How the X (Twitter) Algorithm Really Works in 2026 (And What's Just Guesswork)

Most platforms' algorithms are a black box people guess at. X is the exception: it has published the source code of its recommendation algorithm. So we don't have to guess how reach is decided; we can read it in the code. With one caveat: the code is not the same thing as the model that's actually running.

Let's take three things one at a time: what X has actually revealed, what even that leaves out, and what's just social-media folklore.

What X says about itself

X first opened its recommendation algorithm in 2023. In January 2026 it published an entirely new version: xai-org/x-algorithm is Apache 2.0 licensed and contains the structure of the For You feed: Thunder, which retrieves posts from accounts you follow, and Phoenix, which retrieves and ranks everything else with a Grok-based transformer model.

This is the single most important change, and it makes a large share of the X advice online obsolete.

Hand-tuned weights are history. The old 2023 code showed relative multipliers: a like as one, a repost at roughly double, a reply at well over twenty times. Those are the numbers tip lists still repeat. The documentation for the new release states plainly that X has eliminated every hand-engineered feature and most heuristics from the system. Those multipliers describe a system that no longer exists.

The model predicts fifteen actions. The code names them: favorite, reply, repost, quote, click, profile click, video view, photo expand, share, dwell time, and author follow. There are four negatives: "not interested," block author, mute author, and report. A post's score is a weighted combination of those predictions. The list is concrete and useful, because it tells you what X is trying to predict in the first place.

The weights weren't published. And that's where it stops. The marketing trade publication PPC.land went through the code after release and found not a single numerical coefficient in the package: whether a favorite carries a weight of 0.1, 1.0, or 10.0 goes unstated. The author-diversity attenuation formula and the post-age filtering thresholds aren't disclosed either. The direction is still the same as before, with replies and conversation counting for more than a like. The exact ratio just can't be read anywhere.

Negative signals are part of the score. Blocks, mutes, and reports lower a post's score directly. An active Community Note lowers distribution while it's live.

The first hour decides. The model weighs early velocity. A post that gets twenty replies in the first half hour spreads clearly better than the same number over a day. Early momentum tells the algorithm whether to expand distribution to the For You feed.

What open source doesn't tell you

Code is not the model. The release contains the architecture and a small example model, but not the production weights. Those are the part that actually decides what shows up in your feed. You're seeing the kitchen, not the recipe.

Openness is also communications. Publishing the source code is a competitive advantage for X and an answer to regulatory pressure. It's still a release the platform chose and scoped itself, not an outside audit.

The multipliers still circulating online come from a dead system. Search for X's algorithm and you'll see it. Tip articles published in 2026 confidently give you a formula along the lines of "like ×1, repost ×20, reply ×13.5, profile click ×12" and present it as fresh reading of the open code. That formula is not in the 2026 release. It's the hand-tuned 2023 scoring recycled, precisely what X says it removed. An exact number looks authoritative, which is why it spreads.

What the data suggests

The next tier is weaker than the source code but stronger than a guess. Outside measurements show how the numbers look in practice.

Be careful here, because two different numbers travel under the same name. Adilo's measurement calculates engagement per impression and lands at 3.2% for text posts, 2.09% for images, and 1.34% for video. Social Insider calculates it per follower and lands at 0.12%. Neither is wrong; they just measure different things. So only compare your own numbers against a figure calculated the same way.

Adilo's data holds one surprise: plain text averaged the best engagement and video the worst. Video was the most consistent, though. Its median (0.94%) sits close to its average, while text's median drops to 1.13%. Translated: text produces a few big hits and a lot of silence, video produces a steady middle. The sample is only 3,200 posts, so don't read it as law.

What's just guesswork

This is where many tip lists go off the rails.

"Collect likes, they're what matter." A like is the cheapest signal, and both the old and the new system weigh replies clearly higher. Aim for conversation, not thumbs.

"A reply weighs exactly 13.5 times a like." Or 27 times, depending on which article you read. The fact that the circulating numbers disagree with each other tells you enough: they're the 2023 code recycled, read in different ways. The current release contains no numerical weights at all. The direction holds; the number doesn't.

"Never post a link, X buries it." An exaggeration. A link click is a positive signal the model predicts. A link that sends the reader away instantly shows as poor dwell — but there's no mechanical link ban.

"Tweet 20 times a day, volume wins." It doesn't. One post that starts a conversation you can reply to beats twenty that get skipped.

"Hashtags boost reach." The effect is small. A few relevant tags won't hurt, but they aren't the lever that makes a post fly.

What the algorithm actually rewards

Set the guesswork aside, and the open code plus the outside data point to a few durable principles. They'll still hold next year, because they follow from the structure, not from this month's adjustment.

• Conversation is the currency. A reply you reply back to is the single strongest signal. Post things people want to answer.
• A like is cheap, and the algorithm knows it. Collecting thumbs won't take you far.
• The first hour decides. Be present after posting, don't vanish.
• X is a text platform, unlike the others. Video carries elsewhere; on X, plain text does at least as well in the measurements. Don't port your Reels thinking here unchanged.

How a small business does this without a marketing team

Knowing this isn't the bottleneck for a small business. Applying it week after week, with no dedicated person, is.

X's hook is that first hour: when you post and whether you're around afterwards. Replying is manual work, but content simply not appearing on a regular basis is the usual reason a small business's X account goes quiet. And X is rarely the only channel — the same content needs to reach Instagram, LinkedIn, and elsewhere too. Volume isn't the deciding variable here either, however often that's assumed. Also worth reading: how often you should post on social media.

Steadybeat handles the posting side: you write and schedule posts in one go and send the same content to several channels at once. That frees you to focus on the hour of conversation once the post itself is already handled.

Sources

Verified in August 2026. The algorithm changes, so it's worth checking the platform's own points against the original source.

Platform sources (X publishing on its own product):

• xAI / X, xai-org/x-algorithm. The current Grok-based For You algorithm, opened in January 2026, Apache 2.0. Contains the architecture and a small example model, not the production weights.
• X / Twitter, the-algorithm. The 2023 release. The exact multipliers circulating online come from here, not from the current system.
• Social Media Today: X Shares Insights Into Key Factors That Dictate Post Reach

Independent analysis of the code:

• PPC.land, X's algorithm source code drops: what it reveals about the platform's feed mechanics (21 January 2026). 15 predicted actions, no numerical weights.

Outside measurements (not confirmed by X):

• Adilo, X engagement rate chart. 3,200 posts, engagement calculated per impression.
• Social Insider, Social media benchmarks. 0.12% engagement per follower, full-year 2025 data.