X Algorithm 2026: The Published Weights and How to Grow on X

The X algorithm in 2026 is Phoenix, a Grok-based model that predicts how likely each viewer is to reply to, share, follow from or mute your post, then adds those predictions up using weights X has published on GitHub. A predicted reply is worth 5.0 and a copy-link share 20.0, while a like is worth 0.66. So the way to grow on X is to write posts people reply to and forward, and to reply early and well under bigger accounts.

The short version

• X is the only major platform that publishes its ranking weights. Replies, quotes and DM shares score 5.0, a copy-link share 20.0, a like 0.66.
• The negative weights are bigger than the positive ones: "not interested" is -52, mute -59, report -234. Off-topic bait costs more than it earns.
• Since July 2026, original posts from people you mutually follow get an extra +15 on the reply weight. Follow back the people who engage with you.
• Posts older than 48 hours drop out of For You, and the lift for small accounts only covers a post's first two hours.
• In Buffer's 18.8M-post study, free accounts had a median under 100 impressions per post. Premium accounts got about 600.
• Below about 10,000 followers, replies grow you faster than posts. Work a List of 20-30 accounts and cap yourself at 1-2 replies per account per day.

How does the X algorithm work in 2026?

Since January 2026, the For You feed has been ranked by Phoenix, a transformer adapted from Grok-1. X open-sourced the code in the xai-org/x-algorithm repo, and an August 13, 2026 release added the configuration parameters, including the weights.

The pipeline pulls candidates from people you follow, from embedding-based retrieval of similar posts and from a popular-posts list. It filters out duplicates, posts older than 48 hours and posts you've seen. Phoenix then predicts about 25 actions for you as a viewer, from reply and repost to dwell time and "not interested."

The score is a weighted sum of those predictions. A reranker trades a little score for variety, and visibility filtering can drop a post entirely based on labels, blocks or country rules (README).

The Following tab has been Grok-ranked too since around November 30, 2025, with chronological still in the feed menu (Social Media Today). And ignore the 2023 weights still quoted everywhere (reply 13.5, report -369). They came from the old ranker (CSMaP NYU).

What are the X algorithm weights?

These are the public defaults in home-mixer/params/param.rs, as of October 2026.

The README adds a caveat you should take seriously. The weights "scale predicted probabilities, not raw engagement counts," so a report weighted at -234 doesn't mean one report cancels about 355 likes. Read the ratios as direction. The defaults also move: an earlier snapshot put the like weight at 0.5, and production can differ from the repo (FrameOS).

Bookmarks aren't a predicted action in the code, though Buffer lists "replies and saves" as early signals. Go with the code: design for replies and shares. In practice I'd end posts with a question people can answer in one line, write things someone would send a colleague, and stay inside the topics your followers came for.

What is the bidirectional follow boost?

X ran an A/B test from July 10, 2026, adding 5, 10, 15 or 20 to the reply weight for posts from accounts the viewer mutually follows (BIDIRECTIONAL_BOOST_CHANGE.md).

The +20 version rolled out on July 13. On July 24 X cut it to +15 after feedback that posts from non-followed accounts, like World Cup discussion, felt under-shown. It applies only to original posts, not replies or reposts.

For a mutual, your reply weight goes from 5 to 20. That makes mutual follows a distribution asset. Follow back the people in your niche who reply to you.

How long does a post have to perform on X?

Not long. Buffer describes X testing each post on a small group before expanding reach (Buffer), and the code puts hard limits on it. Candidates older than 48 hours are filtered out of For You. The cold-start boost, which lifts authors with fewer than 200 impressions toward slot 15-16 of the feed, applies only to posts up to 7,200 seconds old, which is two hours (param.rs).

Two more adjustments matter. Each post after your first in a viewer's feed gets multiplied by a decaying factor down to a floor, so posting five times in an hour gives you less per post. Posts from accounts the viewer doesn't follow get discounted (README).

So post when the people who reply to you are online, stay around for the first two hours, and spread posts across the day.

Do you need X Premium to grow?

If you want strangers to see you, yes. Buffer analysed 18.8 million posts from 71,000 accounts between August 2024 and August 2025. Median impressions per post were under 100 for free accounts, about 600 for Premium and above 1,550 for Premium+. Median engagement rate was 0% for free accounts against roughly 0.5% for paid tiers (Buffer).

That data is correlational, like every vendor number in this article. People who pay for Premium also post more and treat X as work. And the open-source ranking code doesn't mention subscription tier at all (README).

But X's own help page says reply prioritization scales with tier, with Premium+ getting the "largest reply prioritization." For a reply-led strategy, that's the feature you're paying for. I'd treat at least the middle tier as the cost of entry.

Does X penalize links in 2026?

X says no. The data says links still hurt.

In October 2025, product head Nikita Bier explained that "posts with links tend to get lower reach" because the browser covered the post and people forgot to like or reply, so X got no signal. X started testing an in-app browser that keeps the post visible (Engadget). On July 28, 2026, Bier told Mark Zuckerberg he no longer needed to put links in replies, and later said "Links are not deboosted. We fixed that issue last year" (PPC Land).

The third-party numbers disagree. Free accounts posting links have had zero median engagement since March 2025, against 0.28% for Premium, per SocialSpool citing Buffer. A PPC Land review of the Phoenix model found link posts got 94% fewer views than comparable posts without links (PPC Land).

The code explains how both can be true. There's no explicit link penalty, but a link click is weighted 0.2 against 5.0 for a reply. A link that pulls someone off X before they reply gives the model little to score.

I'd keep the hook link-free and put the link in the first reply or at the end of a long post.

What content formats work best on X?

Ordinal studied 87,528 X posts from January 2024 to May 2026 and found threads well ahead of everything else (Ordinal).

Ordinal's author cautions that thread results depend heavily on the opening post. Hypefury puts it more bluntly, claiming the hook decides 80-90% of a thread's outcome (Hypefury).

Long posts trade reach for rate. A vendor sample of 672 long-form posts found 500+ character posts averaged 4.37% engagement but 25,844 views, against 2.26% and 56,165 views for posts under 280 characters (Tweetloft). That conflicts with Ordinal on which length wins on rate. Go with the bigger sample: 100-200 characters for daily posts, long-form when you want depth over reach.

For threads, the standard structure is 7-12 posts, one idea each, under about 200 characters per post, with the whole thread published at once (Unilink; SuperX). Beehiiv suggests testing the hook as a single post first and only threading it if it performs (beehiiv).

X is pushing vertical video with a dedicated tab and, since March 2026, a portrait-first player (StratCom Academy). Yet the public defaults weight a video open at 0.02. Post video if you make it anyway, but don't build your plan around it.

How do small accounts grow on X?

By replying. One creator's version is that below about 10,000 followers you grow more by replying than by posting (Postory). The weights back this up: a reply that earns a profile visit and a follow (4.0) is the behaviour Phoenix is built to reward.

Teract's "70/30 rule" says to spend 70-80% of your X time replying to accounts 2-10x your size, on posts with under 20 replies, within 5-15 minutes of posting, and the rest on your own content (Teract). Teract is a vendor and its growth claims are unverified, but Indie Hackers lands on a similar 80/20 reply-to-post split.

Here's the routine I'd run:

• Build one X List of 20-30 accounts. Roughly 40% mid-size accounts, 40% peers near your size and 20% very large accounts for visibility (Postory).
• Do 10-20 good replies a day. Guidance runs from 10-20 to 100 a day, but quality dominates past about ten (Bisonary).
• Cap it at 1-2 replies per account per day (Xpert).
• Answer everyone who replies to you within an hour or two, so the back-and-forth builds (Buffer).

A reply that gets noticed arrives early, adds a specific number, example or counterpoint in 1-3 sentences, reads on its own, and never opens with praise. "First" and "great post" are the pattern the ranker buries. The same rubric works on LinkedIn, covered in how to comment on LinkedIn.

Watch your pace. Bursty or templated replying can trigger "reply deboosting," where your replies sink below "Show more replies" (ShadowbanRadar). Recovery guides say cut volume moderately rather than stopping (X-Deleter). X publishes no reply limit.

How often should you post on X?

More often than anywhere else. Hootsuite recommends 2-3 posts a day and Buffer 3-4, with 3-5 for creators (Hootsuite; Buffer).

Timing data conflicts. Buffer's 8.7 million posts put Tuesday 9 a.m. first, then Wednesday 9-10 a.m., with 9-11 a.m. on weekdays most reliable (Buffer). Sprout Social's March 2026 data favours 12-6 p.m. Tuesday to Thursday, with Saturday worst (Sprout Social).

Use the heatmaps as a starting point and trust your own analytics. Given the two-hour cold-start window and author-diversity decay, two well-spaced posts beat four stacked in one hour. If you have a day job, 2 posts a day and 10 minutes of replies is a sustainable floor; personal branding with a full-time job shows how that fits a five-hour week.

What hurts your reach on X?

The weights give you the list:

• Posts that draw "not interested." At -52, it outweighs a block, and rage-bait gets it from your own followers.
• Links in the hook.
• Burst or templated replies.
• Engagement groups. X stopped paying revenue share on replies after Bier said about half of reply spam came from Telegram groups gaming payouts (PPC Land).
• Hashtag stuffing. Guides range from 0-1 hashtags to "functionally dead" (Teract).
• Deleting and reposting the same post to get a second shot (SocialSpool).

If reach drops off a cliff, check Under the Hood (x.com/i/jf/under_the_hood), X's tool showing visibility labels on your account (README). For how X fits next to LinkedIn, Instagram and Substack, see the four-platform playbook.

FAQ

Is the X algorithm open source?

Mostly. The For You ranking code, Phoenix model, configuration weights and visibility-filtering code are public in the xai-org/x-algorithm repo. X withholds some pieces, including its Grox prompts and some rules, to make the system harder to game.

What does the X algorithm value most in 2026?

A copy-link share (20.0) is the biggest positive weight, followed by replies, quotes and DM shares at 5.0 each and follows at 4.0. Likes are worth 0.66. On the negative side, reports (-234), mutes (-59) and "not interested" (-52) carry the largest weights in the file.

Do links get deboosted on X?

X's product head says no, and the published code has no explicit link penalty. Third-party data still shows link posts getting far less engagement, likely because a link click scores 0.2 and pulls people away before they reply. Keep links out of the main post's hook.

How many replies a day should a small account post on X?

Aim for 10-20 substantive replies a day on a List of 20-30 accounts, with no more than 1-2 replies per account. Spread them out. Bursts of templated replies can get your replies hidden.