X open-sources For You ranking weights: a Report costs 468 likes of reach

Update context: This is the third chapter in explainx.ai's coverage of X's open-source push — after the May 2026 For You algorithm release and July 2026's full-codebase pledge. This time X didn't just publish more code — it published the actual numbers.

On August 13, 2026 at 10:48 PM, the official @XOpenSource account posted that X is "open-sourcing the code that affects a post's visibility in the For You timeline" and shipping a new "Under the Hood" page that shows users the labels applied to their own account or posts. The post drew 2.6M views. A companion table — X's own disclosed ranking weights — is the part builders, marketers, and anyone running an agent that posts to X actually need to read.

The headline number: a single Report costs a post roughly 468 likes worth of visibility. A reply on your own post from someone who follows you back is worth as much as sharing the post via copy link — 40x a plain like.

TL;DR: What's actually new on August 13

What X actually announced

@XOpenSource's post, in full:

Open-sourcing the For You timeline — Today, we are taking another major step in our ongoing efforts to increase transparency. We're open-sourcing the code that affects a post's visibility in the For You timeline, and releasing a new tool that shows people labels applied to their account or posts that might limit visibility... People will be able to check for themselves via a new 'Under the Hood' page – if you've posted 10+ times in the prior month, you'll be able to see the labels applied to your account and posts in that month, and download the data. We're rolling this out as a pilot... Initially it will be available to a randomized test group of eligible accounts that are at least one year old... Follow the work here: github.com/xai-org/x-algorithm.

Two distinct things shipped in that one post, and it's worth keeping them separate:

1. Ranking code — the specific logic and weight configuration that decides whether a post you make actually reaches the For You timeline, not just the general architecture explainx.ai covered back in May.
2. Under the Hood — a user-facing product, not a repo. It's the first time X is showing individual accounts why a label might be suppressing their reach, and letting them download that history.

This is narrower than July's "entire codebase, no exceptions" pledge from Musk — there's still no DMs, ads auction, or moderation backend here — but it's more actionable than either prior release, because it comes with real numbers attached.

The ranking weights: what actually moves your reach

The disclosed weights are relative to a "like" baseline of 1.0x (raw value 0.5 in X's blend). Everything else is scored against that.

Positive signals — what boosts visibility

Tracked but carrying zero weight in the current blend: binary dwell (predicted, not used), profile click (predicted, not ranked), the mutual-follow dwell boost (tested, never shipped), and bookmark (in the action taxonomy and your viewer history, but no ranking weight at all today).

That last one is the most counterintuitive line in the whole disclosure. Bookmarking has always felt like a "save for later, this is good" signal — the code says X's own ranker currently ignores it.

Negative signals — what tanks visibility

A single report doesn't just remove a like's worth of goodwill — it's designed to overwhelm hundreds of positive signals at once. Ten reports on a post outweigh 4,680 likes. That's a deliberate asymmetry: X's ranker treats a report as a much stronger, much rarer, and much more trustworthy negative signal than a mute or a block, which are cheaper actions users take more casually.

What surprises builders and growth people

Three things in this table cut against common X growth advice:

1. A copy-link share beats a repost by 20x. Reposting has been the go-to "get seen" ask for years. The weight table says a plain repost (1.0) is worth a tenth of someone copying your post's URL and sharing it elsewhere (20.0). The algorithm is rewarding distribution evidence — a link only gets copied if someone intends to actually send it somewhere — over the cheaper, more automatable repost click.

2. Replies from people who already follow you back on your own posts are worth as much as the single best action on the whole list. The 15-point mutual-follow boost on top of the 5-point reply base means a reply thread with your existing engaged audience is scored identically to a copy-link share. This rewards creators who reply to their own posts and cultivate two-way-follow communities, not just raw follower count.

3. Bookmarks and profile clicks don't move ranking at all. If your growth strategy leans on "bookmark this thread" CTAs, the disclosed weights say that's not doing anything for reach — it may help the individual reader, but it's invisible to the ranker.

4. Clicks, link-opens, and media expands are all worth less than a fifth of a like. Passive consumption signals (0.05–0.4) are deliberately weak compared to active response signals (2.0+). This lines up with the reply-over-likes hierarchy explainx.ai flagged in the May algorithm breakdown — X's ranker keeps favoring conversation over consumption, and now discloses exactly how much.

How to check your own labels via Under the Hood

Under the Hood is not live for every account yet. Per the announcement:

• Eligibility today: account age of 1+ year, and you're inside the randomized pilot cohort — there's no opt-in button yet.
• Posting threshold: you need 10+ posts in the prior month for the tool to show data on your account.
• What it shows: the specific labels applied to your account or individual posts that may have limited visibility during that month.
• Data export: a download option for your label history, so you can track changes over time rather than relying on a single snapshot.

If your account isn't in the pilot yet, the practical move is watching github.com/xai-org/x-algorithm for the label-definition code, since that repo is where the taxonomy behind Under the Hood's labels will be documented as the rollout widens.

Nobody else is doing this — and X is happy to say so

The loudest reply on the announcement thread, from Joman (@jomangblandino) at 96 reactions, tagged Meta, YouTube, and TikTok directly:

@Meta @YouTube @tiktok_us Y'all don't have the balls to open source your algorithms like this.

It's a fair jab. None of Meta's Feed ranker, YouTube's recommendation system, or TikTok's For You algorithm publish anything close to a numeric weight table — let alone a self-serve tool showing individual users why their own content got suppressed. Whatever the business motive behind X's transparency push (and explainx.ai flagged in July's coverage that a GitHub repo is a snapshot, not a live guarantee), it's currently the only major social platform giving builders and creators actual numbers to work from instead of folklore.

What this means if you build on X, post via API, or run bots and agents

If you're building anything that posts to X — a marketing agent, a growth bot, an agentic pipeline that shares content automatically — these weights are now a spec, not a guess:

• Optimize prompts and playbooks toward replies and link-copies, not likes or reposts. If you're running an AI marketing agent that scores content performance, weight replies and copy-link shares far above like counts in your success metrics — the platform itself does.
• Treat "Report" risk as existential, not incremental. A single report from a real user can offset hundreds of likes. Any agent posting autonomously to X via the API needs a human-review or rate-limit gate before publishing anything remotely likely to get reported — the downside is asymmetric by design.
• Don't rely on bookmarks as a growth signal in dashboards. If a bot or content-agent pipeline is tracking bookmark counts as a leading indicator of reach, the disclosed weights say to stop — it's not in the ranking blend.
• Mutual-follow communities compound. Agents or creators cultivating a smaller, mutual-follow audience that replies get a 40x multiplier on those replies — that's a better ROI target than chasing raw follower growth for one-way audiences.
• Grok and xAI tooling sit next to this stack. If you're already using X's hosted MCP servers or Grok's voice agent builder to publish content programmatically, the same xai-org/x-algorithm repo now documents the exact scoring your agent's output is judged against.

FAQ

Q: Is this the same as the May 2026 algorithm release? No. May open-sourced the For You ranking architecture for the first time in this era. August discloses the specific numeric weights inside that architecture — a much more actionable level of detail.

Q: Does this replace July's "entire codebase" pledge? No, it's narrower. July was Musk's promise to eventually publish X's whole platform, pending a security review. August's release is scoped to visibility-affecting code and a new transparency tool — a concrete deliverable, not the full pledge.

Q: Can I see my own visibility labels today? Only if you're in the randomized pilot: account 1+ year old, 10+ posts in the prior month. X hasn't announced a general rollout date.

Q: What's the single biggest mistake based on these weights? Treating likes and reposts as the goal. Both are near the bottom of the positive scale (1x and 2x respectively) — replies, copy-link shares, and quotes all outrank them by 5–20x.

Related reading

• X's May 2026 algorithm open-source release
• X's July 2026 pledge to open-source the entire codebase
• X hosted MCP servers: Cursor, Claude, Grok connect to the X API
• Grok AI, viral posts, and X's trending engine
• xAI Grok Voice Agent Builder: no-code voice agents
• Build AI marketing agents with Claude: complete tutorial
• BirdClaw: local-first X workspace without the algorithm

Sources: @XOpenSource announcement (Aug 13, 2026) · xai-org/x-algorithm · Reply from @jomangblandino

Ranking weight values, pilot eligibility criteria, and rollout details reflect X's public disclosure as of August 13-14, 2026 and are subject to change as the algorithm is retrained and Under the Hood expands beyond the pilot cohort.