Naano Research · October 2026

The LinkedIn and X algorithms, decoded.

What decides who sees a post in 2026, read from the platforms' own engineering papers and open-source code, then held against the sponsored posts we track at Naano.

A report by Naano · 15-minute read

20K+
LinkedIn posts analysed by our writing AI
2,301
sponsored posts studied
43,984
public reactions and comments studied
1:2
comments to reactions, median per post

01

Key findings

  1. 01

    Both feeds now rank with large transformer models that read a member's history as a sequence. LinkedIn's Generative Recommender reads more than 1,000 past interactions; X's Phoenix model is built on xAI's Grok transformer.

  2. 02

    Neither platform optimises for likes alone. LinkedIn predicts clicks, skips and long dwells next to likes, comments and shares. X scores each post as a weighted sum of more than fifteen predicted actions, negative ones included.

  3. 03

    Attention is measured before engagement. LinkedIn scores down posts it predicts will be skipped; on X, scrolling past without dwelling is one of the negative signals.

  4. 04

    Topic beats audience size. LinkedIn retrieves posts by matching their meaning with a member's profile and history, and X embeds viewers and posts as vectors. The right creator for the right audience matters more than the biggest audience.

  5. 05

    93% of the most-engaged sponsored posts on Naano don't name the brand in their first line, and 68% open with a hook of 80 characters or fewer.

  6. 06

    Only 3% of those hooks are questions. The posts that travel state something.

02

LinkedIn: how the feed ranks posts in 2026

In March 2026, LinkedIn's engineering team described a rebuilt feed. Several separate retrieval sources, from a chronological network index to trending and collaborative-filtering systems, were replaced by one retrieval model: a fine-tuned large language model used as a dual encoder. It turns each member and each post into an embedding and matches them by meaning, including posts from people the member doesn't follow.

What the model reads is text. For a post: its format, its author, its engagement counts and the post text itself. For a member: profile, skills, work history, education and the posts they engaged with. Engagement counts are converted into percentile buckets before the model reads them.

Ranking is done by a Generative Recommender, a transformer that reads more than 1,000 of a member's past interactions as an ordered sequence, each post followed by what the member did with it. It predicts passive actions (click, skip, long dwell) and active ones (like, comment, share). LinkedIn states that demographic attributes are not used to rank.

Dwell time has been a ranking signal since at least 2020. LinkedIn defines a skip threshold below which a view counts as skipped, and lowers a post's score in proportion to its predicted chance of being skipped. In its own words, time well spent is better than more time spent.

  1. 01Millions of posts

    Posts from the member's network and from outside it, refreshed within minutes.

  2. 02LLM retrieval

    A fine-tuned language model turns members and posts into embeddings and matches them by meaning.

  3. 03Generative Recommender

    A transformer reads more than 1,000 past interactions and predicts what the member will do.

  4. 04The feed

    Ranked on predicted clicks, skips, long dwells, likes, comments and shares.

Under the hood

  • Retrieval reads text prompts: a post's format, author, engagement counts and text; a member's profile, skills, work history and the posts they engaged with.
  • Engagement counts are turned into percentile buckets. LinkedIn reports that the correlation between popularity and the model's similarity scores jumped 30-fold.
  • Posts a member saw but ignored are used as hard negatives, to teach retrieval what not to bring back.
  • After the transformer, the ranker adds device type, profile embeddings and aggregated affinity counts, then predicts each action with a mixture of experts.
  • Retrieval runs in under 50 milliseconds over an index of millions of posts.

What it means for a sponsored post

  • The first line decides the skip. A post scrolled past is scored down.
  • The creator's profile is part of the match. A post retrieved for its meaning reaches people whose profiles and history fit the creator's topic: casting is targeting.
  • Comments, shares and long reads count, not only likes.

03

X: how the For You feed ranks posts in 2026

X published the code of its For You feed on GitHub in January 2026 and has updated it since. Posts come from two places: accounts the viewer follows, kept in memory by a service called Thunder, and accounts they don't, found by Phoenix retrieval, which embeds viewers and posts as vectors, by SimClusters and by a list of popular posts.

Phoenix, a transformer built on xAI's Grok, then predicts the probability of every action a viewer could take: like, reply, repost, quote and share; clicks on the post, the profile, a link or a photo; video views and dwell; following the author; and negative actions such as not interested, mute, block, report or scrolling past without dwelling. The final score is a weighted sum of those probabilities, with negative weights on negative actions.

After scoring, the feed is reshaped. An author's second and later posts are discounted, posts from accounts the viewer doesn't follow are multiplied by a factor below one, newer authors are lifted towards a target position, and posts older than 48 hours are filtered out. Visibility filtering runs separately from ranking.

  1. 01Followed and unfollowed accounts

    Thunder keeps recent posts from followed accounts; Phoenix retrieval, SimClusters and popular posts bring in the rest.

  2. 02Phoenix scoring

    A Grok-based transformer predicts the probability of each action, positive and negative.

  3. 03Weighted sum

    Final score = Σ weight × P(action), with negative weights on mutes, blocks and reports.

  4. 04Shaping and filters

    Author diversity, an out-of-network discount, a new-author boost and a 48-hour age limit.

Under the hood

  • The feed is assembled in stages: query hydration, candidate sources in parallel, candidate hydration, pre-scoring filters, scoring, top-K selection, post-selection filters, then blending with ads and follow suggestions.
  • Pre-scoring filters remove duplicates, posts older than 48 hours, the viewer's own posts, blocked or muted accounts, muted keywords and posts already seen.
  • Visibility filtering can allow a post, add a notice, put it behind an interstitial or drop it.
  • Each candidate is scored on its own: a post's score does not depend on the other posts in its batch.
  • A diversity reranker reorders the top posts from their embeddings, so near-identical posts don't sit side by side.

What it means for a sponsored post

  • The first 48 hours are the whole life of a post in For You.
  • Replies, profile clicks and dwell are scored, and so are mutes and reports: bait that earns a click and a mute costs reach.
  • One creator posting twice is discounted. Several creators with one angle each spread further.

04

LinkedIn and X, side by side

LinkedInX
Where candidates come fromOne language-model retrieval matching members and posts by meaning, in and out of networkFollowed accounts, plus Phoenix retrieval, SimClusters and popular posts
Ranking modelGenerative Recommender, a transformer over 1,000+ past interactionsPhoenix, a transformer built on xAI's Grok
What it predictsClick, skip, long dwell, like, comment, shareLikes, replies, reposts, quotes, shares, clicks, dwell, follows and negative actions
Negative signalsA predicted skip lowers the scoreNot interested, mute, block, report and not dwelling carry negative weights
FreshnessPrompts and embeddings refreshed within minutesPosts older than 48 hours are filtered out
DiversityNot detailed in the 2026 publicationAn author's later posts are discounted; a reranker spaces out similar posts
What is publicEngineering articles and a research paperThe feed's source code, under Apache 2.0

05

What our data says: five laws

Law 01

Story first, brand later.

93%of the most-engaged posts don't name the brand in their first line.

Both feeds predict dwell and skip from the opening. A story earns the second line before the brand asks for attention.

Where the brand is first named
  • 8%In the first 80 characters
  • 10%Characters 80 to 200
  • 22%Characters 200 to 400
  • 59%Not in the first 400 characters

Law 02

Short hooks.

60characters in the median opening line. 68% open with 80 characters or fewer.

LinkedIn treats a view shorter than its skip threshold as a skip. A hook that lands in one glance keeps the post on screen.

Opening line length, in characters
Each dot is a sponsored post: opening line length and public reactions
101001,00004080120160+
  • Brand not named in the first 400 characters
  • Brand named in the first 400 characters
  • 80-character hook

Law 03

Short lines.

70%keep their lines to 12 words or fewer on average.

Short lines keep the eye moving, and that is what dwell measures.

Average words per line

Law 04

Statements beat questions.

3%of the most-engaged hooks ask a question. They state something: a number, a scene, a claim to push back on.

A statement invites a reply that adds something, and both feeds score comments and replies as active engagement.

What the opening lines are made of
  • Starts in the first person33%
  • Contains a number32%
  • Uses an emoji15%
  • Pushes back on a belief11%
  • Asks a question3%

Law 05

Weekdays carry B2B.

92%of the most-engaged posts went live Monday to Friday.

Most of the posts that travelled went live on a working day, and on X a post has 48 hours to be seen.

Publication day

06

More patterns in the data

Pattern 01

A few posts carry a campaign.

34%of all reactions went to the top 10% of posts.

Nobody can tell in advance which post will break out. Several creators per campaign spread that risk.

Share of all reactions collected by the best posts
0%0%25%25%50%50%75%75%100%100%Top 10% of posts: 34% of reactions
  • Posts, from most to least reacted
  • If every post performed the same

Pattern 02

Some posts start conversations.

37%of the posts drew more comments than reactions.

Comments are an active signal on both feeds, and a post that takes a position earns them.

Each dot is a sponsored post: reactions and comments
10101001001,0001,000
  • More comments than reactions: 37% of posts
  • More reactions than comments
  • As many comments as reactions

Pattern 03

A personal voice.

74%speak as “I” within their first three lines.

Readers follow people. The brand borrows the creator's voice; it does not replace it.

How the posts are written
  • Speaks as “I” in the first three lines74%
  • Uses an emoji anywhere29%
  • Uses a list31%

Pattern 04

English first, not English only.

21%of the posts that travelled were written in French.

Each market needs creators who write in its language, and one campaign can run in several.

Language of the post
  • 79%English
  • 21%French
Creator's main sector
  • AI41%
  • Marketing16%
  • SaaS13%
  • Sales9%
  • Growth / GTM7%
  • B2B5%

07

Anatomy of a post that travelled

Blueprint of a post that travelled
  • 65-character hook

    68% of the most-engaged posts open with 80 characters or fewer.

  • First person from the start

    74% speak as “I” within their first three lines.

  • 9.5 words per line

    70% keep their lines to 12 words or fewer on average.

Kevin MeyerSponsored post for Synthesia

The best cold call I ever made, I made 40 times first. To no one.

Early in my career, a manager made me rehearse pitches alone in an empty conference room.

Me, talking to chairs.

It felt ridiculous. I wanted to spend that time dialing.

Then a prospect raised an objection I had practiced the day before.

… see more

14556Wednesday
  • A statement, not a question

    Only 3% of the hooks ask a question.

  • Synthesia named after the story

    93% keep the brand out of their first line.

  • Published on a Wednesday

    92% went live Monday to Friday.

Kevin Meyer for Synthesia, measured the way we measure every post. Read it on LinkedIn

Kevin Meyer for Synthesia, line by line.

  1. 01

    The best cold call I ever made, I made 40 times first. To no one.

    A paradox in the very first line. The only way to resolve it is to keep reading.

  2. 02

    Early in my career, a manager made me rehearse pitches alone in an empty conference room.

    A scene from the creator's own career, not a product claim. Readers recognise themselves before they meet the brand.

  3. 03

    Me, talking to chairs.

    Four words of self-mockery. Short lines keep the scroll moving.

  4. 04

    It felt ridiculous. I wanted to spend that time dialing.

    The reader's objection, said before they can think it.

  5. 05

    Then a prospect raised an objection I had practiced the day before.

    The payoff lands before any product is named.

Synthesia is not named in the first 400 characters: the product comes in after the story.Read the full post

08

The Naano playbook

  1. 01Cast for topic, not reach.

    Pick creators whose profile and past posts match your buyers' interests: that is what retrieval matches on.

  2. 02Write the first line for the skip.

    One idea, under 80 characters, no brand name.

  3. 03Tell a scene before the product.

    Let the brand arrive after the story, once the reader has decided to stay.

  4. 04Write in short lines.

    Short lines keep the eye moving, and that is what dwell measures.

  5. 05State, don't ask.

    Give readers something to agree or disagree with; replies and comments are scored on both feeds.

  6. 06One creator, one angle.

    Spread a launch across several voices: X discounts an author's second post, and one voice repeating itself tires its audience.

  7. 07Publish on a working day and stay in the comments.

    On X, a post leaves For You after 48 hours. The first replies are part of the distribution.

  8. 08Keep the post up and measure per creator.

    Naano creators keep sponsored posts live for at least six months, and every creator gets a tracked link.

09

Glossary

Dwell time
How long a post stays visible on screen. LinkedIn counts a view as skipped below a threshold.
Retrieval
The step that picks a few thousand candidate posts out of millions, before ranking.
Embedding
A list of numbers that represents a member or a post, so that similar meanings sit close together.
Out-of-network
Posts from accounts the viewer doesn't follow or isn't connected to.
Hook
The first line of a post, read before anyone decides to stay.
Negative signal
An action that tells the feed a post was unwanted: hide, mute, block, report or scroll past.

10

Questions

Is X's recommendation algorithm open source?

Yes. X publishes the code of its For You feed on GitHub (xai-org/x-algorithm) under the Apache 2.0 licence.

What does LinkedIn's feed algorithm use in 2026?

A language-model retrieval system that matches members and posts by meaning, then a Generative Recommender that reads more than 1,000 of a member's past interactions to predict clicks, skips, long dwells, likes, comments and shares.

Does LinkedIn use demographics to rank posts?

No. LinkedIn says no demographic signal decides what the feed shows. It uses demographic data only to test that no group of creators or members is systematically ranked lower.

How long does a post stay in X's For You feed?

The published code filters out posts older than 48 hours before scoring.

Do negative reactions hurt reach on X?

Yes. Not interested, mute, block, report and scrolling past without dwelling are predicted and given negative weights in the final score.

What makes a sponsored creator post travel?

In Naano's data, the most-engaged posts open with a short statement, speak in the first person, are written in short lines and name the brand after the story.

11

Method and sources

The algorithm chapters quote what LinkedIn and X published themselves: engineering articles, a research paper and the open-source code of the For You feed, read in October 2026. They describe systems that change often; we date every source.

The data chapters use the most-engaged sponsored posts published through Naano by creators who agreed to appear on our site, with public reactions and comments only. The hook is the first line of a post; brand mentions are checked in its first 400 characters. These are patterns in posts that performed, not controlled experiments. The writing-AI figure is the number of LinkedIn posts Naano's writing AI has analysed.

Totals: the sponsored posts Naano has studied and their public reactions and comments. Charts: the 116 most-engaged of them, by creators who agreed to appear on the site, as of November 10, 2026. The writing-AI figure covers all the LinkedIn posts it has analysed.

Next step

Want these laws applied to your next launch?

Naano's campaign team casts the creators, writes the briefs from this research and reports on every post.