What we are working from
Everything below is grounded in what xAI actually publishes. Since 2026 the For You feed has been ranked by Phoenix, a Grok-based two-tower transformer that handles retrieval and ranking in one model, replacing the Heavy Ranker, SimClusters and RealGraph stack from 2023. The repository states the team eliminated every hand-engineered feature and most heuristics.
One caveat we will repeat wherever it matters: the numeric engagement weights are not public. They sit in a params module that has never shipped, in 2026 or in 2023.
Anyone quoting you an exact multiplier is repeating a leak or an inference. The ordering is reliable; the ratios are not. Full breakdown here.
1. A cold start
The most common cause, and the least discussed. Phoenix scores relative to your account's baseline, and an account that has been quiet has a cold audience, fewer people see the first wave, so early engagement arrives slowly, so distribution is capped before most followers ever get a chance.
Test: compare a post after a week of silence against one posted into an active week. If the gap is large, this is your problem, and it is not a content problem.
2. Leading with a link
Dwell time and on-platform engagement are among the nineteen signals the model predicts directly. A link is an instruction to leave, and the post gets evaluated on the behaviour it produces.
Fix: put the substance in the post and the link in the first reply. This costs nothing and is the single highest-yield change most accounts can make.
3. Nothing to reply to
Replies sit above likes in the ranking ordering. A post that is agreeable, complete, and uncontroversial gives nobody a reason to respond, so it collects a few likes and stops.
Fix: make one specific, checkable claim per post. Not manufactured controversy, which trips the negative heads, a real position someone could reasonably dispute.
4. Audience mismatch
Phoenix predicts reaction from a specific person. If your followers came from content unrelated to what you post now, the model has learned an association that works against you.
Test: look at whether your best-performing posts share a topic your recent posts have drifted away from. Follower count is irrelevant here; composition is everything.
5. Negative signals
Four of nineteen prediction heads are negative: not interested, block, mute, report. These are the cheapest way to damage an account and the least visible, nobody tells you they muted you.
Common causes: posting frequency well above what your audience signed up for, engagement bait, and identical copy repeated across accounts, which readers identify as inauthentic within seconds.
What to measure instead
Raw impressions are a poor diagnostic because they aggregate several different things. More useful:
- Median views per post over 30 days. One outlier makes an average meaningless.
- Reply rate as a share of impressions. This predicts future reach better than likes do.
- Profile visits per post. Signals the post made someone curious about you, which is what converts into followers.
- Your own baseline trend. Since ranking is relative, the only comparison that matters is against yourself last month.
Questions we get asked
Why are my views on X suddenly low?
The most common cause is a cold start: ranking is relative to your account's own baseline, so posting into a quiet period means fewer people see the first wave, early engagement arrives slowly, and distribution is capped before most of your followers see it. Other frequent causes are leading with an external link, posts that give nobody a reason to reply, and audience mismatch.
Do external links reduce views on X?
In practice yes, though not because links are penalised by rule. The 2026 ranker predicts dwell time and on-platform engagement directly, and a link is an instruction to leave the platform. Carrying the substance in the post and putting the link in a reply consistently performs better.
Is there a best time to post on X?
Less than commonly claimed. xAI removed hand-engineered features and most heuristics from the ranker in 2026, and the model scores relative to your own audience's behaviour rather than to a global clock. Posting when your specific audience is active matters; universal best-time charts do not.