- Replies are worth roughly an order of magnitude more than likes, and a reply the author engages with is worth several times more again.
- Negative signals (mutes, blocks, "show less often", reports) are catastrophically expensive. One burned reader outweighs dozens of likes.
- The ranker rewards conversation quality per impression, so the winning strategy is writing people want to respond to, not engagement bait.
In 2023 X open-sourced its recommendation system, including the "heavy ranker" model that scores candidate posts for the For You feed. The code that runs in 2026 is not identical to that snapshot, but the incentive structure it revealed has stayed directionally stable, and it is still the best public evidence for one question: what is a post actually rewarded for?
The pipeline in one paragraph#
When someone opens the For You feed, X gathers candidate posts from people they follow and from out-of-network sources (engagement graphs, similarity models), scores each candidate with a neural ranker that predicts the probability of various engagements, multiplies those probabilities by hand-set weights, and sorts. Filters then trim the list (author diversity, seen-before, feedback fatigue). Your post's reach is mostly decided by that weighted score, computed fresh for every viewer.
What the weights revealed#
The open-sourced weighting made the platform's priorities unusually legible. Directionally, and confirmed by years of practitioner testing since:
- A like is the baseline unit. Positive, cheap, low weight.
- A repost is worth a couple of likes. Distribution begets distribution.
- A reply is worth roughly an order of magnitude more than a like. The feed is optimised for conversation, not applause.
- A reply that the author then engages with is one of the most valuable positive events in the system, worth many replies on its own. The ranker treats a real back-and-forth as the strongest evidence of a post worth showing.
- Profile visits and long dwell time count. A post interesting enough that people click through to see who wrote it gets credit for it.
- Negative feedback is brutally expensive. "Show less often", mutes, blocks, and reports carry penalties that dwarf the positive weights. A post that earns 50 likes and a handful of mutes can net out worse than a post with 10 likes and no complaints.
What this means for how you write#
You do not need growth hacks. You need to internalise three consequences of the weighting.
1. Write posts people can respond to#
A perfect, sealed, nothing-left-to-say post earns likes. An opinionated, specific, slightly incomplete post earns replies, and replies are what the ranker pays for. Practical versions: take a position instead of surveying all positions, show your work so people can disagree with a step, ask a real question you actually want answers to (not "thoughts?").
2. Stay for the conversation#
Author-engaged replies being weighted so heavily means the hour after posting is part of the job. Reply to the good responses, quote the best one, answer the objection. This is also why posting time matters less than posting when you can stick around.
3. Never trade a like for a mute#
Engagement bait works on the positive weights and dies on the negative ones. Rage bait, fake controversy, "unpopular opinion" that is actually the most popular opinion: each earns a burst of engagement and a slow bleed of mutes from the exact people you wanted to keep. The math says one burned reader costs you dozens of likes' worth of score, and unlike a like, a mute is permanent.
The boring conclusion that happens to be true#
Strip the weights down and the algorithm is paying for one thing: posts that make the feed feel like a place where interesting people talk. Sounding like a person is a ranking strategy. Generic AI-flavoured content fails here twice: it earns fewer replies because there is no person to respond to, and it earns more "show less often" because readers are increasingly allergic to it. If your drafts have that problem, start with the 8 tells that out a bot.
That is also the thesis Voxly is built on: the way to do well in the ranked feed is not to game the ranker, it is to publish writing that sounds like you and invites a response. The algorithm, for once, is on the writer's side.
FAQ
Is the X algorithm still the same as the open-sourced version?
Not exactly. The 2023 release is a snapshot and X has iterated heavily since, including replacing components with newer ML systems. But the incentive structure it revealed (conversation over passive likes, harsh penalties for negative feedback) has stayed directionally stable, and it remains the best public evidence we have.
Do links kill reach on X?
Posts whose entire job is to push people off-platform tend to underperform on engagement-per-impression, and that is enough to rank them lower. The workaround most writers use: deliver value in the post itself, then put the link in a reply or after the hook has done its work.
Do hashtags help in 2026?
Mostly no. Discovery on X is driven by engagement graphs and content understanding, not hashtags. One topical tag is harmless; a row of them reads as spam to both the ranker and the humans.