For months, our company posts have been pulling 75–80% less reach than they were, and personal posts swing from solid to invisible with no obvious pattern. We’re not alone. AuthoredUp’s analysis of three million-plus posts found 98% of users saw reach decline year-on-year, with median impressions falling around 47%. For larger accounts (50K+ followers) the drops hit 62% on reach and 83% on follower growth. Company pages overall have lost 60–66% of their reach.
This is not a glitch. It’s a deliberate rebuild.
Several things stopped working at once. Polls inflated impressions but the impressions didn’t convert. External links started costing around 60% of reach. Generic “thrilled to announce” company posts went from middling to invisible. AI-written content that worked fine six months ago started disappearing into the void. The frustrating part is that none of this was announced clearly. It rolled out in waves through 2025 and into 2026. We only got real clarity in March.
What LinkedIn finally told us
On 12 March 2026, LinkedIn’s engineering team published “Engineering the next generation of LinkedIn’s Feed” by Hristo Danchev — the most detailed disclosure they’ve ever made. The short version: they’ve replaced thousands of specialised ranking models with a unified LLM-based system, drawing heavily on their 360Brew foundation model (150B parameters, decoder-only, trained on LinkedIn data).
The mechanics matter. Stage one is an LLM-powered retrieval layer that narrows millions of posts to about 2,000 using semantic embeddings. Stage two is a Generative Recommender that processes 1,000+ of your recent interactions as a chronological sequence to rank what you actually see.
In English: the system now reads content, understands topics, and ranks per-recipient based on their professional trajectory. Not per-audience based on network size.
That explains both halves of the problem. Company posts collapsed because follower count no longer guarantees distribution. Personal posts swing wildly because distribution is individualised – the same post lands differently for every viewer based on their own history.
What’s not officially confirmed, despite what most commentary implies: that 360Brew specifically powers the live feed (LinkedIn’s internal notes reportedly say “The LLM-Ranker Was Evaluated and Rejected”); that AI content is actively penalised by name; that specific signals like saves are weighted 5x likes. Chris Essey’s analysis is the cleanest piece on what’s actually published versus what creators have inferred. Worth keeping the distinction when briefing internally — a lot of agencies are stating speculation as fact.
What to actually do, by sector
The structural shifts apply everywhere first: company pages are credibility archives now, not distribution channels. People distribute. Document carousels published natively beat link-out posts roughly 3x. External links cost ~60% reach. Polls and engagement pods are dead or actively penalised.
Manufacturing. This sector actually benefits from the rebuild. Engineers, plant managers, and procurement are exactly the niche, hard-to-reach audiences the new system handles better than the old one. The play is technical depth — capability documents, failure-rate case studies, certification posts — published natively from the company page, paired with Thought Leader Ads from real technical leaders (CTOs, heads of engineering, plant managers), which deliver around 1.7x CTR and 40% lower CPL than corporate ads. Don’t let the marketing team ghostwrite — the model can tell.
Professional services. Counter-intuitively, the best-positioned of the three. The algorithm now favours exactly what consultancies, law firms, and accountancy practices already sell: credibility, considered judgement, expertise earned through work. IR Global’s read is right — write clearly, share real insight from live matters, encourage discussion. The trap is the “evergreen content” programme: LinkedIn is downranking recycled thought leadership that adds no new perspective. Partners writing specifically about live cases and regulatory shifts will beat polished firm content every time. Pick 2–3 topical lanes per person and stay in them so the model can classify expertise.
Tech. Hit hardest because it was furthest into AI-written content and has the most 50K+ accounts (which lost the most reach). The recovery move is founder-led — point-of-view content, lessons from actual experience, honest commentary on what’s failing. The most common mistake is having comms write the CEO’s posts; the model is good at spotting the gap between profile identity and post voice. For wider employee advocacy, give frameworks not scripts — handed-down corporate copy is the most reliably punished tactic.
What changes when you commit to it
Three things shift relatively quickly once you reorient.
- The reach metric stops being meaningful and starts looking misleading. Saves, substantive comments, profile visits, and inbound DMs are closer to what the algorithm itself optimises for, and closer to commercial signal. Reporting these instead tends to improve leadership conversations overnight.
- Variance becomes normal on personal content. A post will reach 800 of exactly the right people instead of 8,000 of mostly wrong ones. That’s the system working, not failing. Stop A/B testing posting times. The gain is marginal compared to the gain from topical consistency.
- The company page becomes useful again, just for a different job. It’s the credibility archive buyers check during due diligence, not the reach engine. Treat it that way and the lower numbers stop feeling like failure.
The honest takeaway
The change is real, it’s structural, and it’s not reversing. That’s annoying, particularly because most of us built the previous strategy on a clear understanding of how reach worked. The good news: the new system is more honest. It rewards genuine expertise, topical consistency, and content that actually delivers value to the people who see it — which is what we’d all claim to have been doing anyway.
The work now is dropping the tactics that became reflexes — engagement pods, link-stuffed posts, polished corporate ghostwriting, vague “thought leadership” recycling — and building the things that work under semantic ranking: specific people with consistent topics, native long-form content, real points of view, and patient measurement of the metrics that actually map to pipeline.
It’s a worse year for the LinkedIn dashboard. It might turn out to be a better one for the LinkedIn pipeline.