Phoenix — xAI's ranker for the For You feed — just shipped its biggest update since launch. It scores every post per viewer against 15 actions. Some lift you. Some bury you. Below: what's changed, what's costing Emily reach right now, and the 7 moves that turn the new system into a tailwind.
For every post + every viewer, Phoenix outputs 15 action probabilities, weights them, sums them. Mutes, blocks, and reports subtract — and a "viral but polarizing" post can net negative.
Phoenix barely reads post text. It identifies accounts by a hash of the engagement around them. Posting a different topic doesn't reset your fingerprint — sustained behavior change does.
The same author appearing repeatedly in a session is attenuated. 8 posts spread across the day score higher than 15 in a 90-minute burst. Cadence is now a real algorithmic input.
Years of political/Israel posting plus the Grammys clip have trained Phoenix to associate her account with a conservative-leaning, often-older cluster. That's a ceiling.
Replies are positive, but the mutes/blocks her posts attract are now a meaningful tax. Accounts with similar positions and less personal-attack energy outperform her on equivalent topics.
The Eilish/Grammy spike likely also generated a "not interested / mute / report" wave. Phoenix logs that against future distribution. The viral moment cost her reach.
On-camera training, NBA podcast access, Israeli UN mission, Modern Orthodox identity — the things that make her non-interchangeable barely appear in her feed. Phoenix pays a premium for distinctness.
Phoenix's heaviest positive weights — video_view, dwell, photo_expand, profile_click — are all video-side. Emily has on-camera training most political accounts don't.
A 30s video that holds attention registers far more positive actions per impression than a tweet with equivalent engagement — and video has a lower mute rate than text-based political commentary.
Goal: widen the audience fingerprint without confusing it. Phoenix handles multiple lanes from one account — if the pattern holds long enough to learn.
Posts that combine two of her lanes in a single thought. They earn engagement from multiple clusters at once, which widens the fingerprint faster than any single-lane post can. Strongest bridge: sports → identity.
Sports audiences are politically heterogeneous in a way political audiences aren't. The same logic works for Israel × media criticism, religion × being young in public, UN job × absurdity of public diplomacy.
Reaction posts attract the same cluster that engaged with the original (reinforcing the fingerprint), carry the original's negative-reaction spillover, and lock to a news cycle already routed through other accounts.
"This is disgusting. They never would have said this about —"
"I'm noticing something. The same people who say —"
"Agree? Thoughts? RT if you agree" is recognized as a pattern and likely triggers "not interested" flags. Replacement: prompts that ask for something specific the reader actually wants to share.
These are identity hooks — the kind of post readers screenshot and quote-tweet with their own answer. Quote-posts are the heaviest reader-side positive signal Phoenix tracks.
The author-diversity scorer penalizes bursts. Replies into larger accounts re-enter the engagement histories of users who don't follow her yet — the fastest way to broaden the fingerprint. What changes is the shape of her posting, not the count.
At Emily's scale, in-network reach is saturated — out-of-network is the actual growth lever. Phoenix has a dedicated OON scorer that rewards a specific kind of content.
Personal-credential openings drive profile clicks (the credential prompts curiosity) and self-contained context lets the post travel beyond her follower graph.
Engagement will feel flat. The fingerprint hasn't moved yet. This is normal — don't panic-revert.
Watch who is replying, not how many. A shift in commenter identity is the leading indicator.
This is the phase where the new shape outperforms the old — including raw engagement.
The Eilish/Grammy archetype: short-term spike, long-term distribution cost via negative-action tagging.
15 posts in 90 minutes is mathematically worse than 8 across 12 hours. The diversity scorer is literal.
The most crowded category on X. Phoenix punishes interchangeability harder than the old system did.
Controversial is a feature anyone can copy. Distinct is the intersection of identity, expertise, and voice only she occupies.
Sports reporter. On-camera trained. Modern Orthodox. Israeli UN mission. Public political stance. NBA podcast host. Young woman in male-dominated rooms. No one else occupies that intersection — and Phoenix pays a premium for non-interchangeable accounts. The work isn't to make her account more careful. It's to make it more clearly the thing only she can be, more often, in the formats the model is built to score.