How the X Algorithm Works in 2026

For publishers, the useful frame is which outcomes X predicts for each viewer, not which hack supposedly unlocks reach.

This guide starts with X's own repository: the August 2026 For You feed overview and Home Mixer parameters. The SocialPilot algorithm guide and SocialRails growth guide provide secondary editorial context, not proof of internal behavior.

One caution before reading the numbers: X says experiments can change many values for part of its traffic. Some prompts and anti-abuse rules are withheld. The repository is the best public map, not a promise that every request uses one frozen configuration.

The For You feed is a pipeline, not one formula

The feed is assembled again for each request. Ranking is only one part of the job.

That sequence, documented in the official repository overview, explains why simple folklore fails. A high-scoring post can still be filtered. A relevant post can lose a slot when the reranker needs more variety. A good post may never enter one viewer's candidate pool.

X pulls candidates from inside and outside your follow graph

Thunder supplies recent material published by followed accounts. Phoenix retrieval matches a representation of the viewer with representations of posts. SimClusters infers communities from audience engagement patterns and finds posts associated with similar groups. Together, the recommendation sources give creators a path beyond their existing follower graph.

The current parameter file sets maximum results of 1,200 for Thunder and 1,000 for Phoenix, with both Phoenix and SimClusters enabled. Those are ceilings, not a claim that every request receives exactly 2,200 candidates.

Phoenix receives a sequence of the viewer's recent actions as context. The current maximum sequence length for both retrieval and scoring is 1,024. Explicit and implicit engagement signals are enabled, with a separate cap of 15 signals per type. This is stronger evidence than the often repeated claim that X only looks at a person's last 128 engagements. The code exposes a larger maximum, although it does not tell us that every stored action has equal influence.

For publishers, the practical reading is modest: topic and audience fit help retrieval. An account need not cover one subject forever, but each post needs enough context to reach people likely to care about it.

Filters narrow the pool before scoring

The public pipeline removes several kinds of candidates before Phoenix assigns action probabilities. According to the first-party filter documentation, that includes:

• duplicate copies returned by multiple sources
• candidates published more than 48 hours earlier
• material authored by the viewer
• items the person has already encountered or received in the session
• posts from blocked or muted accounts
• posts matching muted keywords
• inaccessible subscriber-only posts
• some out-of-network replies, reposts, and adult-content candidates

The 48-hour rule is concrete evidence that recency affects For You eligibility. It is not evidence for a universal minute-by-minute freshness multiplier across every X surface. Search, profiles, notifications, and other timelines may use different systems.

Distribution is not only a contest for a higher score. X also avoids repeatedly serving the same post to one person.

Phoenix predicts actions, then X assigns value to them

Phoenix does not output one vague measure called relevance. It estimates probabilities for actions such as liking, replying, reposting, quoting, sharing, following the author, opening media, spending time, selecting "Not interested," muting, blocking, or reporting. The ranking scorer multiplies each predicted probability by a configured weight and adds the results.

The table below shows selected defaults from the August 2026 ranking parameters.

These coefficients need careful reading. They are applied to predicted probabilities, not awarded as fixed points after a person acts. X notes in the code that weights reflect both the value of an action and how common that action is. A rare event can need a larger coefficient to affect the total. A reply coefficient above the favorite coefficient therefore supports the idea that replies can be valuable, but it does not establish a fixed exchange rate between one reply and a certain number of likes.

Negative feedback carries much larger negative coefficients. The model explicitly prices the risk of a report, mute, block, or "Not interested" response. Provocation can attract visible replies while also raising the probability of costlier negative feedback. Public engagement totals alone cannot show that tradeoff.

X adjusts scores for repetition, network distance, and discovery

After the weighted sum, more rules change the order.

Mutual-follow original posts get a reply-value lift

One 2026 change is explicit in the ranking scorer. When the author and viewer follow each other, an original post gets a 15.0 boost on top of the 5.0 base reply weight, for an effective default of 20.0. Replies and reposts are excluded. X's change note records the experiments behind the July update, giving first-party support to SocialPilot's reporting on mutual follows.

Repeated authors lose strength within the slate

Author diversity is enabled. The scorer derives a multiplier from a 0.5 decay parameter and a 0.25 floor. Under the published formula, the second candidate from one author gets 0.625, while later candidates approach 0.25. This is repeated-author control inside one ranking slate, not a fixed daily quota.

Both secondary guides warn against flooding the feed, which is sensible editorial advice. But the current code shows score decay for successive candidates from the same author, not a published rule that every creator receives a set number of daily slots.

Out-of-network posts face a discount

The current Home Mixer defaults set the out-of-network factor to 0.75. The pipeline also applies out-of-network treatment to certain replies and reposts from followed accounts. Recommendations can still rank, but they generally need enough predicted value to overcome that discount.

New authors can receive a controlled lift

The cold-start defaults cover posts from authors below 1,000 impressions, with a maximum follower count of 1,000 and a maximum post age of 24 hours. Their target slots are 15 and 16. This is an opportunity for discovery, not guaranteed reach. Eligibility still depends on the rest of the pipeline.

The reranker trades some score for variety

VMRanker is enabled and uses a diversity method that considers post embeddings. Its job is to avoid a run of highly similar neighbors, even when those posts all scored well. A publisher should not assume that several near-duplicate posts can occupy adjacent high positions simply because the topic is working.

Visibility rules can override ranking

X separates relevance ranking from visibility. A different system considers account and post labels, the viewer's blocks and mutes, protected or suspended status, subscription access, country, settings, and other conditions. It can allow a post, place it behind an interstitial, or drop it.

Some rules exist only for recommendations beyond the person's follow graph. A post may remain visible to followers while being excluded from recommendation inventory. The official visibility description is the clearest reason not to treat follower reach and non-follower reach as the same problem.

X has also published an Under the Hood account report for aggregate visibility-related labels, with availability expanding over time. If distribution changes abruptly, that report is more useful evidence than guessing about a secret reach penalty.

What the public code does not prove

The two secondary guides capture common creator concerns, but several strong claims go beyond the primary evidence available here.

Use the SocialPilot guide and SocialRails guide as idea banks for experiments, not as substitutes for code or your own analytics.

A practical X strategy based on the code

The repository points toward a less theatrical publishing plan.

1. Write for a recognizable audience need. Retrieval uses recent behavior and community similarity. Make the topic, claim, and intended reader clear enough to match.
2. Earn meaningful actions without bait. Replies, shares, quotes, follows, and clicks all enter the value model. Empty outrage may also increase mutes, blocks, and reports.
3. Give each post a complete job. A post should teach, report, argue, entertain, or invite a useful response. Do not rely on a format label to carry weak material.
4. Avoid near-duplicate bursts. Repeated-author decay and diversity reranking reduce the case for publishing several similar posts at once.
5. Treat links as a testable variable. Compare matched groups of link and non-link posts at similar times. Record impressions, meaningful engagement, profile actions, and conversions. Do not move every link to a reply because a guide promised a guaranteed penalty.
6. Use your audience's timing, not a universal schedule. The code establishes a 48-hour For You cutoff, not one best hour or daily volume. Your analytics should set the cadence.
7. Watch negative feedback and eligibility. If reach falls, inspect content quality, repetition, audience fit, and any available visibility labels before blaming one hidden switch.

Review groups of comparable posts rather than one outlier. Aggregate results contain noise from audience mix, timing, competition, and experiments.

Frequently asked questions

Is the X algorithm fully open source in 2026?

X publishes substantial code for For You retrieval, ranking, filtering, model training, and visibility systems. It also says some model prompts and anti-abuse rules are withheld. Deployment configuration and live experiments can differ from the checked-in defaults. "Publicly inspectable" is more accurate than "every production detail is exposed."

Do replies matter more than likes on X?

The current default reply coefficient is higher than the favorite coefficient, so predicted replies carry more value in the weighted score. That does not create a fixed conversion rate between actual replies and likes. The model predicts many actions at once, then later adjustments and filters still apply.

Does X penalize links?

The reviewed first-party files do not show a blanket penalty for merely containing a link. They show a positive default weight for the predicted action of opening one. Link posts can still perform differently because of audience behavior, content quality, model features, or experiments. Test the claim on your account rather than treating it as guaranteed.

Does X Premium increase reach?

The reviewed first-party files do not establish a universal Premium reach multiplier. Premium may change product features and audience behavior, but those effects should not be described as a code-confirmed ranking bonus without stronger evidence.

How often should I post on X in 2026?

The code does not prescribe a universal number. Publish at a pace that preserves quality and lets you compare results. Repeated-author decay suggests that several similar candidates can compete with one another inside a feed request, but it does not amount to a daily cap.

Can a small account reach people who do not follow it?

The architecture gives small accounts a route to non-followers. Phoenix and SimClusters source recommendations beyond the follow graph, and the parameters include a limited cold-start lift for eligible newer authors. It does not guarantee impressions, because scoring, reranking, visibility, and competition still decide what survives.

Sources and method

Primary evidence: xai-org/x-algorithm at commit a389166, especially its For You feed overview, Home Mixer defaults, ranking scorer, and mutual-follow change note.

Secondary editorial references: SocialPilot, How the X Algorithm Works in 2026 and SocialRails, How to Grow on X in 2026. Claims from these guides are labeled as secondary when the first-party code does not confirm them.