A fresh read of X's recommendation code — the Grok-powered model it calls Phoenix — surfaced the point weights behind the For You feed on August 17, 2026. The share is the top signal, a like the weakest, and the penalties are applied per user, not platform-wide.
2026-08-19 · by Moe Ameen
On August 17, 2026, Social Media Today highlighted a fresh analysis of X's open-sourced recommendation code — reported by Business Insider's Truman Dickerson — that attaches hard numbers to how X's current ranking model values each type of engagement. That model, the Grok-powered system X calls Phoenix, replaced the platform's older hand-tuned heuristics earlier in 2026. Rather than following fixed rules, it predicts how likely a given user is to take more than a dozen different actions on a post — like, reply, repost, share, follow, mute, block, report, dwell, video completion, and more — then scores and ranks each post from those predictions.
The point weights that surfaced tell a clear story. A share of a post's URL — copying the link and passing it along, on or off X — is worth roughly 20 points, about 40 times the value of a like. A reply, a quote, or an internal DM share sits around 5 points each; a follow the post earns is about 4; a repost roughly 1; and a like just 0.5, the weakest positive signal in the set. The negatives dwarf all of them: a report is about -234 points, a mute -59, a "not interested" tap -43, and a block -31. The important nuance is that those penalties are applied per user, not globally — one person muting or reporting you damages your reach with that person, not across the whole platform.
The reading X analysts drew is that Phoenix is built to reward content people actively pass along — especially content shared into other formats and off the platform — and that it has quietly demoted the like as a value signal. The practical advice that follows is to study what a specific audience will genuinely forward to their own networks, rather than optimizing for passive taps. One honesty note: X re-tunes Phoenix continually and does not publish every current weight, so treat the exact point values as a directional read of the code rather than fixed, permanent rules.
The single most useful thing in this data is that the share outranks everything else, and that is a production problem, not a posting trick. You cannot make a like-bait post score like a share; you have to make things people actually want to pass on, in the format each audience shares. Point [Kompozy](/) at one source — a talk, a client call, a strong opinion — and, governed by your [Persona Brief](/glossary/persona-brief), it produces the shareable shapes Phoenix now rewards: Clipped Shorts and [Persona Shorts](/glossary/persona-shorts) people re-share as native video, brand-exact [Carousel Posts and quote graphics](/glossary/output-buckets) built to be screenshotted, and [Persona Tweets](/glossary/persona-tweet) written to earn quotes and replies instead of passive taps.
Because the highest-weighted action is redistribution across formats and off-platform, the winning move is to put the same idea in front of every audience in the shape they'll forward — which is exactly what Kompozy's multi-format generation plus [Autopilot](/glossary/autopilot) scheduling across the eight social platforms plus blog and email does, from one queue behind a per-post review gate. Keep replying and quoting yourself, because that human conversation is a signal no tool should fake. Let Kompozy handle turning one idea into the dozen shareable assets that give the [algorithm](/glossary/algorithm) something to reward — on X and everywhere else at once.
According to analysis of X's open-sourced Phoenix ranking code highlighted on August 17, 2026, the highest-weighted positive action is sharing a post's URL — copying the link and redistributing it, including off-platform — at roughly 20 points, about 40 times a like. Replies, quotes, and DM shares (~5 points) and follows (~4) also outweigh likes, which sit at just 0.5. The clear ordering is shares over replies over reposts over likes.
Phoenix is the Grok-powered ranking model behind X's For You feed, which replaced the platform's older hand-tuned heuristics earlier in 2026. Instead of fixed rules, it predicts how likely a user is to take more than a dozen different actions on a post — like, reply, repost, share, follow, mute, block, report, dwell, video completion, and more — and scores each post from those predictions.
Not globally. The analysis shows negative actions carry heavy weights (a report about -234, a mute -59, "not interested" -43, a block -31), but they are applied per user — one person's report or block reduces your reach with that person, not across the whole platform. That said, a consistent pattern of negative reactions from many viewers still adds up against a post.
No. X re-tunes Phoenix continually and does not publish every current weight, so the specific numbers are best treated as a directional read of the open-sourced code rather than fixed rules. What stays stable is the ordering — shares outrank replies, which outrank reposts, which outrank likes — and the emphasis on content people actively redistribute.