X growth playbook: what the open ranking code rewards

Updated 21 September 2026. X open-sourced parts of its ranking pipeline (the `xai-org/x-algorithm` repository), and that code tells us more about what actually gets a post distributed than any "growth hack" thread does. This guide is honest about the boundary between verified ranking incentives and editorial judgment. Anyone claiming to know X's full deployed ranking weights is guessing. What you can do, and what most growth advice does not do transparently, is build your posting habits around the incentives that are actually verifiable in the code, and treat everything else as a hypothesis to test on your own account.

1. What the open-sourced ranking logic actually tells us

Reviewed from the public `xai-org/x-algorithm` repository (the home-mixer request pipeline and its ranking scorer).

**Verified, read directly in the code:**

• Ranking is per-viewer prediction, not a universal post score. A post does not have one "algorithm score." It has thousands of viewer-specific predicted-engagement outcomes, one per person whose feed it might reach.
• The heavy ranker optimizes predicted engagement across reply, repost, quote and like-type actions. As of the weights published 13 August 2026, the relative pull of each action (from `home-mixer/params/param.rs`) is approximately: favorite 0.5, reply 5.0, quote 5.0, share via copy link 20.0, follow the author 4.0. Replies and reposts are structurally worth far more than passive likes, and a copy-link share is worth the most of all. Negative actions pull hard in the other direction: not interested about -43, mute about -59, report about -234.
• For comparison, the 2023 version of the algorithm (the older `twitter/the-algorithm-ml` release) weighted things differently: like 0.5, retweet 1, reply 13.5, and reply engaged by the author 75. Those numbers are historical and no longer reflect production. If you read older "algorithm leak" content online, check whether it is quoting the 2023 weights before you trust it.
• Negative feedback is a first-class input. Mutes, blocks, reports and "show less" decisions fight against distribution harder than positive signals push it. This is why bait and rage-bait patterns are the costliest thing you can post.
• Author diversity and out-of-network adjustments exist in the pipeline: a feed throttles how much one author can dominate it, and posts from accounts you do not follow face a higher bar to get in. In practice, your post competes in strangers' feeds on predicted usefulness, not follower count, but the bar for a stranger's feed is genuinely higher than for your own followers'.
• X says the repository mirrors production and is updated roughly every four weeks. Treat any weight, including the ones above, as a snapshot, not a permanent constant.

**Editorial inference, ours, labeled as such:**

• First-hour engagement from your own followers likely determines whether a post escapes to feeds outside your network. We cannot verify the exact window from the code, but the mechanism (candidates get sourced and ranked using early feedback signals) makes early engagement plausibly important. The practical consequence: replying to your own post's early comments, quickly and substantively, is one of the highest-leverage five minutes you can spend after publishing.

2. The dimensions worth optimizing for, and why each exists

Each of these maps to a verified ranking incentive or a negative-feedback risk you can act on directly.

| Dimension | Ranking logic | What a strong post does | |---|---|---| | Hook | Visible content decides whether someone scrolls past; the fold is the first impression | First line under about 90 characters, concrete (a number, a name, a specific moment) | | Shareability | Reposts and copy-link shares carry a post into new networks, and copy-link is the single highest-weighted action | One self-contained, forwardable sentence near the top | | Reply invite | Replies are among the highest-value predicted actions, but bait is punished by negative feedback | At most one genuine, specific question | | Quote-worthy | Quote posts are their own distribution event | An attributable position, not a summary of someone else's point | | Follow reason | The follower graph seeds all future distribution | Three or more concrete specifics: numbers, tools, outcomes | | Dwell fit | More substantive content earns more time on screen; threads multiply the number of ranking surfaces | 120 to 280 weighted characters, or split into a thread | | Link handling | Posts with an outbound link in the body are reach-suppressed and, under X's current API pricing, cost the poster money to send | No outbound link in the body; put it in the first reply instead | | Bait risk | Negative feedback outweighs positive bumps by a wide margin | No hollow openers, emoji stacks, all-caps, or follow and repost asks | | Topic focus | Interest-graph classification concentrates the right audience | One idea per post, at most one or two hashtags |

Different goals call for different emphasis: an account optimizing purely for reach will lean on shareability and hook; an account building buyer trust for a B2B product will lean harder on follow reason and dwell fit. Both read from the same underlying incentives.

3. Practices distilled from competitors and practitioners

These are practitioner consensus, not measured facts from the ranking code. We adopt the ones that survive the honesty bar above.

• Write the hook first, then earn it. Hook libraries are a structural pattern worth learning from, not something to copy verbatim.
• One idea per post. Multi-idea posts split topic classification and dilute every ranking surface at once.
• The reply is the growth surface. The single best strategy for a small account on X is substantive replies on larger, adjacent accounts within hours of their post going up. This is also what X's own ranking wants to amplify: replies are its native currency.
• Quote when you have a standalone take; reply when you are adding to someone else's. A quote post is its own distribution event.
• Threads are for earned depth. Each child post is a separate ranking surface with its own predicted engagement, so only split into a thread when the material genuinely needs the room.
• Native articles are an authority asset, not a reach play. They rank in article-specific surfaces and tend to have a longer shelf life than a normal post.
• What to avoid, even though some tools still offer it: engagement pods, automated likes or replies, follow-for-follow loops, manufactured controversy, "best time to post" claims with no account-specific evidence behind them, and any virality score presented as a guarantee. X's automation rules and its heavy weighting of negative feedback make all of these liabilities, not shortcuts.

4. How this should be tuned over time

The weights above are a snapshot from 13 August 2026. X has said it updates the public repository roughly every four weeks, and the ranking model itself changed materially in January 2026 (a newer transformer-based ranker replaced the older one). Do not treat any specific number as fixed. Re-check the source before making it a rule you build a whole content strategy around, and prefer comparing your own account's outcomes across formats and time windows over trusting any third party's claimed weights, including this one.

5. What "using the algorithm responsibly" looks like in practice

Reading the ranking code is legitimate and public. What is not legitimate, and what draws enforcement, is automating the actions the ranking rewards: auto-liking, auto-following, auto-replying with keyword triggers, or bulk actions designed to farm engagement. X's current API terms explicitly prohibit keyword-triggered replies, auto-DMs and bulk follow and unfollow behavior, and treat automated likes and retweets as engagement farming. A tool that reads what is already on your screen and drafts a reply for a human to review and send is a fundamentally different, and safer, category from a tool that acts on your account without you in the loop. If a growth tool promises to post, like, follow or message on autopilot, that promise is the risk, not the feature.