LinkedIn rebuilt its feed with AI. What should you post now?

On 12 March 2026 LinkedIn's engineering team published a description of how it had rebuilt the feed, and the short version is that both halves of the machine were replaced. Finding candidate posts is now done by a language model that reads meaning, and ordering them is now done by a model that reads your last thousand interactions as a sequence rather than judging each post on its own. It is the most consequential change to that platform in years, and almost nothing written about it since has quoted the actual post.

We have written before about the model specific side of LinkedIn, in Copilot prompts for LinkedIn posts. This is the other half of that story: not how to write the post, but what the machine on the other side is now doing with it.

What LinkedIn actually said changed

A feed is two problems. First, out of hundreds of millions of posts, find a few thousand that could plausibly interest this person. That is retrieval. Second, put those few thousand in an order. That is ranking. LinkedIn replaced both, and they were replaced for different reasons.

The embedding idea is worth a plain explanation, because it is what everything else follows from. Every post, and every member profile, gets converted into a long list of numbers that represents what it is about. Two things that mean similar things end up near each other in that space, even when they share no words at all. LinkedIn's own example is that an electrical engineer interested in small modular reactors can be matched to relevant posts without exact keyword overlap.

LLM-generated embeddings capture deeper professional interests beyond keyword matching.

The ranking half is the less discussed and arguably more interesting change. The old system scored each candidate post independently, which LinkedIn's engineers say missed the sequential patterns in how professionals consume content. The replacement treats your feed history as a sequence, and the post is explicit about the scale: more than a thousand past interactions, used to model how your interests move over time rather than what you clicked this morning.

Why this is good news if you have four hundred followers

For most of LinkedIn's history, reach was a function of your connection graph. You posted, your first degree connections saw some of it, and if enough of them engaged quickly it leaked into second degree feeds. A small business had a hard ceiling that no amount of quality could raise, because the ceiling was the size of the network, not the merit of the post. It is the same ceiling that made hashtags feel necessary on other platforms, and it has come down for the same underlying reason.

A retrieval layer that works on meaning does not have that constraint in the same way. LinkedIn's description is explicit that the system surfaces relevant content from outside a member's immediate network. In other words, the question moved from who do you know to what is this about and who is that for. That is a structurally better deal for a specialist with a small following, and a structurally worse one for a large account posting generalities.

What it means for what you write

Three practical consequences, in descending order of how confident we are about them.

• Be about one recognisable thing. Semantic retrieval works by proximity, and a post that is about pricing, hiring and a conference in the same six paragraphs sits in the middle of nowhere, near nobody. This is the closest thing to a direct implication the post supports.
• Be about it for months, not for a week. Ranking now reads your history as a trajectory. Consistency of subject is not a discipline thing any more, it is a legibility thing. We have made the general case for consistency over volume elsewhere and this is a specific reason it holds here.
• Stop optimising for words. Repeating a phrase four times to make sure the algorithm catches it was always slightly silly, and against a system that embeds meaning it does nothing except make you sound like a press release. Say the thing once, clearly.

There is a fourth that we hold more loosely. If the ranking model is trained on engagement quality rather than raw likelihood of a like, then the comment-bait post, the one asking you to type YES to receive a template, has a worse time of it. That is a reasonable reading and it is also the reading that everybody wants to be true. LinkedIn does not say it. We are not going to pretend otherwise.

What we cannot tell you

The engineering post is about infrastructure. It talks about retrieval latency, training speed and recall improvements, because it was written for engineers by engineers. It does not contain a creator playbook, it does not name a single content format, and it does not say what gets suppressed.

So there are things that are being confidently asserted across the internet right now which have no source. Whether external links are penalised. Whether the first hour of engagement still determines everything. Whether video is favoured. Whether posting on a Tuesday morning matters. Every one of those may be true and none of them is in the document. The reason we are labouring this is that a rebuilt system is exactly the moment when old folklore gets recycled with a new date on it, and a small business acting on folklore spends real hours on it. Instagram publishes far more than LinkedIn does, and we read all nine of its ranking system cards, including the one where Meta says an LLM scores your post.

What to do this month

• Write down, in one sentence, the subject your LinkedIn account is about. If you cannot, that is the finding, and what should a small business post about is the piece to read before the rest of this list.
• Look at your last ten posts and count how many are about that subject. Six or fewer means the account is not legible to a system that models interests over time.
• Rewrite your profile headline and about section in plain description rather than in slogans. They are embedded too, and a slogan means nothing in a semantic space. It is the same edit, for the same reason, as the name field on an Instagram profile.
• Drop the keyword repetition. Say it once, in the words you would actually use, which is easier if you have written your voice down somewhere.
• Give it a quarter before concluding anything. A ranking system that reads a thousand interactions is not going to reassess you in a fortnight.

None of that is exotic, and that is rather the point. If the profile behind the posts is still a default one, the headline is the field to fix first, and the examples in circulation are measurably the wrong length. The rebuild did not introduce a new trick to exploit. It removed some of the value of the old tricks and increased the value of being clearly about something, which is the advice that was already true and is now also mechanically true. The same shift is happening in the assistants, where whether you get named at all turns on being describable in one sentence. If the problem is that being clearly about something every week is more writing than you have time for, that is a different problem, and we have written honestly about whether a tool is worth it for a small business.