For many mid-sized businesses LinkedIn is the most important organic B2B channel: around 19 million people in Germany use the network every month. That same channel changed fundamentally in 2026, because LinkedIn now ranks the feed with a large language model. It sounds like a technical detail. The consequences are anything but. Posts that visibly came straight out of an AI tool lose reach. Posts with a recognisable voice and a clear subject gain it.
Plenty of marketing leads are noticing this in an uncomfortable way: since they started producing content with AI, reach has gone down rather than up. This piece sets out what sits behind that technically, which signals actually count in 2026, and how to keep using AI without being penalised for it.
1. What changed in the LinkedIn feed in 2026
Until a few years ago the LinkedIn feed ran on a fairly mechanical logic: follower count, hashtags, likes, posting frequency. If you knew the rules, you could play them. LinkedIn has replaced exactly that mechanism, step by step, with a model that reads posts for their content rather than sorting them by counted values.
Three shifts matter for companies. First, topics beat reach. The feed assigns posts to topic clusters and shows them to people interested in that topic, rather than pushing them at your network wholesale. Second, depth beats frequency. A post that gets saved, or that draws a properly argued comment, weighs far more than ten likes. Third, LinkedIn is moving noticeably harder in 2026 against automated interaction and AI comment spam.
Put briefly: the 2026 feed does not assess how often you post. It assesses whether a person with a recognisable position is writing about a clearly defined subject. That is good news for specialists in mid-sized firms, and bad news for content production lines.
2. 360Brew: how LinkedIn's language model reads posts
Behind the rebuild sits a model LinkedIn has documented itself: 360Brew, described in the paper "360Brew: A Decoder-only Foundation Model for Personalized Ranking and Recommendation" (arXiv, January 2025) by the LinkedIn team around Hamed Firooz. It is a decoder-only model with roughly 150 billion parameters, trained on LinkedIn's own data.
The decisive difference from classic recommendation systems: instead of hundreds of hand-built numerical features, the model reads profiles, post text and user behaviour as language. It understands what a post is about — and whether that fits the author's existing topic profile. According to the publication, the model covers more than 30 prediction tasks: feed, job recommendations, search and ads.
In practice that means the feed recognises subject-matter proximity without you helping it along with hashtags. And it recognises just as readily when a text is formally clean but substantively interchangeable. The same logic has long since taken hold outside LinkedIn as well; the parallels are set out in our piece on E-E-A-T and AI authority.
3. The path of a post: check, test audience, expansion
LinkedIn does not communicate the details officially, but observations from practice have lined up for years. A post essentially runs through three stages:
- Quality check (seconds): automatic classification for spam, clickbait, pod patterns and missing relevance. Posts flagged here are throttled from the outset.
- Test audience (the first 60 to 90 minutes): the post is initially shown only to a small slice of the network. This phase decides everything that follows.
- Expansion (hours to days): on good signals the post runs into second- and third-degree contacts and into topically related feeds.
The practical consequence: the first hour matters more than the rest of the day. Publish in the morning and reply to comments only in the evening and you give away the window in which LinkedIn is measuring. For German-speaking B2B audiences, experience puts the most active slots on Wednesday and Thursday mornings.
A word of caution on engagement pods. Coordinated like-and-comment groups are conspicuously easy for a semantic model to spot: the same people, the same order, comments with no content in them. The short-term reach effect is increasingly bought at the price of lasting throttling.
4. Which signals actually count in 2026
The order of engagement signals has shifted markedly. The available industry analyses produce roughly this picture for 2026:
- Saves carry the most weight. Someone filing a post away for later is making the strongest possible statement about its value.
- Substantive comments come next — actual arguments, questions and objections, not one-word approval.
- Dwell time follows: how long someone stays with a post, including the click on "see more".
- Likes come last. A post with 40 substantive comments clearly beats one with 400 reactions.
For content planning that is a concrete instruction. Formats people want to keep — checklists, frameworks, step-by-step guides, comparison tables — are structurally better placed than snippets of opinion. And posts that ask a genuine question get more out of the test phase than those collecting agreement.
5. Why unedited AI text costs reach
This is the core of the problem, and it is not ideology but a consequence of the model architecture. A semantic ranking model compares a new post with a great many similar posts. Text that comes out of a language model without editing resembles thousands of others: the same sentence patterns, the same lists of three, the same closing question that leads nowhere. It carries no information that does not already exist, and it is ranked accordingly.
On top of that comes a second layer. LinkedIn tightened its action against bot activity and AI spam in comments in 2026, and is reworking how comments appear in the feed so that discussion beneath a post carries weight again. Automated commenting tools that dispense generic approval by the minute are therefore not merely ineffective — they are a risk to your own profile's visibility.
The four most common reach killers:
- External links in the post text. Analyses by Sprout Social put this at around 60 per cent less reach for 2026. The link belongs in the first comment — or the value belongs directly in the post.
- More than three hashtags. A model that understands language needs no keyword crutches. Hashtag cascades now read more like a spam signal.
- Unedited AI text: generic, interchangeable, with no figure or experience of its own — and therefore worthless to a semantic model.
- Topical randomness. Posting about recruitment today, sustainability tomorrow and tax law the day after builds no topic profile for the model to latch onto.
The answer is not less AI but a different division of labour. AI handles research, structure, variants and repurposing. The human supplies the specific figure from an actual project, the objection from a customer conversation, the position you also have to defend. How to organise that as a process is set out in our piece on the AI-supported editorial workflow for smaller companies.
6. Formats: what works in the 2026 B2B feed
Because dwell time and saves are weighted so heavily, the order of formats shifts too. Document posts — multi-page PDF carousels — lead the field, because people page through them and therefore measurably stay with the post longer. Posts whose main purpose is an external link bring up the rear.
These are guide values, not guarantees. The figures come from public analyses by various providers and vary considerably by industry and network size. As at August 2026.
Three format recommendations for mid-sized company presences. First, PDF carousels of five to twelve pages for guides and comparisons. Second, text posts of roughly 800 to 1,300 characters with a genuine hook in line one — the first two lines decide the click on "see more". Third, short videos under 90 seconds, with subtitles as a rule, because the feed is consumed largely without sound.
7. What this means for mid-market marketing
The good news first: semantic ranking favours subject-matter expertise, and there is plenty of that in mid-sized firms. A managing director who has been building special-purpose machinery for 15 years has more to say than any content tool. The problem is rarely the substance. It is the missing process for making it visible regularly.
Topic profile before topic variety
Settle on two or three core topics and stay with them for at least a quarter. The model needs repetition to form a profile. Three to five posts a week are enough; more tends to spread reach out rather than increase it.
People beat company pages
Personal profiles achieve structurally more organic reach in B2B than company pages. The most durable arrangement for a mid-sized firm: the company page carries facts, job ads and advertising, while reach comes through two or three people from the team writing about their own specialist field.
Use AI where it does not show
Idea generation, research, outlining, carousel layouts, subtitles, turning one blog article into five formats — that is where AI saves real hours without the feed penalising it. The technical side of distribution, and how a multi-channel set-up is built, are separate subjects with their own mechanics; what matters here is that the visible surface still carries a human signature.
8. A 60-minute check of your LinkedIn presence
If you would rather build LinkedIn systematically than post on instinct, arrange a free initial conversation. We look at your recent posts and your topic profile together.
9. Conclusion: the feed rewards substance again
The 2026 LinkedIn rebuild looks at first glance like a setback for AI-supported marketing. It is in fact a correction. A feed that understands language can recognise interchangeability, and it does. What gets penalised is not the use of AI but the delivery of unedited AI output.
For mid-sized companies that is more opportunity than risk. If you have genuine expertise, focus it on two or three topics and use AI to get it into good formats faster, you stand better in 2026 than any company with a high posting frequency and empty text. The work shifts away from producing text — AI does that — and towards choosing what should be said at all.
Our assessment: the decisive LinkedIn metric in 2026 is not the number of posts but the ratio of saves and substantive comments to impressions. Keep that ratio in view and you are steering the channel. Count posts and you are merely administering it.
Also worth reading: social media as a source in AI answers.
10. FAQ: the key questions on the 2026 LinkedIn algorithm
Does LinkedIn penalise AI-generated posts?
LinkedIn does not prohibit AI content. But the semantic ranking model recognises when a text is interchangeable and carries no information of its own, and such posts are distributed less widely. Enrich AI drafts with your own figures, examples and a clear position and you are at no disadvantage.
What is 360Brew?
360Brew is the decoder-only foundation model documented by LinkedIn, with roughly 150 billion parameters, described in an arXiv paper from January 2025. It reads profiles, posts and behaviour as language rather than as numerical features, and covers more than 30 prediction tasks — among them feed, search, job recommendations and ads.
Do external links really cost reach?
Analyses by Sprout Social put posts with an external link in the main text at around 60 per cent less reach for 2026. LinkedIn confirms no specific figure, but the direction is not seriously disputed in practice. The usual solution: put the link in the first comment and deliver the value in the post itself.
How often should a smaller company post on LinkedIn?
Three to five posts a week per profile is a realistic and effective range. A higher frequency tends to spread reach across more posts without raising overall visibility. More important than the number is topical consistency across several weeks.
