X's ranking weights, straight from the published code

The short answer

Concentration of conversation beats volume of approval. The ranking model predicts what each person will do with each post, multiplies those predictions by published weights, and adds them up. Anything that sparks a real exchange outranks anything merely agreed with.

The pipeline

1. Find candidates. About 1,500 posts per request, roughly half from accounts you follow and half from beyond.
2. Rank. A model predicts how likely you are to like, repost, reply or click, and combines those predictions into one score.
3. Filter. Removes blocked and muted accounts, posts you have seen, and too much from a single author.
4. Mix. Blends the ranked posts with ads, follow suggestions and other items.

The weights

These come from the home timeline ranking README in the-algorithm-ml, published in April 2023. xAI's 2026 code keeps the weighted-sum structure but does not publish a matching table, so the 2023 figures remain the clearest public guide.

The negative weights are larger than the positive ones

This asymmetry is the most under-reported fact about the platform. Optimising for approval has a low ceiling and a hard floor: a post that annoys people loses more reach than a bland post ever gains.

What changed in the 2026 release

xAI published a new ranking codebase in January 2026 and has updated it roughly every four weeks since. The architecture moved rather than the principle:

• Phoenix, the retrieval and ranking service, ranks posts with a transformer model instead of the 2023 neural network, predicting likes, replies, reposts, quotes, shares, clicks, dwell, video views, follows and negative actions.
• Home Mixer builds the For You feed; Thunder holds recent posts from accounts you follow in memory.
• A decay stops one author dominating your feed, out-of-network posts are discounted, and new authors get a boost.
• Grox classifiers label posts as spam, adult content or violent media as they are published.
• Since November 2025 the Following feed is sorted by predicted engagement and relevance rather than strictly chronologically — chronological order is still available.

The direction of travel is clear: hand-tuned weights matter less each release, and what compounds is content a real person wants to answer. Worth noting for anyone selling engagement tactics.

Two facts that decide reach before content

Published weights are not the only variable, and pretending otherwise would be dishonest. Buffer's study of 18.8 million posts from 71,000 accounts found paying accounts get around 10 times the median reach of free ones — under 100 impressions for free accounts against close to 600 for Premium.

Links are worse. Buffer found median engagement on link posts from non-Premium accounts has been stuck at zero since March 2025, while Premium link posts managed about 0.28%. If you are on a free account and your link post did nothing, that is the explanation, and the workaround is to put the link in a reply rather than the main post.

What the weights reward, in practice

• Ask something worth answering. A reply is worth 27 likes; a reply the author engages with is worth 150.
• Be there after you post. Conversation depth is the highest-weighted thing in the file.
• Write text-first. In Buffer's 2026 analysis of 45M+ posts, text-only posts had about 20% more engagement than video on X.
• Create bookmark-worthy reference posts — data, frameworks, checklists — since saves are treated as a positive signal.
• Never give anyone a reason to mute you. The negative weights dwarf the positive ones.

Where Surfolks fits

Surfolks is a read-only feed covering X and LinkedIn. It collects public posts from the people you choose, keeps a history you can filter by person and date, and scores posts against your goals. It does not collect X likes or bookmarks, does not bypass protected accounts, and never posts, replies, or sends DMs — so nothing in the weights above is something you have to game. The per-platform boundary is spelled out at what we collect, and the LinkedIn equivalent is in how the LinkedIn feed ranks.