The For You feed isn't a mystery box — it's a pipeline with known stages and measurable preferences. Here's how candidate sourcing, ranking and filters actually work in 2026, which signals carry real weight, what reliably kills reach, and how to turn all of it into publishing decisions instead of folklore.
Every For You feed is built in three passes. First, candidate sourcing: X pulls roughly half the pool from accounts you follow and half from out-of-network posts surfaced by engagement graphs and similarity. Second, ranking: a model scores each candidate on predicted engagement — will this person reply, dwell, repost? Third, filters: dedupe, author diversity, and safety rules trim the list. Your job as a creator is simple to state: survive sourcing, score high in ranking, and never trip a filter.
Not all engagement is equal. Early velocity leads: the first 30–60 minutes decide whether a post gets pushed to wider out-of-network pools. Replies outweigh likes by a large margin — a conversation is a far stronger prediction signal than a tap. Dwell timecounts even without a click: a reader pausing on your post tells the model it held attention. Profile visits and follows from a post rank near the top too. Likes matter, but they're the cheapest signal on the list — optimize for the expensive ones.
X is pushing hard into video, and the ranking reflects it: native uploads get their own surfaces (the video tab, immersive feeds) and favorable treatment in For You. Dwell time on video is easy for the model to measure — watch seconds are unambiguous — which is exactly why it's rewarded. This doesn't mean every post needs a clip. It means when you have visual material, upload it natively rather than linking out to YouTube, and front-load the first three seconds the way you'd front-load a hook.
Since Grok began annotating posts, tone is a ranking input, not just a vibe. Posts classified as informative, constructive, or genuinely funny get a tailwind; posts classified as rage-bait, dunking, or manufactured outrage get throttled — they still earn replies, but those replies count for less. This closed the oldest loophole on the platform: picking fights used to be free reach. It isn't anymore. The practical rule: write posts you'd be comfortable having read back to you, because a model effectively is.
X Premium gives a real but modest ranking boost — meaningful in-network, smaller out-of-network — plus reply prioritization, longer posts, and eligibility for revenue sharing. It is a multiplier on content the model already likes, not a substitute for it: a verified account posting weak material still stalls. If you're posting seriously and daily, the boost compounds enough to justify the subscription. If you post twice a month, it changes almost nothing. Buy it when your volume deserves the multiplier.
Three reliable reach killers. External links in the first post: X deprioritizes posts that send readers off-platform — put the link in a reply, or use your profile. Engagement bait: “repost if you agree” and follow-for-follow patterns are explicitly detected and demoted. Dead-hour posting: because early velocity gates everything downstream, publishing when your audience is asleep means the post fails its first test before anyone good sees it. None of these are shadowbans — they're the scoring working as designed.
Most “algorithm tips” circulating on X are folklore — someone's single lucky post generalized into a law. Your last 60 days of posts are a better guide than any thread about leaked code. Pull them and ask: which formats earned replies, not just likes? Which hours produced fast early engagement? Which posts drove profile visits and follows? That's the ranking model's opinion of your account, expressed in numbers. (For the timing half of that analysis, see our best-time guide.)
Every specific weight in this guide will drift; X ships ranking changes constantly and announces almost none of them. What doesn't drift is the direction: the model keeps getting better at rewarding posts people genuinely want to read and discuss. So build a system that's robust to re-weighting — write for replies and dwell, post into your proven hours, keep links out of the first post, review your numbers monthly — and let the trick-chasers rebuild their playbook every quarter while yours keeps working.
GrowthX predicts each draft's like, retweet and view range calibrated on your own last 60 days, finds your proven posting windows, and schedules to the second — so every post takes its first velocity test at full strength.
Try GrowthX →