Bangermeter — X Algorithm Scorer

Scores posts with X's published For You ranking weights and the real scorer arithmetic. Every factor cited to source.

See what X's algorithm actually rewards — on every post, as you scroll. HOW TO USE IT 1. Scroll x.com. Every post picks up a badge with two scores, 0–100. 50 is a typical post. 2. Click the badge. You get a plain-English breakdown of why it scored that way. 3. Start typing a post. A live meter scores your draft before you send it. That is the whole thing. Everything below is why you can trust the numbers. WHERE THE NUMBERS COME FROM On August 13, 2026, X published the real ranking weights for the For You timeline: the exact numbers that decide how far a post travels. Bangermeter puts them on every post in your feed. No leaked screenshots. No "I reverse-engineered the algorithm" threads. The weights come straight from X's published source, and every factor in the breakdown is cited to the line of code it came from. WHAT YOU GET • A score badge on every post as you scroll. Two numbers, 0–100, where 50 is a typical post. – Content Appeal: what the ranking signals would predict from the content alone. – Earned Engagement: how the weights value what the post actually earned, per view. • Click the badge for a plain-English breakdown. Which signals fired and why, what each reply and repost is actually worth, the rescoring factors that applied, and how fresh the post is. Every section has a "Show the math" expander with the full weight × probability table if you want the numbers. • A live meter in the compose box. Score your draft as you type, with hints as they trigger — asks a question, substantive text, bare link, hashtag pile, engagement bait. • A weight table in the popup. All 29 ranking heads, their published values, and which ones the extension can measure versus estimate. • Reply mode. Reply drafts are scored as replies (they take a ×0.75 factor in the published math), and the breakdown explains what X publishes about replies: out-of-network replies never reach For You at all, replies never get the mutual-follow boost, and replies to accounts over 250K followers (as of 30 September) are quality-scored by a Grok model whose rubric X withholds. • Draft comparison. Save up to three variants of a draft and pick the best-scoring one before posting. Variants live in memory only. • Local score history. Opt in from the popup and opening a breakdown logs the scores, the post's link and a short text snippet to a list in your browser, so you can see which formats score higher over time. Off by default, capped, local-only, one-click clear. • Under the Hood import. In X's transparency pilot? Import the report you download from x.com/i/jf/under_the_hood and see the visibility labels X itself applied — parsed locally, stored locally. • A posting-cadence note. Post twice into the same stretch of feed and the panel shows the author-diversity attenuation production applies (second post ×0.625, decaying to a ×0.25 floor). SOME OF WHAT THE WEIGHTS SAY • Copying a post's link carries 20, the heaviest positive coefficient X publishes. It is also one of the rarest actions a reader takes, which is part of why the number is so large. • A reply carries 5. From someone you mutually follow, on an original post, 20. • Sharing to DMs carries 5, and nothing in the visible counts shows you it happened. • Profile clicks are worth nothing. That head used to pay, and X zeroed it. • Muting hurts nearly twice as much as blocking. • Scrolling straight past is a scored penalty. • Bookmarks are not a scored action at all. • Links do get credit. The "link penalty" was always folklore. HOW IT'S HONEST ABOUT WHAT IT KNOWS The weights are X's. The probabilities are estimates. X predicts how likely you are to reply, share or mute a post using a model nobody outside the company has. Bangermeter derives three of those probabilities from real counts — likes, replies and reposts — and estimates the rest from what's in the post. Every head in the popup is labeled as measured, estimated, or zeroed by X, so you always know which is which. It also matters that these weights multiply a predicted probability, not a count. A report is not a 234-point deduction; that number is the coefficient on how likely a viewer is to report you. Anything claiming these are per-action point totals has it wrong. There are no invented numbers anywhere in this extension. Circulating figures that failed source-tracing — "bookmark 20x", "links minus 30-50%", "block minus 120" — were left out rather than guessed at, and none of them matched the real values when those were published. This is a relative score against a typical post. It is not a prediction of reach, and nothing can honestly promise you one. PRIVACY • No data collection. None. • No network requests. The extension never phones home, because there is no home. • No account, no sign-in, no analytics, no tracking. • The only permission is storage: your settings sync with your Chrome profile, and the optional score history and imported Under the Hood report stay in this browser's local storage. Nothing is transmitted anywhere, and both are one click to delete. Everything runs locally in your browser. The source is public and reviewable. OPEN SOURCE MIT licensed, with the full research record and 362 automated tests that check every published weight against X's source individually. github.com/itsryanlenk/bangermeter WHAT'S NEW IN 0.10.4 • Five weights changed in X's config between its 6 and 13 October syncs, and this release catches up. Likes went from 0.5 to 0.66 in two steps: 0.6 in the 12 October sync (published 13 October), then 0.66 in the 13 October sync (published 14 October). Video open fell from 0.07 to 0.02 and dwell time rose from 0.004 to 0.006 per second in the 12 October sync (published 13 October). Post-unexplored fell from 0.02 to 0.015 in the 6 October sync (published 7 October). And "not interested" got another 10% harsher, −47.52 to −52.272, in the 13 October sync (published 14 October). • The coefficients moved relative to each other. The reply coefficient is now about 7.6 times the like coefficient, down from 10, and about 30 times on an original post from a mutual, where X adds 15 to the reply weight (down from 40). The repost coefficient is about 1.5 times the like coefficient, down from 2. These multiply X's predicted chance of each action, not counts, so as X's own code comment warns, they do not mean a reply is worth 7.6 likes. • Which numbers this follows, and why. X publishes its weights in two files, and on 16 October they disagree on those five: home-mixer/params/param.rs has the new values, and vm-ranker/params.rs, the scorer's own file, still lists the old ones. The scorer polls the same live config file that param.rs is copied from, and both files use the same setting names, so the extension follows param.rs. How the scorer handles a setting missing from that config lives in a library X has not published. Three of the four new heads below, and the post-unexplored switch, appear only in vm-ranker/params.rs, so the extension takes those from there. X does not publish its per-user experiments or overrides, so no outside tool can see those. • Four new heads, all at 0, so no score moves: video continuation, user video continuation and profile visit time (5 October sync, published 6 October), and react with video (12 October sync, published 13 October). Two video-carousel weights also appeared in the 14 October sync (published 15 October). They order the video carousel, not For You, and the carousel is off by default as of the 16 October sync, so the extension does not score them or list them. • Smaller corrections. Post-unexplored is in-network by default, but since the 5 October sync (published 6 October) a switch can extend it out of network (it ships off); the breakdown used to call it hard-wired. And the net-negative ceiling is now 0.000897, from 0.000896, as of the 13 October sync (published 14 October). The Under the Hood import now recognizes BRAZIL_ELECTION_LEGAL, a label X added to its public list on 2 October ("In Brazil, post hidden from recommendations"). And the "Visibility limited" details are re-checked against X's 17 October code: X's published code still usually removes these posts from For You, as before. X's newer notice rules, published 6 and 8 October, sit at a level its published For You code does not use.