In October 2025, Elon Musk announced that X would rip out its entire feed-ranking system and hand the job to Grok. Not a tweak. A full replacement of the rule-based “if this, then that” heuristics that have run every major social feed for a decade, in favour of an AI that reads and watches every post directly. Musk said the platform was “aiming for deletion of all heuristics within 4 to 6 weeks,” with Grok processing over 100 million posts and videos a day.
That was nine months ago. It’s not a rumour or a plan anymore: it happened, mostly on schedule, and it’s had a bumpy, well-documented run since. If you’re a creator on X, the version of this story worth knowing isn’t the announcement. It’s what shipped, what broke, and what changed about how visibility and pay actually work today.
It shipped, with a catch nobody flagged at the time
On January 19, 2026, X open-sourced the new recommendation code on GitHub, confirming the system now runs almost entirely on a Grok-based Transformer model rather than hand-tuned rules. No more manually weighted rules like if engagement passes a threshold, boost the post. Instead, a single model ingests a user’s behaviour history and outputs a relevance score.
But developers who dug into the release found something the October announcement didn’t mention: paid verification is wired directly into the base score. Verified accounts (X Premium, $3/month and up) start with a ceiling around +100. Unverified accounts are capped at around +55 before content quality even comes into play. In other words, the “new account problem” (posting something great and nobody seeing it) didn’t disappear. It just changed shape: instead of needing a large following, you now need a subscription to compete on equal footing. That undercuts the original pitch that visibility would hinge solely on the content itself.
Critics also noted the code release itself was incomplete: it shows how signals flow through the system but withholds the actual weighting constants, prompting some to call it “transparency theatre.” No outside analyst can fully reconstruct how much a like, reply, or dwell-second is actually worth.
Reply farming killed the cheat code
The original pitch implied Grok’s semantic reading would simply reward substance over engagement bait. What actually happened is narrower and more interesting: X discovered that close to half of its reply spam was coming from people coordinating in outside Telegram groups specifically to farm the revenue-share program. In response, Bier announced that payouts would shift to “Verified Home Timeline impressions” rather than reply volume, and that low-effort replies (generic thanks, emoji strings, templated responses) would no longer count toward earnings even if they technically boosted engagement numbers.
The algorithm does appear to weight dwell time (how long someone actually pauses on a post) and shares-via-DM as stronger quality signals than raw reply counts. So the “reply guy” growth hack that’s worked on X for years is genuinely less effective now, but it was killed mainly to stop a specific fraud pattern, not because Grok developed some abstract taste for substance.
Payment chaos played out in public
The idea that X’s underpayment problem is a quiet, unaddressed issue doesn’t hold up. It’s been litigated publicly, repeatedly, over months:
- In late December 2025, Musk responded to a creator’s call to raise payouts with “Ok, let’s do it,” publicly looping in Bier.
- In January 2026, X formalized that shift, moving payouts to verified, timeline-based impressions rather than raw reply counts.
- On March 25, 2026, X announced another revenue-sharing overhaul (this one prioritizing impressions from a creator’s home region), and it landed so badly with the creator community that Musk paused the new structure within six hours of its launch.
That’s not stagnation. It’s a public commitment to raise payouts followed by two formula changes in three months, one of which had to be reversed almost immediately. Worth knowing if you’re trying to plan around X as an income source: the rules have genuinely not settled, and probably won’t for a while.
Human-defined quality is no longer theoretical
The idea that an algorithm’s sense of “quality” reflects the choices of whoever trained it isn’t abstract anymore. Since Grok took over ranking duties, it’s also been at the center of several serious incidents: a well-documented episode where the chatbot began generating antisemitic content and referred to itself as “MechaHitler,” a bug that made it repeatedly bring up a South African conspiracy theory in conversations that had nothing to do with the topic, and, separately, a surge in AI-generated non-consensual sexualized images that drew investigations from California’s Attorney General and the UK’s Ofcom, along with regional restrictions in Malaysia and Indonesia. X has since restricted some of Grok’s image-editing features in response.
None of that is a direct statement about how the ranking algorithm treats your posts. But it’s concrete evidence for the underlying point: whatever “quality” means to Grok is a product of specific engineering and moderation choices made under public and regulatory pressure, and that change over time, not some neutral, discovered truth about good content.
What the algorithm rewards as of mid-2026
Independent analysis of the system over the spring points to a fairly consistent pattern: the ranking now favours engagement depth (long, substantive replies, time spent actually reading a post, conversations that continue) from accounts that behave like real, consistent humans, within a recognizable niche topic. A February 2026 API restriction and a March 2026 enforcement wave specifically targeted bot-like posting patterns: identical phrasing, mechanical timing, mass actions. The detection is behavioural, not about the topic itself, so it catches cheap automation while missing well-paced, varied activity.
Practical takeaways, updated
- Verification is now a structural floor, not just a nice-to-have. If you’re not verified, you’re capped before content quality is even considered.
- Stop optimizing for reply volume. It no longer pays, and low-effort replies can now actively work against you.
- Consistency of topic matters as a ranking input, not just a branding choice: accounts that scatter across subjects appear to reset their own classification.
- Expect more change, not less. The payout formula has already shifted twice since January, and one of those changes was reversed within hours. Build for a platform that’s still actively rewriting its own rules, not one that’s settled into a final state.
- Don’t assume “AI quality scoring” means neutral or apolitical. The same system that makes distribution decisions is built by a company that has had to walk back some of its AI’s outputs under regulatory pressure.
The October 2025 headline claiming that X would replace engagement heuristics with an AI that reads every post proved true. What nobody could have written accurately back then is how messy, political, and still unfinished that transition would look nine months in.