On March 12, 2026, LinkedIn published the most detailed account it has ever given of how its Feed actually works: “Engineering the next generation of LinkedIn’s Feed,” written by Hristo Danchev, Senior Staff TPM at LinkedIn. It’s a rare thing — a platform’s own engineering team explaining, in technical detail, the architecture behind ranking decisions that usually stay hidden.
Read closely, two of the specific choices LinkedIn’s engineers made are structurally identical to decisions already built into Clue Labs, before this post existed.
LinkedIn Algorithm Engineering: What Actually Shipped
The post describes a full rebuild: a multi-source retrieval system replaced with a unified two-layer architecture. First, a dual-encoder retrieval model — a shared LLM that reads both the member and the candidate post and compares them for relevance — pulls a pool of candidates. Second, a transformer-based Generative Recommender (GR) model ranks that pool.
This is LinkedIn algorithm engineering at a level of detail platforms rarely publish: architecture diagrams, training details, and the specific problems each design choice was built to solve.
How LinkedIn Feed Ranking Works Now
The retrieval layer is where the interesting engineering decisions live. Instead of matching posts to members by keyword or by who they follow, the dual encoder reads meaning — processing a member’s history and a candidate post’s content and scoring them by relevance in embedding space.
LinkedIn reported the retrieval system now runs with sub-50ms latency, using a custom Flash Attention variant that delivers an additional 2x speedup, alongside hard negative sampling that produced a 3.6% recall gain. That’s the plumbing. The two decisions worth Clue Labs’ attention sit inside it.
Percentile Engagement Scoring: The Fix That Boosted Retrieval by 15%
LinkedIn’s engineers found that raw engagement numbers — a post with, say, 12,345 views — tokenize badly for a language model and correlate poorly with actual relevance. Their fix was percentile engagement scoring: converting a raw count into a percentile rank, wrapped in a token like <view_percentile>71</view_percentile>, telling the model a post sits above average without handing it an unstable raw number to reason over.
The result, per LinkedIn’s own reporting: a 30x improvement in correlation between popularity features and embedding similarity, and a 15% improvement in recall@10 — meaning the top 10 retrieved posts were measurably more relevant.
This mirrors a principle already built into how Clue Labs tracks and measures KPIs. We don’t report performance as a standalone raw count either — it’s always weighed relative to a distribution, not presented as a number on its own. It’s the same underlying fix LinkedIn just described: raw counts don’t compare meaningfully across accounts, sectors, or time periods by themselves. We didn’t model this after LinkedIn. We arrived at the same principle, independently, for the same underlying problem.
Passive vs Active Engagement: Distinct Targets, Not One Blended Score
The second decision is just as telling. LinkedIn’s ranking model splits engagement into two separate towers: one predicting passive signals — clicks, skips, long-dwell — and a second predicting active signals — likes, comments, shares. Rather than training toward one blended “engagement” number, the model treats these as distinct prediction targets, each carrying different information about whether a post is actually working.
That’s the same logic behind Clue Labs’ KPI-differentiator approach in the Clue Score. We don’t score “engagement” as a single number either — we score specific, distinct signals, because a post that gets read slowly and never shared is telling you something different than a post that gets shared instantly and abandoned.
The AI Recommendation Architecture Clue Labs Already Built
Put the two together and what LinkedIn published is a confirmation, from the inside, of an AI recommendation architecture built on two principles Clue Labs already treats as foundational: performance is measured relative to a distribution rather than presented as a standalone raw count — the same principle behind how we track and measure KPIs — and engagement is a set of distinct behaviours to predict separately, not one number to blend together.
Why Independent Convergence Beats Imitation
This isn’t “we studied the algorithm and made educated guesses.” It’s independent convergence — arriving at the same structural solutions LinkedIn’s engineers landed on, for the same underlying problem, without having seen their post first. That’s a stronger claim than modelling-after. It’s arriving at the same answer.
It’s also a direct rebuttal to the “the algorithm is basically random” narrative that circulates every time a low-effort post outperforms months of considered work. It isn’t random. It’s a legible, engineered system, built on exactly the kind of structured signal design Clue Labs already uses. The unpredictability creators feel is a visibility problem, not a randomness problem — and understanding the architecture underneath it is the whole premise of what we’re building.
Source:
Hristo Danchev, “Engineering the next generation of LinkedIn’s Feed,” LinkedIn Engineering Blog, March 12, 2026. linkedin.com/blog/engineering/feed/engineering-the-next-generation-of-linkedins-feed
