Every post you write now goes to two readers.
The first is a person. They open it, read it, decide whether it was worth their time. Nothing new there — we've been writing for that reader forever.
The second reader showed up recently, and it isn't a person. It's a model. It reads your text, works out what it's actually about, and decides whether to surface it or quote it in an answer. You're writing for both now. They want slightly different things — but less different than you'd fear.
LinkedIn is the easiest example because they published how it works. Their engineering post describes the new way content gets picked: instead of the old pile of separate pipelines, there's now an LLM retriever that turns every post and every profile into a vector of meaning, then pulls a couple thousand candidates out of millions by how close those meanings sit. That's the first point where a model reads your text. Before anything gets ranked, something has already worked out what you wrote about.
Here's the part almost every write-up got wrong. The retrieval runs on an LLM, yes. But the final ranking of those candidates isn't done by a language model — it's a separate transformer called Feed-SR. In their own paper, LinkedIn says plainly that they tried an LLM ranker and chose not to ship it: the transformer scored better and cost less to run. So "the feed is one giant LLM now" isn't what happened. The LLM sits exactly where it's decided whether your text gets considered at all — at the reading-for-meaning step.
There's one catch worth mentioning. Online, everyone calls this system "360Brew" and recites its exact rules. 360Brew is a pre-production research model from a separate paper whose every version on arXiv is marked withdrawn, and LinkedIn never confirmed it runs the feed. Their own Feed-SR paper contradicts the story directly. Somewhere along the way, someone attached a catchy name to a real engineering paper, and everyone else repeated it. If a source can't get the name of the engine right, treat its "exact rules" as a guess too. Read the primary source, not the recap of a recap.
The name isn't the important part. The shift is. The feed used to be gameable with mechanics — keywords, early-hour posting windows, pods trading likes. Now there's a model at the door that reads for meaning, and you can't fool it with the tricks that fooled a counter. It reads roughly the way an assistant reads a question you asked it.
That's where the two readers start looking surprisingly similar. The assistant looking for the answer to a professional question and the feed retriever deciding whether to surface your post are doing the same job: reading the text for whether there's a specific point about a specific thing. What lands with the first lands with the second. Neither gets much from smooth, empty writing; both reach for text where it's clear what it's about and who it's for.
This doesn't mean you should start writing for the algorithm. If anything, it points in the opposite direction. Writing for the machine used to mean writing worse: stuff in keywords, pad to length, staple on five hashtags. Now the machine reads like a reader, so writing for it means writing more clearly. Don't try to please the algorithm. Try to be understood. One concrete point beats ten blurry ones. An example someone can quote beats a paragraph of generalities. A topic you can be pinned to beats trying to appeal to everyone.
There is one trap, though. Optimizing to "get quoted by a model" is a slippery road, and at the end of it waits exactly the AI slop I spent last time learning to scrub out — the confident, weightless prose that says nothing. The irony is that the model at the door doesn't like it either; it tells specific from smoothly-empty about as well as a person does. Push that too far and you end up writing for neither. The only thing that seems to work for both is having something specific to say.
LinkedIn just happened to show how the system works first. Everything that retrieves and quotes you — search, assistants, feeds — is moving the same way, from "how many signals did you stack" to "what is this actually about." What people buried as SEO back in 2016 didn't die. Search just stopped counting and started reading.
So yes, you've got two readers now. But write honestly and to the point, and the second one isn't extra work. It wants the same thing the first one does.
Have you caught yourself writing with one eye on how a model will read it? Does the text come out better or worse for it?
Top comments (6)
The idea that we're now writing for both human and model readers, as you pointed out, highlights the importance of clarity in our writing. The fact that LinkedIn's Feed-SR transformer is responsible for the final ranking of candidates, rather than an LLM, is particularly interesting - it shows that while models are being used to understand the meaning of our text, traditional ranking algorithms are still being used to determine relevance. I've found that in my own writing, focusing on making concrete points and providing specific examples has not only improved my ability to communicate with human readers, but also seems to have improved the visibility of my content in feeds. Do you think this shift towards model-driven content ranking will ultimately lead to a more authentic and engaging online discourse, or are there other potential implications we should be considering?
Thanks — you picked up on the part most people gloss over.
That split is actually reassuring. The model figures out what the post is about, but the final ranking is still handled by a separate model. If one giant LLM did everything, people would inevitably start writing for the LLM itself. With the split, the model at the door is mostly answering a simpler question: "Is there actually a point here?" The easiest way to pass that test is to have one.
On the "more authentic" part, I'm a bit more cautious. Clear writing probably has an advantage now, but the other side of that is that people are already building tools that imitate substance. They produce text that looks specific without actually saying much. So I don't think better conversations happen automatically. It feels more like an arms race: retrieval models get better at telling substance from polished emptiness, while generators get better at making polished emptiness look like substance. My guess is that having a real point will stay the easier strategy, simply because it's less work than faking one.
The other thing I'd watch is homogenization. Ranking by meaning has a cost. Models surface things that are close to patterns they've already learned, which can quietly work against ideas that are genuinely unusual or don't resemble anything else yet. A simple like counter had plenty of flaws, but it didn't care whether an idea was familiar or weird. So the question isn't only whether clear writing gets rewarded now. It's also whether the range of ideas that gets surfaced slowly becomes narrower.
Have you noticed that narrowing in your own feed, or does it still feel wide open?
Great perspective, Eugeniya! I really like the point that the shift is not about writing for machines instead of humans — it is about making ideas clearer, more specific, and easier to understand.
The distinction between retrieval for meaning and ranking for relevance is especially interesting. As AI systems become better at understanding context, I think the advantage will move away from optimization tricks and toward having real experience, original insights, and concrete examples behind the content.
I have noticed the same pattern in AI engineering and developer communities: the posts that create the most meaningful discussions are usually the ones that explain a real problem, a real implementation, or a lesson learned from building something.
Would love to connect and exchange ideas around AI, agent workflows, content intelligence, and developer ecosystems. Happy to collaborate and share experiences as we explore where this space is going.
Thanks, glad it landed. You put your finger on the thing I keep circling back to — the posts that get people talking are the ones with an actual problem or a scar behind them. The rest just doesn't hold either reader's attention anymore. That's mostly what I'm here to write, so, more of it coming. Good to have you in the comments.
Thanks, Eugeniya! I completely agree with your point about the “scar behind the idea.” In my experience, the strongest technical content usually comes from solving a real problem, debugging a difficult issue, or learning something through building — not from trying to optimize for a ranking system.
The homogenization point is also important. As AI-driven retrieval improves, there is a risk that systems favor patterns that are already familiar because they are easier to classify. Preserving space for unconventional ideas, experiments, and early-stage thinking will be a key challenge for future content platforms.
I think the best balance comes from combining strong AI understanding with human curiosity. Models can help people discover relevant knowledge, but communities still need builders sharing original experiences, failures, trade-offs, and lessons that cannot be reduced to generic patterns.
I’m looking forward to seeing more of your work around AI workflows and content ecosystems. There are many interesting challenges ahead as AI changes how we create, discover, and share knowledge.
You're welcome :) I'll try to prepare more material on this topic.