Grok's 2026 Ranking Changes: What They Mean for X Replies

X's reply ranking is not the same system it was a year ago. In January 2026, X tore out its old recommendation engine and replaced it with a Grok-powered model, then adjusted that model again in July to push replies from people you actually know higher up the list. If a thread's "best" order has felt different lately, that is not you imagining things — the model choosing it changed twice this year.

The January rewrite: ranking runs on Grok now

For years, X's reply ranking leaned on a set of hand-built rules: weight likes here, discount bot-like accounts there, nudge verified replies up a little. In January 2026, X released a new recommendation system built on a Grok transformer model and published the underlying code. Instead of a checklist of separately tuned signals, the model is trained end-to-end to predict how likely you are to engage with a given reply, reading the reply itself rather than scoring it against a fixed list of rules.

The practical effect is less about any single reply and more about consistency. A rules-based system tends to reward the same kinds of replies every time — a viral joke, a fast dunk, a verified badge. A model trained on predicted engagement can weigh those same signals differently thread to thread, which is part of why the "best" order on two similar threads can now feel less predictable than it used to.

The July fix: mutuals move up the list

Six months later, X made a narrower change specifically to reply threads. Its product team said the ranking model had been missing a signal for "mutuals" — accounts that follow each other — which meant replies from strangers with high engagement routinely outranked replies from people you actually know. The July update corrected that, so a reply from someone who follows you and whom you follow gets weighted higher in the ranked order than it would have a few months earlier.

This is a narrower fix than the January rewrite, but it is the one most readers will notice day to day. A thread that used to surface the loudest strangers first can now show a friend's reply near the top instead, even if that reply has a fraction of the likes. The stated goal was to make busy reply threads feel less like a contest between strangers with the biggest following and more like an actual conversation among people who already know each other, which is a different aim than the January change was solving for.

What this means for the replies you actually see

Two things follow from these changes that are worth knowing as a reader, not just as someone curious about the mechanics:

• "Best" is less of a fixed formula now. A model trained on predicted engagement can weigh the same reply differently depending on the thread and the model's current state, whereas the old rules-based system was more consistent by design. Do not assume last month's intuition about what rises to the top still applies cleanly.
• Your own follow graph now matters more than it did. If you follow and are followed by a lot of the people replying on a given thread, you will likely see a noticeably different "best" order than someone reading the same thread who follows none of them. This was always somewhat true; the July update made it more pronounced.

Neither change touches the deeper mechanics covered in why some X replies look hidden or collapsed — quality filtering, mutes, and pagination all still work as a separate, earlier step before any ranking model runs.

Why two people can see a completely different order

Because the current model folds your follow graph directly into its prediction, the same thread genuinely renders differently for different readers now more than it used to. That was already true under the old system to a lesser degree, but a model tuned to reward mutuals makes the gap wider: two people opening the identical thread, one with a dozen mutual followers replying and one with none, can land on two "best" orders that barely overlap in their first few rows.

This is worth knowing mainly so you stop assuming a reply "should" be higher just because it looks obviously good to you. It might be ranked lower for you specifically because you and the replier are not mutuals, while the same reply sits near the top for someone else reading the same thread.

It also changes what "checking the best replies" is useful for. On a thread where you know a lot of the people replying, the ranked order now does a reasonable job of surfacing takes from people whose judgment you already trust. On a thread full of strangers — a breaking story, a viral post from an account you do not follow — the mutuals boost has nothing to work with, and the order falls back closer to the engagement-driven ranking that was already in place before July. Knowing which kind of thread you are in tells you how much to trust the default order before you have scrolled past the first few replies.

The SkimFast way

None of this changes what a reader can actually control: switching between ranked and chronological order, and doing it without losing your scroll position in the feed. SkimFast opens a thread's full reply list in a popover right in your timeline, sorted and searchable in place, so when X's own ranking model shifts again — and it likely will, given how much it has already changed in one year — you are not relearning a new set of assumptions about what "best" means. You are just flipping sort order and reading what is actually there.

A quick way to read past any model

X's ranking will keep changing; that much is already established by two rewrites in a single year. What stays useful regardless of which model is running underneath is a habit: check "newest" on anything time-sensitive, read a few rows past the top three on anything contested, and treat a reply's rank as a hint about the current model's preferences, not a verdict on whether it is correct. The mutuals update makes the ranked order more personal than before, which is exactly why a chronological read is worth reaching for whenever you want the thread itself rather than a model's guess about your taste. SkimFast keeps that switch one click away, so whichever version of X's ranking is live when you read a thread, seeing past it costs you nothing more than a tap.