X Algorithm Weights in October 2026: How Posts Rank

The X algorithm ranking weights for October 2026

The table uses X's published parameters at commit 6258a15. Ratios are rounded against the like weight of 0.66. This is a dated look at the published settings, which X can override through configuration and experiments.

Selected weights, rather than the complete ranking formula. The last column compares coefficients; it does not convert real engagements into likes.

The mutual follow calculation adds 15 to the ordinary reply weight of 5. Its eligibility check requires the viewer and author to follow each other and the candidate post to be an original. It increases the weight of a predicted reply to that original post.

For a writer, the useful question is why someone would send the post to another person. A comparison that helps a colleague make a decision is a good reason. So is an explanation that answers a question their team keeps asking.

What those weights actually mean

The published scoring code combines weights with predicted probabilities and, for some attention signals, predicted continuous values such as reading time. A simplified action contribution looks like this:

Action contribution = action weight × predicted probability

In a hypothetical example, the model gives a reader a 10% chance of liking your post and a 1% chance of copying its link. The like contribution is 0.66 × 0.10 = 0.066. The copied link contribution is 20 × 0.01 = 0.20. Even with a lower predicted probability, the copied link contributes more in this example.

Change the reader and those probabilities can change. Someone who regularly sends practical advice to colleagues may respond differently from someone who mostly watches videos. The same post can therefore receive different scores for different people.

The code explains that reports are much rarer than likes, which is part of why their coefficient is so large. Negative feedback matters, but reading its weight as a universal reach deduction gets the mechanism wrong.

How a post gets into the For You feed

According to X's description of the feed pipeline, ranking happens after candidate collection and initial filtering. Posts from followed accounts compete alongside recommendations from elsewhere.

1. Collect candidates. The system retrieves posts that might suit the viewer.
2. Apply filters. It removes ineligible candidates, including duplicates, previously seen content and posts older than the configured age limit.
3. Predict responses and score. The model estimates likely actions, then combines those estimates with their weights.
4. Adjust and select. Further ranking steps account for factors such as repeated authors before posts become part of the final feed.

The snapshot configures a scoring result size of 2,800. Your post can be one of many candidates considered for a feed build. Scoring well helps, but visibility rules can still remove it before it reaches the viewer.

Original posts and replies have different discovery paths

The reply and repost filter removes those candidates when the viewer does not follow the author, unless a configurable exception allows certain replies retrieved through Phoenix. That exception is disabled in this snapshot's defaults.

This filter governs recommendations in the For You feed. Readers can still find a reply by opening the conversation it belongs to.

Publish original posts that make sense to someone who has never seen your account. In replies, add something that belongs in the existing conversation: an example, a useful correction or a question the author can answer.

Why the first two hours deserve your attention

The October parameters include a discovery mechanism for eligible posts, with a follower cap of 50,000, a two hour age window, a 200 impression threshold and target slots around the sixteenth position.

These settings give eligible posts a chance at an initial lift. They do not guarantee a feed position or 200 views for every smaller account.

Publish when you can answer people afterward. If someone asks for the example behind your claim, give it to them. If they spot an exception, discuss it. Their response may be the start of a more interesting conversation than the original post.

That does not require watching the screen for two hours. Choose a time when checking in is realistic. Our guide to the best time to post on Twitter explains how to find those windows for your own audience.

The pipeline also filters posts older than 48 hours in this published version. Treat that as a rule for this recommendation path, not an expiry date for the post itself. People may still find older posts through profiles, search and direct links.

Posting more often has diminishing returns within one feed

X also describes an author diversity adjustment that reduces the scores of additional posts from the same author. Publishing three times does not mean all three posts compete at their full original scores in one viewer's feed.

If you have three observations from the same experiment, consider whether one complete explanation would serve the reader better. Split them into separate posts when each has enough substance to stand alone.

Do bookmarks and group chats help your ranking?

Bookmarks have no dedicated term in the published action weight mapping. This calculation awards no separate bookmark contribution, though it cannot establish how saved posts are used elsewhere in X.

A bookmark still tells you someone wanted to return to the post. A checklist or reference can be useful for that reason alone, even without a direct ranking reward.

The scoring code also distinguishes engagement on posts served in Home from direct navigation. Opening a post from a group chat does not create the same Home ranking impact described here. Sharing useful work with a colleague can have value without being an algorithm tactic.

How to write posts people want to discuss and share

Before publishing, ask who would find the post useful and what they could do with it. A few changes often make the answer clearer.

1. Start with a specific observation. Replace “consistency matters” with what happened when you changed something, what surprised you and what another person could learn from it.
2. Include the detail people would ask for. A result becomes more useful when readers know the conditions. Explain the audience, the time period or the tradeoff behind it.
3. Invite a conversation you can contribute to. A focused question about a decision gives people somewhere to start. Be ready to explain your own answer.
4. Put the useful point early. Give readers enough context to understand the answer without making them hunt for it.
5. Review several comparable posts. Look for patterns in the subjects people discuss and share. One unusually good or bad result cannot tell you which part of the ranking system caused it.

For example, a consultant could publish “Three ways to improve onboarding.” A more useful post might explain the one question customers kept asking, the sentence the team changed and what happened next. The second version gives a reader something concrete to discuss with a teammate.

When reviewing results, separate reach from response. Twitter impressions tell you how often posts were displayed. The replies themselves tell you whether people understood the point, challenged it or wanted more. Use both when planning the next piece.

Common questions about the X algorithm weights

Are shares worth more than likes on X?

In this October snapshot, the copied link coefficient is about 30 times the like coefficient. The actual scoring contribution also depends on the predicted chance of that reader taking each action.

Does replying to a mutual follow get a 30 times boost?

The comparison describes the predicted reply contribution when ranking an eligible original post from a mutual follow. The eligibility check excludes candidates that are themselves replies or reposts. It is not a 30 times multiplier on the reply you write.

Are these the exact settings for every X user?

No. These are published defaults from a specific October version. X documents configuration overrides and experiments, so the public code cannot establish the exact settings behind every feed request.