# How the LinkedIn Algorithm Works in 2026 (From LinkedIn's Own Engineering)

**Quick answer:** LinkedIn's feed works in two steps. First it pulls candidate posts using LLM-generated embeddings that match what a post is about to what you care about, from inside and outside your network. Then a transformer ranker scores those posts against more than a thousand of your past interactions, with signals refreshed within minutes (LinkedIn Engineering, March 12, 2026).

I ghostwrote LinkedIn posts for about six years, for 50 to 60 founders, coaches and consultants, and those posts have passed 100 million views between them. A lot of that job was explaining why a post died, usually to a client holding a theory from some random guide on the internet: a spam-filter flowchart from years ago, and a "golden hour" nobody can source.

The crazy thing is, LinkedIn's engineers published how the new feed works in March 2026, and LinkedIn researchers have published two related models on arXiv about it.

But is that really how the LinkedIn algorithm works?

This guide sticks to those sources, labels what's confirmed and what's rumour, and tests the popular rules against 24,745 real posts from Oiti's database of live LinkedIn content.

**By [Aitijya Sarkar](https://www.linkedin.com/in/thewallandthem/)**, Founder - Oiti | LinkedIn ghostwriting | 100M+ views - Updated 2026-10-03

## How the LinkedIn algorithm works in 2026

The LinkedIn algorithm finds posts you might care about, ranks them by how likely you are to engage, and updates both as people react. LinkedIn's March 12, 2026 engineering post describes three parts.

- **Retrieval.** A post's text is turned into an embedding by a large language model, a set of numbers that captures what the post is about. The feed then pulls posts whose meaning sits close to your interests, whether or not you follow the author.
- **Ranking.** A transformer model LinkedIn calls a Generative Recommender reads your history in order, "more than a thousand" past interactions, and predicts what you'll do with each candidate: stay and read (a long dwell), like, comment or share.
- **Freshness.** Nearline pipelines refresh the embeddings and signals "within minutes", so the feed picks up new posts and new reactions fast.

![Flowchart of LinkedIn's 2026 feed: posts from your network, who you follow and suggested content pass through one retrieval step using LLM-generated embeddings, then a Generative Recommender transformer that reads over a thousand of your past interactions and predicts clicks, skips, long dwells, likes, comments and shares, then your feed. A dashed loop shows three nearline pipelines refreshing signals within minutes. A separate note says 360Brew, a 150B decoder-only research model, is not named in LinkedIn's post.](https://www.ghostwriting-ai.com/posts/diagrams/feed-pipeline.webp)

*Retrieval finds candidates by meaning, ranking orders them by your history, and nearline pipelines keep both fresh within minutes. 360Brew is research LinkedIn hasn't tied to the feed.*

Source: [LinkedIn Engineering: Engineering the next generation of LinkedIn's Feed (Mar 12, 2026)](https://www.linkedin.com/blog/engineering/feed/engineering-the-next-generation-of-linkedins-feed).

Two things follow for anyone posting.

Your followers aren't a fixed audience, because retrieval reaches past your network by topic. And each reader's history counts for more than anything you do the moment you hit Post.

## What's confirmed, what's research and what's rumour

Only a handful of claims about the LinkedIn algorithm come from LinkedIn itself. Most of what circulates is creator folklore. Here's where each popular claim comes from.

| Claim | Status | Source |
| --- | --- | --- |
| The feed retrieves posts using LLM-generated embeddings | Confirmed by LinkedIn | LinkedIn Engineering, Mar 12, 2026 |
| A transformer ranker reads 1,000+ of your past interactions | Confirmed by LinkedIn | LinkedIn Engineering, Mar 12, 2026 |
| The ranker predicts long dwells, likes, comments and shares | Confirmed by LinkedIn | LinkedIn Engineering, Mar 12, 2026 |
| Signals refresh within minutes | Confirmed by LinkedIn | LinkedIn Engineering, Mar 12, 2026 |
| LinkedIn fine-tuned Meta's LLaMA 3 (1B and 3B) for feed retrieval | Backed by research | arXiv 2510.14223 (Oct 2025) |
| 360Brew is a 150B-parameter, decoder-only ranking model | Backed by research (paper since withdrawn) | arXiv 2501.16450 (Jan 2025) |
| 360Brew is the model running your feed today | Creator rumour | Forbes (Jan 2026) and creator guides; LinkedIn hasn't said so |
| A 60 to 90 minute "golden hour" decides reach | Creator rumour | Repeated by SocialBee without a source; Buffer notes LinkedIn never confirmed it |
| A save is worth about 5x a like | Creator rumour (third-party study) | AuthoredUp, cited by Hootsuite |
| Posts get sorted into spam, low-quality, good and expert buckets | Creator rumour | SocialBee, no source given |
| Links in the post kill reach | Creator rumour; my data: 0.97x likes | Oiti analysis of 24,745 posts |
| Hashtags boost reach | Creator rumour; my data: no consistent effect | Oiti analysis of 24,745 posts |

Source: [LinkedIn Engineering (Mar 12, 2026)](https://www.linkedin.com/blog/engineering/feed/engineering-the-next-generation-of-linkedins-feed)

Other sources: [arXiv 2501.16450: 360Brew](https://arxiv.org/abs/2501.16450), [arXiv 2510.14223: Large Scale Retrieval for the LinkedIn Feed using Causal Language Models](https://arxiv.org/abs/2510.14223), [Forbes (Jan 12, 2026)](https://www.forbes.com/sites/jodiecook/2026/01/12/the-linkedin-algorithm-changed-again-heres-whats-new-for-2026/), [SocialBee](https://socialbee.com/blog/linkedin-algorithm-explained/), [Hootsuite](https://blog.hootsuite.com/linkedin-algorithm/).

## Is the LinkedIn algorithm called 360Brew?

LinkedIn has not named its feed algorithm, and its March 2026 engineering post never mentions 360Brew. The name comes from a LinkedIn research paper, and LinkedIn snake oil salesmen started attaching it to the live feed afterwards.

Three different things get lumped together under that name, so here's what each one is.

- **360Brew is a research model.** The paper (arXiv 2501.16450, January 2025) describes a 150-billion-parameter, decoder-only model for personalized ranking and recommendation, built by LinkedIn researchers and able to handle more than 30 prediction tasks with one model. Every version is now marked withdrawn, with an arXiv admin note saying the submitter didn't have the right to agree to the licence, and the abstract itself calls 360Brew a "research pre-production model". LinkedIn has never officially confirmed that it runs the production feed.
- **The LLaMA 3 retrieval model is a different paper.** That one (arXiv 2510.14223, October 2025) fine-tunes Meta's LLaMA 3, at 1 and 3 billion parameters, to create the embeddings the feed uses to retrieve suggested posts from outside your network. It's a much smaller model, and it isn't 360Brew.
- **The Generative Recommender is what LinkedIn has confirmed.** It's the ranking model LinkedIn's March 2026 engineering post describes, and the one LinkedIn has officially said ranks the live feed.

So what's it called? LinkedIn hasn't branded the system. "360Brew" is a research name creators borrowed, and until LinkedIn says otherwise, it's a research model, not the feed you scroll.

## What changed in the LinkedIn algorithm in 2026

The big change is that the feed now matches posts to readers by meaning, not only by who they're connected to. Older guides described a pipeline that filtered spam, tested a post on a few connections, then widened distribution if it did well. The March 2026 system retrieves posts by topic and ranks them against each reader's long history, updating within minutes.

For creators, three things shifted.

- **Topic consistency counts for more.** If your posts keep circling the same subjects, the embeddings get a clear picture of what you write about, and the feed can match you to people who read about it.
- **Reach beyond your followers is normal.** Out-of-network retrieval means a good post on a subject can find readers who have never heard of you.
- **Reading time is a target.** Long dwells are one of the outcomes the ranker predicts, so a post people read to the end earns something even without a like.

## I tested the popular rules on 24,745 posts

Most folk rules about the algorithm don't show up in the data: timing, hashtags, links and formatting style all landed within about 10% of normal. I compared each post with its own author's median, so big accounts don't swamp small ones. A score of 1.00 means "exactly as well as this author usually does."

| Popular rule | What my data showed | Likes vs author's median | Comments vs author's median |
| --- | --- | --- | --- |
| Links in the post hurt reach | No real effect | 0.97x | 0.96x |
| Use 3 to 5 hashtags | No consistent pattern; 86.5% of posts used none | 0.92 to 1.07x | 0.93 to 1.03x |
| Post on Tuesday to Thursday | No day stands out | 0.97 to 1.01x | not measured |
| Post at the right hour | Flat across the working day | 0.96 to 1.04x (05:00 to 18:59 UTC) | not measured |
| Write one sentence per line | No difference from paragraphs | 0.98 to 1.02x | 1.00 to 1.02x |
| End on a question | Likes flat, comments slightly up | 1.00x | 1.08x |
| Short posts do better | The opposite: under 300 characters did worse | 0.83x | 0.73x |
| Post as often as you can | A second post within 12 hours did worse | 0.84x | 0.85x |

Source: [Oiti analysis of 24,745 LinkedIn posts, 2026](https://www.ghostwriting-ai.com/posts/best-time-to-post-on-linkedin)

Two findings did hold up. Posts under 300 characters got 17% fewer likes and 27% fewer comments than their author's usual, so a post needs enough in it to react to. And posting again within 12 hours of your last post cost 16% of likes and 15% of comments.

I watched the short-post problem play out with a founder who wanted one-liners because "nobody reads on LinkedIn." The one-liners got a polite trickle of likes. Then he let me post a long story about letting his first hire go, which he nearly vetoed, and it got more comments than anything else he posted that year.

| Gap since previous post | Rel. likes | Rel. comments |
| --- | --- | --- |
| Under 12h (3,763 posts) | 0.84x | 0.85x |
| 12 to 24h (6,710 posts) | 1.00x | 1.00x |
| 1 to 3 days (9,271 posts) | 1.02x | 1.00x |
| 3+ days (4,544 posts) | 1.04x | 1.02x |

*Median likes and comments relative to the author's own median, grouped by hours since that author's previous post, with a reference line at 1.0.*

Baseline: 1.00x = Author's typical post.

Source: Oiti analysis of 24,745 LinkedIn posts

The 12-hour finding matches the one mistake I stopped making. A consultant I wrote for wanted two posts on launch day, one at 8am and one after lunch. The morning post was the better of the two, and it stalled about when the second one went up. Since then I've kept launches to one post a day.

Past 12 hours the gap stopped mattering: a post every day did as well as one every few days. My guide to the [best time to post on LinkedIn](https://www.ghostwriting-ai.com/posts/best-time-to-post-on-linkedin) has the full day and hour breakdown, and [LinkedIn hashtags](https://www.ghostwriting-ai.com/posts/linkedin-hashtags) covers the hashtag numbers in detail.

## LinkedIn algorithm myths, checked

Most of the rules people repeat about the LinkedIn algorithm have no source behind them. The five you'll hear most:

- **The golden hour.** LinkedIn has never confirmed that your first 60 to 90 minutes decide everything. Chris Donnelly, whose team studied 300,000 posts, told Forbes "there has never been a golden hour". LinkedIn's own post has signals refreshing within minutes and ranking tied to each reader's history, so a slow start isn't fatal. A fintech founder I wrote for blocked out the hour after every post to answer comments, and once had me move a post because he had a dentist appointment. Then a Friday-afternoon post of his did almost nothing all evening and turned into his best post of the quarter by the next Wednesday. He still blocks the hour. I've just stopped arguing.
- **Hashtags drive reach.** In my data, they showed no consistent effect by count or position. LinkedIn also removed hashtags from profiles in February 2024. Use one or two if they help a reader find the topic, and expect nothing more. A leadership coach I worked with ended every post with ten hashtags, a leftover from Instagram. During a quiet month we dropped them to see what would break. Nothing broke, and she never asked for them back.
- **Links get buried.** Posts with a link got 0.97x their author's usual likes. If there's a penalty, it's a few percent. Nothing like the collapse people describe. A B2B SaaS founder I wrote for in 2021 wouldn't move links into the comments because he thought it looked shady. His link posts did roughly what his other posts did. The flops were the ones that read like an ad, and those flopped with or without a link.
- **Every post is sorted into four quality buckets.** Nobody cites a source for this one, and LinkedIn's engineering post talks about predicted engagement, with no buckets anywhere.
- **Saves beat everything.** The "5x a like" figure comes from a third-party study (AuthoredUp), not from LinkedIn. The engineering post lists long dwells, likes, comments and shares as ranking targets and never ranks one above the others.

## How to write for the 2026 LinkedIn algorithm

Write about a few subjects, make posts worth finishing, and give people something to reply to. The ranker is built to spot exactly that.

Across six years of client work, the accounts that grew all did the same unglamorous thing: they picked two or three topics and kept going. One client, a CFO who wanted to write about fundraising, hiring and her marathon training, stalled for months. When she narrowed to cash-flow mistakes she'd seen in early-stage startups, her posts started reaching finance people she'd never met. That's topic retrieval at work.

- **Stick to your subjects.** Retrieval matches posts to readers by meaning. A clear topic gives it something to match.
- **Earn the "see more".** A first line that gets people to expand the post buys you reading time. The [LinkedIn hook generator](https://www.ghostwriting-ai.com/tools/linkedin-hook-generator) gives you options for free.
- **Don't go too short.** Under 300 characters, posts underperformed on both likes and comments. [LinkedIn character limits](https://www.ghostwriting-ai.com/posts/linkedin-character-limits) has the length data.
- **Invite a reply.** Posts ending on a real question got 8% more comments.
- **Leave 12 hours between posts.** Schedule the second idea for tomorrow; [how to schedule a LinkedIn post](https://www.ghostwriting-ai.com/posts/how-to-schedule-a-linkedin-post) shows how.

## How to check whether the algorithm is showing your posts

1. **Open a post's analytics** Click the impressions count under any post to see who it reached.
2. **Compare impressions with your usual** One post is noise. Look at your median over the last month and judge new posts against it.
3. **Read who replied, not just how many** Comments from people in your field mean the feed matched your topic to the right readers.
4. **Change one thing at a time** Try longer posts, or a tighter topic, for a few weeks before judging.

For what counts as a normal number, see my guide to [LinkedIn impressions](https://www.ghostwriting-ai.com/posts/linkedin-impressions).

## The cheat sheet

![Cheat sheet sorting LinkedIn algorithm claims. Confirmed by LinkedIn Engineering (Mar 12, 2026): LLM-embedding retrieval, a transformer ranker reading 1,000+ past interactions, predicted long dwells, likes, comments and shares, signals refreshed within minutes. Backed by research: LLaMA 3 (1B and 3B) retrieval (arXiv 2510.14223) and 360Brew, a 150B decoder-only model (arXiv 2501.16450, withdrawn). Creator rumour: 360Brew runs the feed, a 60 to 90 minute golden hour, saves worth 5x a like, quality buckets, links kill reach (my data: 0.97x likes), hashtags boost reach (no consistent effect).](https://www.ghostwriting-ai.com/posts/infographics/linkedin-algorithm.webp)

*What LinkedIn has confirmed about its feed, what comes from research papers, and what is creator rumour.*

## The bottom line on the LinkedIn algorithm

Without the folklore, the LinkedIn algorithm does two jobs. It finds posts whose subject matches what a reader cares about, from inside or outside your network. Then it ranks them using more than a thousand of that reader's past interactions, guessing whether they'll read, like, comment or share. That's what LinkedIn has published so far.

What moves reach is boring.

Stay on a few subjects so retrieval knows who to show you to. Put enough in each post to be worth finishing: anything under 300 characters got 17% fewer likes and 27% fewer comments in my data. Leave 12 hours between posts, and end on a real question when you have one. LinkedIn has never confirmed a golden hour, hashtags showed no consistent effect, and links cost a few percent at most, so I'd stop worrying about all three. If you still want a posting slot, my guide to the [best time to post on LinkedIn](https://www.ghostwriting-ai.com/posts/best-time-to-post-on-linkedin) has the hours.

The client I think of most here was a founder with a spreadsheet of algorithm rules: post at 8:07, three hashtags, link in the first comment, reply within the hour. His reach barely moved whether he followed the sheet or ignored it. It moved the month he stopped writing about everything and wrote only about the hiring mistakes he'd made.

As far as I can tell, the feed had no opinion on his rules. It just finally knew what he was about.

## Frequently asked questions

### What is the LinkedIn algorithm?

It's the system that picks which posts appear in your feed and in what order. In 2026 it retrieves posts using LLM-generated embeddings that match topics to your interests, then ranks them with a transformer model that reads more than a thousand of your past interactions, according to LinkedIn's engineering team.

### What is the LinkedIn algorithm called?

LinkedIn hasn't given it a public name. Its engineers call the ranking model a Generative Recommender. "360Brew" is the name of a 150-billion-parameter research model from a LinkedIn paper (arXiv 2501.16450, now withdrawn), and LinkedIn has not said it runs the live feed.

### How does the LinkedIn algorithm work in 2026?

In two steps. Retrieval finds posts whose meaning matches your interests, from inside and outside your network. Ranking predicts whether you'll read at length, like, comment or share, based on your history, and those signals refresh within minutes.

### What has happened to the LinkedIn algorithm?

LinkedIn rebuilt the feed around large language models. Its March 12, 2026 engineering post describes retrieval built on LLM embeddings and a transformer ranker that reads each member's long history, a long way from the connection-first model older guides still describe.

### Is there a golden hour on LinkedIn?

Not one LinkedIn has ever confirmed. The 60 to 90 minute rule is repeated widely without a source, and LinkedIn's own description of the feed has signals refreshing within minutes and ranking tied to each reader's history, not a single launch window.

### Do hashtags still work on LinkedIn?

Not in any way I could measure. In my analysis of 24,745 posts, every hashtag count sat within about 8% of the author's usual likes, and 86.5% of posts used none.

### Do links hurt reach on LinkedIn?

Barely. Posts with a link in the body got 0.97x their author's usual likes and 0.96x their usual comments, a few percent at most.

### Is 10,000 impressions a lot on LinkedIn?

Yes. Across a 6,651-post dataset, the median post got 581 impressions and only one post in ten passed 13,551, so 10,000 puts a post near the top tenth. My guide to [LinkedIn impressions](https://www.ghostwriting-ai.com/posts/linkedin-impressions) has the full benchmarks.

## The algorithm rewards a clear topic, week after week

Oiti is the AI clone for LinkedIn content. It learns what you believe and the way you actually talk, then drafts posts on your subjects that read like you typed them, so staying consistent stops depending on how busy your week is.

**[Start free](https://tryapp.ghostwriting-ai.com/sign-up)** - 7 days free.

## More on how LinkedIn works

- **Guide — [Best time to post on LinkedIn](https://www.ghostwriting-ai.com/posts/best-time-to-post-on-linkedin)**: What 24,745 posts say about days, hours and posting gaps
- **Guide — [LinkedIn impressions](https://www.ghostwriting-ai.com/posts/linkedin-impressions)**: What counts as an impression, and what a normal number looks like
- **Guide — [LinkedIn hashtags](https://www.ghostwriting-ai.com/posts/linkedin-hashtags)**: How many to use, and why the count barely matters
- **Free tool — [LinkedIn post generator](https://www.ghostwriting-ai.com/tools/linkedin-post-generator)**: Draft a post on your topic in a minute