LinkedIn’s new AI retrieval system gives low-connection accounts 3-4x larger performance gains than established ones, turning a well-written profile into the fastest path out of obscurity.
Every LinkedIn strategy post starts the same way: build your network first. Get to 500 connections. Engage consistently for months before you expect results. The implication is clear: new accounts are at a disadvantage, and you need to grind your way out of it.
LinkedIn’s own engineering research says the opposite.
The old algorithm punished new accounts
Until recently, LinkedIn’s feed relied heavily on network signals. Collaborative filtering. Trending content in your industry. Engagement patterns from people similar to you. All of these systems need data to work, and a new account doesn’t have any.
A fresh profile with 50 connections and no engagement history was essentially invisible. The algorithm couldn’t match you to relevant content because it didn’t know what “relevant” meant for you yet. So it showed you generic posts from big accounts, you didn’t engage because none of it was interesting, and the system learned nothing. A cold start that stayed cold.
This is the world most LinkedIn advice was written for.
What actually changed
In late 2025, LinkedIn replaced its patchwork of retrieval systems with a single fine-tuned LLaMA-3 model. Instead of relying on network graphs and behavioral history, this system reads your profile text, generates a 3,072-dimensional embedding of your professional identity, and matches it against content embeddings using cosine similarity.
The shift is fundamental. The old system asked: who do you know, and what have they engaged with? The new one asks: what does your profile say about what you care about?
For established accounts with years of engagement data, this is a modest improvement. For new accounts, it’s transformative.
The numbers LinkedIn published
LinkedIn’s research paper on the Causal LLM retrieval system includes a breakdown that most people skimmed past:
| Metric | Overall Gain | Low-Connection Users |
|---|---|---|
| Revenue | +0.8% | +3.29% |
| Daily Unique Professional Interactions | — | +1.17% |
Low-connection users saw gains 3-4x larger than the overall population. The system that was supposed to improve the feed for everyone disproportionately helped the people who had the least to work with.
Why new accounts benefit more
The reason is structural, not accidental.
The old retrieval pipeline needed behavioral signals that new accounts hadn’t generated yet. The new one needs exactly one thing: a well-written profile. A clear headline, a specific summary, experience descriptions that use the natural vocabulary of your field. That’s enough for the LLaMA-3 model to generate a sharp embedding and start matching you with relevant content.
There’s a second system reinforcing this. LinkedIn’s ranking engine (a separate sequential transformer called the Generative Recommender) processes your last 1,000+ interactions to predict what you’ll engage with next. For accounts with a long history, the profile signal is one input among many. But for accounts with fewer than 10 interactions, a separate Qwen3 profile embedding delivers its largest measurable benefit: over 2% improvement in long-dwell prediction accuracy.
When you have no behavioral history, your profile does all the talking. And the new systems are actually good at listening.
What this means practically
The conventional advice to grind through a cold start period is based on a system that no longer exists. A new LinkedIn account with a precisely written profile now enters a feed environment where the algorithm can match it to relevant content from day one.
This doesn’t mean new accounts magically get massive reach. Distribution still depends on creating content that earns engagement, building a network worth engaging with, and showing up consistently. The algorithm can match you to the right content, but it can’t make people care about what you post.
What it does mean: the barrier to entry dropped significantly, and almost nobody has updated their strategy to reflect it.
The real cold start problem is profile quality
If the retrieval system generates your professional identity from your profile text, then profile optimization isn’t a nice-to-have. It’s the entire mechanism by which new accounts escape the cold start.
A vague headline like “Helping businesses grow” produces a blurry embedding that matches weakly with everything. A specific one like “B2B demand generation for developer tools” produces a sharp embedding that matches strongly with exactly the right content and audience.
For established accounts, engagement history compensates for a mediocre profile. For new accounts, nothing compensates. The profile is the embedding. The embedding is the retrieval. The retrieval is whether anyone sees you at all.
The cold start problem didn’t disappear. It just moved from “you need to build your network first” to “you need to write your profile well first.” One of those takes months. The other takes an afternoon.