The algorithm files · plain English, no jargon

LinkedIn replaced its algorithm with a 150-billion-parameter AI. Here's what that means for your reach.

Ezekiah Livingston LinkedIn strategist. I run LinkedIn end to end for B2B and deep-tech founders, and I watch this algorithm for a living. Find me on LinkedIn.

Published July 5, 2026 · Updated September 5, 2026 with the AI-slop crackdown numbers and the "report as slop" button.

So what actually happened?

For years, LinkedIn's feed ran on a pile of separate machine-learning models. One scored your post, another guessed what you'd click, another handled job recommendations, and so on. Each one leaned on crude signals: IDs, tags, engagement counters.

Then LinkedIn's research team built 360Brew: a single foundation model with about 150 billion parameters, trained on LinkedIn's own data, that handles more than 30 of those ranking and recommendation tasks by itself. Confirmed That's not creator gossip, it's from LinkedIn's own research paper, published January 2025.

The part that changes everything: 360Brew doesn't look up your post in a spreadsheet of signals. It reads it. Like a language model, because it is one. It reads your profile, your history, your post, and the relationship between all three, then asks a very human question: "Is this person credible on this topic, and who would genuinely care about this?"

In other words, LinkedIn hired a 150-billion-parameter editor. And the editor reads everything.

What changed for your reach

Here's where I have to be honest with you about what's confirmed and what's the community reading tea leaves. Anyone who tells you they know the exact weights is selling something.

Reach dropped, hard, for most people Observed

Independent trackers and a lot of very grumpy creators reported organic reach falling somewhere around 40 to 50 percent year over year after the rollout. LinkedIn hasn't published official numbers. What's consistent across every report I trust: the drop wasn't evenly distributed. Broad, play-it-safe content fell off a cliff. Specific, opinionated, niche content held its ground, and in plenty of cases grew. If you're trying to figure out which of several possible causes applies to your own account, run the reach-drop diagnostic.

The model reads meaning, not hashtags Confirmed

Because 360Brew processes text semantically, the old tricks aged badly overnight. Keyword stuffing, hashtag salads, "agree?" bait: the model sees through them the way you see through a bad sales email. It's judging whether your post actually says something to a definable audience.

Your profile is now part of your content Confirmed

This is the one almost everyone misses. The paper describes building ranking decisions from textual prompts that include member profile information and interaction history. Your profile isn't a static business card anymore. It's context the model uses to decide whether your post on, say, vector databases comes from someone credible or someone cosplaying. A vague profile quietly taxes every single post you publish.

Depth signals beat vanity signals Observed

Creator-side testing keeps pointing the same direction: saves and long, substantive comments correlate with distribution far more than reactions do. Some trackers put a save at several times the weight of a like. Take the exact multipliers with salt, but the direction fits how the model works. A save says "this has lasting value." A three-sentence comment says "a real person actually engaged with the idea." A thumbs-up says almost nothing.

Where things stand right now

This story didn't stop at the rollout. 2026 turned the model's quiet judgment into open policy: LinkedIn went from reading generic content to publicly punishing it. The sequence so far:

• January 2025LinkedIn publishes the 360Brew paper: one 150B-parameter model replacing dozens of task-specific rankers. Confirmed
• Through mid-2026Reach reshuffles as the new stack settles. Trackers put the median decline around 40-50% year over year, concentrated in generic content. Observed
• May 20, 2026LinkedIn announces changes explicitly targeting low-quality AI-generated posts, comments, and automation tools. Flagged content isn't removed; it's suppressed beyond your immediate network. Confirmed
• July 30, 2026Every post in the feed gets a report option: "Seems like AI slop." Distribution of a flagged-and-confirmed post drops platform-wide. Confirmed
• September 2026Press reports put the slop button past a million uses, and flagged content's views down roughly 40% within weeks. Observed

The AI-slop crackdown, with the actual numbers

If you've heard someone claim LinkedIn "killed 90% of AI posts," that's the telephone-game version. The real, sourceable numbers are less tidy and more interesting:

The mechanics matter more than the numbers. Nothing gets deleted, and AI-assisted writing isn't banned. What happens is quieter: content the system classifies as slop keeps reaching your immediate network and stops traveling beyond it. Your mother still sees it. The stranger who might have hired you doesn't. Most people flagged this way never find out, because a suppressed post looks exactly like a mediocre one.

How does it decide? LinkedIn hasn't published the classifier, but the reported behavior lines up with what 360Brew was already doing: reading structure and meaning. The tells are the pattern, not the tool: uniform sentence rhythm, tidy setups, rhetorical questions, zero verifiable detail. Which means the same post fails with both audiences, the model and the humans, for the same reason:

The slop-proofing test, before you post: does it contain one detail only you could know (a number, a name, a cost)? Does it take a position a reasonable peer could dispute? Would you say the first line out loud to a colleague without wincing? Three yeses and no classifier on earth files you under slop, because you aren't. The fastest fixes live in the hook-writing guide: the opening line is where both the model and the humans decide.

How to survive it (the playbook)

None of this requires posting daily or dancing for the machine. It requires being legible to a model that reads like a person. Here's what I'm doing for my clients and my own profile:

1. Pick a lane and repeat yourself. 360Brew builds a picture of what you're credible on from everything you've written. Five posts on one theme build that picture. Five posts on five themes build nothing. Being "known for one thing" is now a literal, mechanical input to your reach.
2. Make your profile agree with your posts. If you post about AI infrastructure but your headline says "Helping brands tell stories," the model sees a mismatch, and so does every human who clicks through. Fix the headline, the About, the Experience framing. Same story everywhere. (This is exactly what my profile optimization work exists for.)
3. Write posts worth saving, not just liking. Frameworks, hard-won numbers, checklists, contrarian takes with receipts. Before publishing, ask: would anyone bookmark this to use later? If the honest answer is no, it's a comment, not a post. Document posts are the easiest format for earning saves; the carousel guide covers how to build one that gets read.
4. Earn real comments, ignore reaction counts. End with a question only your actual audience can answer, take a position someone can push back on, and reply like a human when they do. Ten substantive comments beat two hundred silent likes.
5. Kill the engagement bait. Pods, "comment YES for the guide," tag-a-friend chains. The old algorithm was fooled by these. The new one reads them for what they are, and the people you actually want to reach always could.
6. Consistency over volume. One sharp post a week, on your lane, beats daily filler. The model has a long memory of your text. Filler dilutes your topical identity; the exact thing your reach now depends on. One of my clients grew from roughly 10 views a month to over a million total impressions posting once a week. The frequency wasn't the trick. The consistency of the story was.

What's confirmed vs. what's speculation

Quick sanity check, because the LinkedIn-guru industrial complex is having a moment with this update:

Confirmed by LinkedIn's paper: 360Brew exists, it's a ~150B-parameter decoder-only foundation model trained on LinkedIn data, it replaced many task-specific ranking models, it works from text (profiles, activity, content), and it generalizes to tasks it wasn't specifically trained on.

Confirmed by LinkedIn's 2026 announcements: the May 20 policy targeting low-quality AI posts, comments, and automation tools; suppression-not-removal as the penalty; the "Seems like AI slop" report option on every post since July 30; and the company's own 94% detection-accuracy claim, offered without a disclosed false-positive rate.

Observed by the community, not confirmed: the exact size of the reach decline, the ~40% view drop on flagged content, the million-plus report tally, precise signal weights ("saves are 5x likes"), and any claim about exactly when your specific post gets shown. Directionally useful, numerically fuzzy.

Safe to ignore: anyone selling a "360Brew hack," and any tool promising to make AI drafts "undetectable." You can't hack a model that reads meaning. You can only mean something.

The uncomfortable takeaway

Every era of this platform rewarded a skill. The last one rewarded gaming distribution. This one rewards actually being someone specific: a clear profile, a consistent point of view, and posts with enough substance that professionals keep them.

Which is annoying, because that's harder than hacks. And also good, because it's harder than hacks. The bar went up, and most of your competition won't clear it.

Questions people keep asking me

Primary source: 360Brew: A Decoder-only Foundation Model for Personalized Ranking and Recommendation (LinkedIn AI, arXiv, Jan 2025). Crackdown reporting: TechCrunch (July 30, 2026, the "Seems like AI slop" report option), The Next Web (the 94% detection-accuracy claim), Forbes (Aug 10, 2026, creator guidance), and aggregated press coverage of the May 20, 2026 announcement. AI-share estimates: Pangram and Originality.AI studies, which disagree (40% vs 81% of long-form posts) and are both labeled Observed. Reach observations aggregated from public creator reports and independent algorithm trackers; they are labeled "Observed" above because LinkedIn has not confirmed them.