Nobody outside LinkedIn knows the ranking function. LinkedIn is unusual, though, in that its engineering team publishes papers about the machinery. That gives you an architecture, not a formula, and the two are often confused.
The short answer: LinkedIn splits the problem into retrieval (which posts are even candidates) and ranking (what order they appear in), and as of 2026 both stages are model-driven rather than rule-driven.
What LinkedIn says it optimises for
In “Engineering the next generation of LinkedIn’s Feed,” published 12 March 2026, LinkedIn Engineering describes replacing several separate retrieval sources with a single semantic retrieval system built on a dual-encoder language model, and replacing the ranking stage with a sequential transformer it calls the Generative Recommender.
The stated split, in LinkedIn’s framing: “retrieval determines which posts reach the ranking stage, ranking determines what a member actually sees.”
The ranking model reads a member’s history in order, processing over 1,000 past interactions with causal attention, meaning it never looks ahead. Notably, LinkedIn says training was filtered to posts that received positive engagement, dropping scrolled-past items.
Note: Figures here were verified against the LinkedIn Engineering blog post “Engineering the next generation of LinkedIn’s Feed” (12 March 2026) and LinkedIn Engineering’s earlier post on feed dwell time, as of September 2026. Platforms change these without notice.
What signals does LinkedIn rank on?
Ordered by the emphasis in LinkedIn’s own writing.
| Signal group | What LinkedIn names | Notes |
|---|---|---|
| Active engagement | Likes, comments, shares, extended reading time | The positive signals the ranking model is trained on |
| Passive behaviour | Dwell time, scrolling past without engaging | LinkedIn Engineering has a dedicated paper on dwell time as a ranking input |
| Sequence of your interactions | Over 1,000 prior interactions, in order | The core input to the Generative Recommender |
| Semantic match | LLM embeddings of member and post, compared by cosine similarity | Decides candidacy, before ranking |
| Profile context | Industry, skills, geography | Used as features |
| Engagement counts | Encoded as percentile buckets rather than raw numbers | Blunts the advantage of very large accounts |
| Recency | Explicit recency signals | Present, but not described as dominant |
The percentile-bucket detail is the most useful thing in that list. It implies LinkedIn deliberately does not feed raw follower or like counts into ranking, which is the opposite of how most people describe the platform.
What changed most recently
Stated by LinkedIn: the retrieval and ranking rebuild, March 2026. Dwell time as an explicit ranking input, published earlier by the same team.
Widely reported, not confirmed: the claim that LinkedIn tests a post on a small slice of your network and expands from there (“the golden hour”). LinkedIn Engineering does not describe a staged rollout in these terms. What it describes is a candidate-generation and ranking pipeline that runs per member, per session. Early engagement plausibly matters because it changes what the model has seen, but the tidy hourly thresholds people quote are not sourced.
Practitioner heuristic: the belief that external links suppress reach. See below.
Myths this kills
“The algorithm hates links.” This is the single most repeated claim about LinkedIn and it has never been confirmed by LinkedIn. There is no published statement that a post containing a URL is demoted. What is plausible, and consistent with LinkedIn’s own description of ranking on dwell and engagement, is that a post which sends people off-platform accumulates less of both. That is an outcome, not a penalty, and the fix is to write a post worth reading on its own rather than to hide the link in the comments.
Shadowbans. No confirmed mechanism. LinkedIn does restrict accounts and remove content under its policies, and it tells you when it does. An unexplained reach drop is more often a change in what you posted. See what a shadowban actually is.
Engagement pods. Unfounded as a strategy and directly at odds with the published architecture. If engagement counts enter ranking as percentile buckets, a pod is buying you a bucket boundary at best. Worse, pod engagement comes from people whose interaction history does not match your target audience, which is exactly the signal a sequence model learns from.
Posting time is decisive. Overstated. LinkedIn’s audience is concentrated in working hours, so the floor is real, but a one-hour shift is not what separates a post that travels from one that does not. Our best time to post on LinkedIn guide and the LinkedIn posting time tool handle the timezone maths. Cadence matters more, covered in how often to post on LinkedIn.
What actually moves distribution, in priority order
- Whether people stop and read. Dwell time is a documented input. The first two lines before the “see more” fold decide it.
- Comments with substance. They are active signals and they extend a post’s life.
- Topical clarity. Semantic retrieval matches meaning, so a post about one clear thing is easier to route than a post about four.
- Relevance to your actual network’s interests. The model is built from each viewer’s own interaction sequence.
- Reposts and shares with commentary.
- Consistency. Not a ranking signal. It is how you get enough samples to learn anything.
What to do
Write the post so the first two lines are worth the click on “see more.” Say one thing. Answer the comments, because replies are active engagement from exactly the audience the model is watching. Put the link wherever it reads naturally.
Scheduling LinkedIn alongside everything else
Cadence is the part of this you can systematise. BulkPublish schedules LinkedIn personal profiles and company pages from the same calendar as the other 14 platforms, with a per-channel queue so a LinkedIn post does not have to follow an Instagram schedule. The calendar feature page shows the layout.
The parts you control
Format, hook, consistency. LinkedIn has told you more about its machinery than most platforms, and the honest conclusion from reading it is that the machinery mostly rewards a post people actually stop for. That is not a hack, and there is no version of it you can buy.