LinkedIn's algorithm now favours relevance, not just follower count

LinkedIn now shows you posts from strangers more often than it shows you posts from people you follow, and it has built a model specifically to decide which strangers deserve the slot. Social Media Today reports that LinkedIn's new "Profile Relevance" system uses large-language-model-generated embeddings, mathematical representations of what a post is actually about, matched against embeddings of what a given user cares about, to surface content from outside a member's network. The company frames this as an improvement to recommendations. It is more accurate to call it a demotion of the follow graph.

The follower count was already a vanity metric. Now it's barely a signal.

For a decade, the operating theory of LinkedIn growth was straightforward: accumulate followers, post consistently, and your reach compounds because the people most likely to engage are already sitting in your audience. That theory required the feed to privilege the social graph, showing people what their network posted. Profile Relevance inverts the logic. The embeddings model is comparing the semantic content of a post, its topic, tone, and claims, against a model of a user's interests built from their behaviour, not their connections. Whether the poster is a first-degree connection becomes one input among many, not the gating factor.

This matters more for institutional accounts than personal ones, because institutional accounts have historically leaned hardest on follower accumulation as a proxy for authority. A multilateral agency with 400,000 followers built that number over years through partnership announcements and conference recaps. Under a relevance-weighted system, that number buys less automatic distribution than it used to. What buys distribution is whether a specific post's content maps closely enough to what a specific reader already demonstrably cares about, based on what they've engaged with, not who they follow.

The mechanism rewards specificity and punishes the generic update.

Embeddings work by clustering meaning. A post about Basel III capital requirements sits in a different semantic neighbourhood than a vague post about "navigating regulatory change." The first has sharp enough content to match confidently against users who've engaged with capital-adequacy content before. The second is too diffuse to match against anything with confidence, so the model has less reason to push it to anyone outside the existing follower base. This is bad news for the corporate-communications habit of writing posts vague enough to seem relevant to everyone. Under a relevance model, vague is invisible. Specific is findable.

That has a direct implication for financial-services and industrial communicators, who tend to default to the broad, risk-averse register precisely because it feels safer than taking a position. The safer post is now also the less distributed one. A post from an industrial group's sustainability lead that names a specific regulation, a specific plant retrofit, or a specific emissions figure gives the embeddings model something concrete to match. A post that says the company is "committed to a sustainable future" gives it nothing.

Reach outside the network was already happening quietly.

LinkedIn has been expanding out-of-network distribution for several years, first through the "suggested" posts that interrupt the chronological-ish feed, then through more aggressive surfacing of creator content to non-followers. Profile Relevance is the infrastructure that makes that expansion defensible at scale: instead of guessing broadly at what might interest a user, LinkedIn can now justify showing a finance executive's post to a reader who has never followed them, because the model has evidence the content matches that reader's demonstrated interests. The practical effect for anyone building a personal brand on the platform is that the addressable audience for any given post is no longer capped by who clicked "follow." It is capped by how precisely the post's subject matter can be modelled and matched.

That is a genuine opening for institutions that have been slow to build large followings but have deep, narrow expertise: standard-setting bodies, central-bank economists, policy researchers at philanthropic foundations. Their content was always too technical for mass reach under a follower-driven feed. Under a relevance-driven one, technical is exactly what the model is built to find a home for, provided the audience exists and has shown interest in adjacent material.

The practical shift for anyone posting on behalf of a bank, a UN agency, or an industrial conglomerate is to stop optimising for broad appeal and start optimising for legible expertise. Write the post that only makes sense to people who already care about the subject. The model's job is now to go find them.