How LinkedIn's feed decides who sees you

What's actually running

In March 2026, LinkedIn published its most detailed feed disclosure yet. The pipeline is a single LLM-based dual-encoder that finds candidate posts by nearest-neighbor search on embeddings, then ranks them with a transformer called the Generative Recommender¹. Embeddings refresh within minutes; the pipeline runs in under 50 milliseconds for more than 1.3 billion members.

The ranker doesn't score posts one at a time. It reads a sequence of more than a thousand of a member's past interactions, alongside profile fields, device type, and affinity features¹. Format is one input among many. LinkedIn discloses no boost or penalty tied to it.

The dwell-time signal has a longer paper trail. In 2020, LinkedIn's engineers disclosed two dwell measurements (on the feed and after the click) plus a P(skip) model predicting whether dwell time falls below a skip threshold. Adding dwell features lifted the ranker's offline AUC by as much as 10%². The architecture has since been rebuilt; the doctrine hasn't. Clicks alone are noisy; time spent is the truer signal.

The multipliers nobody can source

LinkedIn's last on-the-record statement on format was a denial. In 2019, Pete Davies, then LinkedIn's Senior Director of Product for the feed, wrote: "Despite the rumors, the algorithm doesn't favor any particular format"³. Six years and two rebuilds later, LinkedIn has never reaffirmed or retracted that position.

Meanwhile, creators repeat the opposite claim: carousels earn about 1.9× the reach of other formats, documents outperform images by roughly 3×, video over three minutes reaches 1.21× the median. Every number traces to Buffer, Oktopost, or similar studies. None cite LinkedIn, and none reconcile with the 2019 denial. The 2026 engineering post names format as an input the ranker sees, not a lever that shifts distribution¹.

Two things can be true. Format can correlate with reach without the algorithm rewarding format: a carousel holds the scroll longer, so its dwell time is higher, so its engagement is stronger. That's a signal effect, not a format multiplier. When a guide quotes 1.9× as fact, it's inventing a number LinkedIn has never confirmed.

The 2026 change most guides missed

In May 2026, LinkedIn changed how it distributes generic AI-generated content. Laura Lorenzetti, LinkedIn's VP of Product, laid out the policy. Posts the system flags as generic or inauthentic aren't removed. They stay fully visible to the poster's first-degree connections. What changes is reach: flagged posts are suppressed from recommendations beyond that circle⁴. In testing, LinkedIn says the system correctly identifies generic content 94% of the time. It doesn't say by how much reach drops.

Lorenzetti's post welcomes AI-assisted writing that carries an original angle or sparks conversation. What gets suppressed is content produced at scale with no added judgment. That is the same pattern LinkedIn's 2020 P(skip) model was already built to punish. The policy formalizes what the ranker was already doing.

The signals LinkedIn has named

LinkedIn names its positive signals outright. The engineering post lists likes and shares among the interaction types the ranker embeds and reads as sequences¹; the public Help Center adds reactions, comments, reposts, and sends as the specific engagement types that shape what shows up in your feed⁵. The Help Center also names three negative signals: Not interested, unfollow, and Report post — each reduces similar content in that member's feed⁵.

What LinkedIn hasn't disclosed: how those signals weigh against each other, what counts as a "long dwell" in minutes, or any decay curve for how fast a post's reach fades. Claims like "the first 90 minutes decide reach" or "5–10% of your network sees it first" trace to creator blogs, not LinkedIn.

Two things to act on today

Two places to start:

• Write posts that hold attention long enough to earn a long dwell. Long dwells and comments are two of the four engagement signals LinkedIn names as ranker inputs. Chapter 3 walks through the post structures that keep a reader on the screen past the "See more" fold.
  Read Chapter 3 arrow_forward
• Check whether your Followable layer is the one costing you. The Scorecard's routing table flags the layer scoring lowest across the five: Findable, Followable, Verifiable, Credentialed, Introducible. Four minutes, no signup.
  Take the Scorecard arrow_forward