How the X Algorithm Ranks Replies?

In March 2023, X published a large part of its ranking stack as the twitter/the-algorithm repository. Inside the home-mixer scored-tweets parameters, the heavy ranker carried a set of weights that told you, in plain arithmetic, what the system was optimizing for. It is still the clearest public evidence of how the X algorithm ranks replies against every other action a reader can take.

A reply was weighted 13.5. A retweet was 1.0. A like was 0.5. A reply that the original author then engaged with was 75. On the other side of the ledger, negative feedback such as a block, a mute or a "show less often" tap was -74, and a report was -369.

One caveat, and it matters: these are the weights in the 2023 open-source release. X has changed the live system since then and does not publish current values, so treat them as direction and relative priority rather than as today's exact math. In the current public repository these defaults are zeroed out and the real values live in configuration that is not public.

Even as direction, the ordering is striking. A reply was worth about 27 likes. A reply the author answered was worth about 150 likes. A report was the heaviest value in the whole published set, positive or negative, which tells you how expensive one annoyed reader was meant to be. Those parameters sat inside a much larger pipeline, and it helps to understand how the recommendation algorithm works overall before treating any one number as a lever you can pull.

The 75 weight on an author-engaged reply is the single largest positive value in that published set, and it points at something specific. The system was not rewarding the reply itself. It was rewarding evidence that the reply produced a real exchange between two people.

That changes what a good reply looks like. A polished one-liner that earns applause is a dead end by this logic. A reply the author can answer in ten seconds is not.

In practice that means writing something answerable. Point at one exact claim in the post rather than the post as a whole. Add a constraint, a counter-case, or a detail the author is likely to want to confirm or correct. Ask for one thing, not three.

It also means accepting that most replies will not get an answer, and that is fine. The asymmetry works in your favor: a handful of author-answered replies carries more weight than a long run of replies that collect a few likes each.

X does not distribute a post to your followers and stop. Structurally, it ranks candidates in For You by predicted engagement, which is why follower count alone is a poor predictor of reach and why a reply from a small account can surface above one from a large account in the same thread.

Some of the behaviors that feed that prediction are visible to you directly. Detail expands, when someone taps the reply to open it in context. Profile clicks, when they want to know who wrote it. Bookmarks, which tend to correlate with people coming back. New follows that land after a reader has read the reply. Quote posts, which create a new entry point into the same conversation. Others, such as how long a reader lingers on a post, are part of the signal set the ranking code was built around but are never exposed in your own analytics, so treat them as context rather than as a dial.

A useful habit is to compare two of your own replies with similar like counts and look at the columns underneath. The one with more detail expands and profile clicks is usually the one still visible in the stack an hour later. That gap is a large part of why some replies take off while near-identical ones sink without trace.

None of this requires a trick. It requires giving the next reader something to do: a specific point to agree with, extend, or push back on.

Weights are only half of how the X algorithm ranks replies. They apply once a reply is already in the candidate pool. Before that, relevance filtering decides whether your reply is a plausible continuation of the conversation at all.

Posts are doing different jobs. Some are starting an argument. Some are asking for a concrete tactic. Some are reacting to news. Some are personal. A reply built for the wrong job can still be well written and still land nowhere, because it does not give readers the kind of response the thread was set up to produce.

The fix is unglamorous. Match the frame of the post, then add one notch of specificity on top. Reuse the wording the post and the nearby replies are already using rather than paraphrasing it into your own vocabulary. If you cite something external, put the takeaway in the reply so nobody has to leave the thread to evaluate it.

On sensitive subjects, one clarifying line up front is worth writing. The published weights put negative feedback at -74 and a report at -369, and both are easier to trigger with an ambiguous reply than with an explicit one.

Most confusion about reply performance comes from comparing numbers that use different denominators.

Engagement rate has two common definitions. Per follower is what the big industry benchmark reports use, and they generally place the cross-industry median around 0.029%, down from roughly 0.035% in the prior edition, with a spread from about 0.009% for media accounts to about 0.072% for sports teams. Per impression is what X's own analytics shows, dividing engagements by impressions, and it is commonly reported under about 1%, with anything above 1% widely treated as strong. Text posts and threads commonly read higher than video on that measure.

Two other numbers are routinely misread next to it. X's native engagement rate is not link click-through rate, because it counts every interaction, including detail expands and profile clicks, so it will always look larger than the click number for the same post. Reach is the bigger surprise: a typical organic post is commonly reported to reach only low single-digit percentages of an account's followers, with recent summaries citing roughly 3-4%, and verified or Premium accounts are commonly reported to get wider distribution.

Read against those figures, the case for replies is easier to see, and it is the reason the question of replies pulling extra distribution keeps coming up: replying puts you inside a conversation that already has an audience, instead of asking For You to find one for you from scratch.

The awkward part of the reply-weight logic is the cold start. A thread with no responses gives the ranker almost nothing to score, and nobody wants to be the first person talking into an empty stack.

That is a real problem worth solving deliberately on your own threads. Getting a few substantive responses under the opening post early gives later readers something to react to, and it is the reason many accounts choose to seed the first genuine replies so a thread has something to build on before the conversation is left to run on its own.

Paid promotion works on the same principle when the targeting matches the intent of the post. It puts a reply in front of people who have context, early, while the thread is still forming. Poorly targeted spend does the opposite: impressions from readers with no context produce shallow taps at best, and the negative feedback weights punish the rest.

The durable version of this is a broader X growth setup where replies, original posts and audience building support each other rather than being run as separate campaigns. Replies bring you into other people's audiences. Original posts give those new readers a reason to stay. Neither works well alone.

That is the practical summary of how the X algorithm ranks replies: conversation is weighted far above reaction, the author's response is the highest-value outcome available to you, and relevance decides whether any of it is scored in the first place.

No. The 13.5 reply, 0.5 like and 75 author-engaged values come from the March 2023 open-source release, and X has changed the live system since without publishing current numbers. In the current public repository those defaults are zeroed out, so use them as relative priority, not as today's exact math.

Both, but for different reasons. Large accounts put your reply in front of more readers, while smaller accounts are far more likely to answer, and an author-engaged reply carried the highest published weight in the 2023 code. A mix of the two tends to beat chasing only the biggest posts.

Likes are the cheapest action a reader can take, and a reply that reads as a complete thought gives no reason to look further. Replies that earn profile clicks usually hint at specific experience or a detail the reader would have to visit your account to get, rather than closing the loop in the reply itself.