Your X Playbook Is Running on 2023 Numbers. The Algorithm Changed Its Brain.

TL;DR: In January 2026, X replaced its For You ranking system with a new one, Phoenix, and in August it published the current production weights in its open code repository. The numbers every growth playbook still quotes — a reply is worth 13.5 likes, an author reply is worth 75, a profile click is worth 12 — belong to the system that was switched off. In the current ledger, a reply scores 5, a profile click scores 0, and the 75-point author-reply bonus no longer exists. The most valuable action is someone copying your link to send privately, at 20. I should note the source honestly: these figures come from two independent readers who cloned and verified the public code file by file, not from my own clone — the values agree across both, and the structural facts below do not depend on any single number.

Start with the reply farm, because it is the most expensive habit the old table created. Under the 2023 weights, flooding replies under large accounts looked rational: each reply carried the second-highest value in the system. Under Phoenix, a reply is 5 — still useful, no longer dominant — and, more importantly, the For You feed by default filters out replies from accounts the reader does not follow. A stranger mostly sees your reply only if they open the original post's thread. That is rented attention in someone else's room, not distribution. It can be worth doing deliberately. It is not a growth engine, and counting replies as if they were reach is accounting fraud against yourself.

The profile funnel died even more quietly. A profile click went from 12 to exactly 0. Not reduced — zeroed. Every post designed to tease people into visiting a profile in order to convert them is now optimizing for an action the ranking system does not reward at all. If a post is good, the reward the system actually pays for is different: a follow, scored at 4, happens on the post itself, or it does not happen.

What the new ledger actually pays for

The highest single weight, 20, belongs to the least gameable action: someone copying your link to share it privately. You cannot farm that in a group chat — the code's own comments note that coordinated engagement has no ranking effect, because weights multiply predicted probabilities, not counts. A weight is not an exchange rate. It is a lever on a prediction about one specific reader. That is also why every "one reply equals N likes" conversion table, old or new, is a misreading: there is no conversion, only probabilities.

The one large lever a small account can actually build is conditional. When two accounts follow each other and one of them replies under the other's original post — not a reply, not a repost, an original — the reply weight rises from 5 to 20, with a +15 mutual-follow boost. The sequence is written into the code: first become mutuals with peers in your niche, then have real conversations under their original posts. Replying to strangers earns no boost at all.

Two more mechanisms punish the volume playbook directly. Posts from the same author in one feed load are discounted: the second post keeps 62.5 percent of its score, the third 44 percent. And a small account's cold start is a narrow window: only original posts from authors under 50,000 followers, under two hours old, with fewer than 200 home impressions, are eligible to be tested on strangers, and the test is driven by like rate in those first two hours. A post's life is then hard-cut at 48 hours. Volume does not widen that window. It splits it.

What I changed the same day

I run a personal AI operation for one person — myself — and I post about building it. Two days ago I cut my own X cadence to at most one original post a day, five in any rolling seven days, on instinct. The code turned out to agree: a second post the same day is born at three-fifths strength. So the cadence stays, and now it has a mechanism, not just a preference.

The rest of the playbook changed with it. Replies and profile clicks are deleted as KPIs; I track mutual follows gained, replies under mutuals' original posts, and follows earned instead. I am building a short list of 15 to 25 peers in standards, AI infrastructure, and founding — people worth following first — because the mutual boost only pays on original posts, and only between mutuals. Every post now has to answer one question before it ships: who would copy this link, and to whom? And one red line is now explicit: in July 2026, X removed 42,000 accounts for chatbot auto-replies, and its automation rules ban keyword-triggered bulk replies. My replies are drafted with AI help and sent one by one, by hand. Farming replies was already a bad trade. Automated, it is an account risk.

The table will change again. The structure is the durable part: private hand-offs are worth most, mutuals beat strangers, originals beat volume, and counts are not currency. Build for the structure, and the next weight change is a footnote. Build for the 2023 numbers, and you are sending medicine to a system that no longer exists.

— Jiahui Miao, founder of vertciti, building a personal AI operating system for one.