How the X Algorithm Actually Weighs Replies
Every few months someone posts a screenshot claiming to know the exact multiplier X uses for reply engagement versus likes. It always has a suspiciously precise number attached, and it's almost never sourced to anything real. Here's what's actually documented, versus what's a reasonable guess, versus what's just made up.
This matters because a lot of reply advice floating around is built on confident-sounding claims nobody can verify. I'd rather tell you what's actually known, flag what's speculative, and let you make your own call, than repeat a number I saw in someone else's thread.
What's actually documented
In 2023, Twitter published a portion of its recommendation algorithm's source code publicly. That release showed the broad architecture: a candidate generation stage that pulls in possible tweets to show you, followed by a ranking stage that scores those candidates using a mix of signals, including predicted engagement probability across several types of interaction (likes, replies, retweets, dwell time, and a few others). It confirmed that replies are one of several engagement types the ranking model is trained to predict and weight, and that conversations (replies to replies) factor into scoring for a tweet's overall visibility.
What it did not do is hand over the exact live weights currently in production, and those weights are not static even if they were published once, because ranking models get retrained and adjusted continuously. So "the algorithm favors replies" is a fair, documented statement. "Replies are weighted exactly X times more than likes" is not something anyone outside the company can honestly claim right now.
What's reasonable to infer, flagged clearly as inference
Dwell time (how long someone stays on a piece of content before moving on) is a general industry-wide ranking signal, and it shows up in the parts of the released code covering ranking features. It's reasonable to infer that a reply thread which keeps people reading, scrolling through multiple replies, expanding a thread, benefits from that same kind of signal. This is an inference from documented mechanics, not a confirmed specific behavior for replies. I want to be upfront that this is the speculative part.
Similarly, "conversation depth," meaning a tweet that spawns genuine back-and-forth rather than a pile of disconnected one-off replies, is plausibly rewarded because sustained engagement of any kind is the general thing these systems are built to detect and amplify. But X hasn't published a specific "conversation depth score" as a named, confirmed ranking feature. Treat this as an educated read of publicly available architecture, not a verified fact.
What's genuinely unclear or unpublished
Whether replying frequently to other accounts helps your own tweets get algorithmic reach isn't something X has confirmed as a ranking mechanic. There's a separate, well-established social effect where people you've interacted with are more likely to engage with you back (this is documented in the reciprocity piece on this blog), which produces a similar real-world result through human behavior rather than an algorithmic rule. It's worth being precise about which of those two things is actually happening, because they get conflated constantly.
Most confident claims about "the algorithm" online are pattern-matching dressed up as certainty. The honest position is: some things are documented, some things are reasonable inference from documented architecture, and a lot of what gets repeated is neither.
Why this matters for how you write replies
None of this changes the practical advice much. Whether or not there's a specific algorithmic reply bonus, a reply that reads full thread context and adds something specific still gets more human engagement than one that doesn't, and human engagement is the actual thing that then feeds any ranking system regardless of the exact weights. That's the mechanic covered in more tactical detail in five reply frameworks worth stealing: write for the humans reading the thread first, and the algorithmic outcomes tend to follow from that rather than the reverse.
My actual take
I think the obsession with reverse-engineering exact algorithm weights is mostly wasted effort for individual accounts. The released source code confirms the categories of signals in broad strokes, and that's genuinely useful context. But chasing an exact multiplier that isn't published, and that changes anyway, is time better spent writing a reply that a human would actually want to respond to. The algorithm rewards genuine engagement because that's what it's built to detect. Skip the exact-number treasure hunt and optimize for the human on the other end instead.
FAQ
Yes, in 2023 Twitter released a portion of its recommendation and ranking code publicly on GitHub. It showed the general shape of the system, including that replies and engagement signals feed into ranking, but it did not expose live production weights or the full current ranking model, which changes over time.
There's a well-documented social reciprocity effect (people you've engaged with are more likely to engage back), and separately, active conversation participation is understood to be a positive signal in engagement-based ranking generally. Whether X specifically weights your outbound reply volume as a ranking input for your own tweets isn't something the company has confirmed publicly, so treat that specific claim as plausible but unverified.
Dwell time is a documented signal used broadly across recommendation systems in the social media industry, including in X's released ranking code. It's reasonable to assume it applies to replies and threads, but X hasn't published exact reply-level dwell time weighting.
No. Replies are generally scoped to the conversation they're posted in and to the profile of the replier, they don't get pushed into the main For You feed the way original tweets and quote tweets can. A reply's reach mostly comes from people already viewing that specific thread, plus whatever visibility the replier's own profile carries.
Be skeptical of anyone stating a precise multiplier unless they can point to an actual source. X has confirmed the general categories of signals it uses, but has not published exact weights, and those weights change over time regardless.
Related reading
Write for the human, not the algorithm
Ekoreva reads the thread and suggests replies people actually respond to.
Add to Chrome (free)