tags to those directly relevant; use 1–3 hashtags at most. There is no confirmed “5 tag limit” in LinkedIn’s engineering sources, but the principle of avoiding spam-like patterns is sound.
Engagement Behaviour
How to behave on-platform to build your member embedding and early post velocity
✓
Be present and reply to comments in the first few hours after posting. Early engagement velocity influences further distribution. A thoughtful reply extends the engagement window and signals to the system that a genuine professional conversation is taking place — confirmed by LinkedIn engineering.
✓
Write substantive comments on others’ posts in your niche. These count as Professional Interactions and contribute to your positive engagement history, which shapes your member embedding at retrieval stage. “Great post!” adds nothing — multi-sentence responses do.
✓
Only engage with content you genuinely find valuable. The retrieval model trains on your positive PI history specifically — it learns your interests from what you actually engage with. Hollow engagement (liking things you didn’t read) adds noise to your profile embedding.
✓
Think carefully before posting filler content. Posts that receive zero engagement from people who were shown them generate hard negatives. Over time, this trains the system to distribute your content more narrowly. A post that doesn’t land has a cost beyond just zero impressions.
✓
Create genuinely useful content that earns saves. A save signals the viewer intends to return — this is evidence of lasting value. It is a stronger behavioural signal than a like because it represents intentional deferral, not a momentary reaction.
✓
Track saves and profile views as your primary qualitative metrics — not likes. These are the signals most consistent with how the engineering systems define genuine engagement value. Follower count and like volume are increasingly poor proxies for actual reach quality.
Tactics to Abandon
Confirmed penalised behaviours from primary sources
✓
Stop using engagement pods. LinkedIn’s VP of Product confirmed in November 2025 that eliminating pod effectiveness is an explicit goal. The system detects coordinated activity rings — same-cluster accounts engaging rapidly after publish. Lempod was banned and removed from the Chrome Web Store.
✓
Remove third-party comment scripts and browser plugins. LinkedIn confirmed in February 2026 that comments posted through third-party scripts are removed from the “Most Relevant” section. Any automation that mimics engagement is identifiable and penalised.
✓
Stop treating likes as your primary success metric. The engineering systems that matter — retrieval and ranking — optimise for Professional Interactions (long dwell, comments, reposts, saves), not likes. Chasing like counts means optimising for a signal that the system largely ignores.
✓
Stop posting at high volume with low quality. Each low-performing post (generating hard negatives) has a compounding cost on your future distribution. Fewer, better posts is the correct model — confirmed by how the training loop works in the research papers.
✓
Stop scattering across unrelated topics. Because both retrieval (via cosine similarity on embeddings) and ranking (via sequence trajectory modelling) depend on topical consistency, frequent topic-switching degrades performance at both stages simultaneously.
Long-term Strategy
Compounding your authority across both retrieval and ranking over time
Allow 90 days. Multiple secondary sources indicate it takes approximately 90 days of consistent, aligned posting for the system to fully build your topical profile and optimise your distribution. This is consistent with how LLM embedding systems accumulate signal over time, though the 90-day figure itself is not in LinkedIn’s engineering papers.
✓
Plan content around a consistent topic cluster for at least 90 days. Both retrieval (member embedding accuracy) and ranking (sequence trajectory signal) improve as you build a body of topically aligned work. This is a compounding advantage — early consistency pays dividends continuously.
✓
Treat LinkedIn as a discovery engine for non-connections. The confirmed engineering goal of the LLM retrieval system is to surface relevant content from outside your network. Consistent topical alignment means you are continuously eligible for discovery by people who have never heard of you but are interested in your niche.
✓
Optimise for LinkedIn Search alongside the feed. The same retrieval system that powers the feed also powers LinkedIn Search. Posts and profiles with clear, specific language rank in search results and notifications — content lifespan extends well beyond the initial feed window.
✓
Accept that overall raw reach may be lower but more targeted. This is confirmed by LinkedIn’s own A/B test data: the LLM retrieval system improved relevance significantly, particularly for smaller-network members. A smaller, highly engaged niche audience is the system’s intended output — it is not a side effect.
✓
The system rewards substance, expertise, and specificity — which are competitive advantages for genuine specialists and practitioners over generalist content creators. The shift from engagement hacks to semantic relevance favours people who actually know what they are talking about.