What X's open-sourced algorithm actually says about shares
The short answer: X published the code behind its For You feed, and the highest-weighted positive action in it is not a like, a repost or a reply — it is sharing a post by copying its link. Most of what has been written about these numbers since is wrong, and the source code says so itself, in a comment written to head it off.
The published default weights
From home-mixer/params/param.rs in
xai-org/x-algorithm,
Apache 2.0. The repository went public in January 2026, but this file — the one holding the
actual numbers — was only added on 13 August 2026. Each weight multiplies the model's predicted
probability of that action; together they make the score that ranks a post.
| Action | Weight |
|---|---|
| Share via copy link | 20.0 |
| Reply, when the author follows you back | 5.0 + 15.0 |
| Reply | 5.0 |
| Quote post | 5.0 |
| Share via direct message | 5.0 |
| Follow the author | 4.0 |
| Share (generic) | 2.0 |
| Repost | 1.0 |
| Like | 0.5 |
| Click | 0.4 |
| Open a link | 0.2 |
| Profile click | 0.0 |
| Dwell | 0.05 |
| Report | -234.0 |
These are the defaults in the published configuration. They are runtime parameters, so they can be changed without changing the code.
Verified against the source on 28 August 2026. These numbers move: X updated this file on 26 August and the dwell weight went from 0.0 to 0.05. So treat any undated write-up of these weights — including this one, if you are reading it much later — as possibly stale, and check the file.
Two things almost every write-up gets wrong
1. The weights are not count equivalences
You will read that "one report cancels 468 likes". The code contains a comment written specifically to correct that:
"One common misinterpretation is that you can read these weight ratios as count equivalences ... this is incorrect because the weights apply to the predicted probabilities rather than raw counts."
So it is not that one copy-link share outweighs forty likes. It is that the ranker values the likelihood of a copy-link share far above the likelihood of a like — which means the system is tuned to promote content it expects people to copy and send on.
2. Engagement only counts if the post was served in a timeline
The same comment notes that navigating directly to a post has no ranking impact. Sending your own post round a group chat for likes does not do what people think it does — by this mechanism it does nothing at all.
Why you have seen different numbers: there are two repositories
Nearly every wrong figure in circulation traces back to one thing — X has open-sourced its ranking twice, and most write-ups are quoting the 2023 release as though it described today's feed.
- 2023 —
twitter/the-algorithm-ml, the "Heavy Ranker". Still online, still indexed, still quoted. - January 2026 —
xai-org/x-algorithm, the Grok-era replacement. This is the current one, and the source of the table above.
There is a subtlety here that catches careful writers, not just careless ones.
When the 2026 repository first went public in January it contained the ranking
code but not the ranking numbers: weighted_scorer.rs referred
to each weight by name — including the copy-link share — while the file defining their values
did not exist. Anyone writing between January and August 2026 could therefore see which
signals counted, but not what any of them were worth. The values were published on
13 August 2026. That seven-month gap is why so many correctly-sourced, 2026-dated articles
still carry the 2023 figures: for most of this year they were the only numbers available.
The 2023 weights, from that repository's own README, were:
| 2023 weight | Value |
|---|---|
| reply_engaged_by_author | 75.0 |
| reply | 13.5 |
| good_profile_click | 12.0 |
| good_click | 11.0 |
| good_click_v2 | 10.0 |
| retweet | 1.0 |
| fav | 0.5 |
| report | −369.0 |
Line those up against what people repeat and almost all of it decodes:
- "Replies ×13.5", "profile clicks ×12", "link clicks ×11" — the 2023 numbers, unchanged, presented as current.
- "Bookmarks ×10" — this is 2023's
good_click_v2, which meant "clicked into the conversation and stayed at least two minutes". It was never bookmarks. Bookmarks are not in either blend. - "A reply is 27× a like" — 13.5 divided by 0.5. A ratio between two 2023 weights, repeated as though it were a weight itself.
- "A report cancels 468 likes" — 234 divided by 0.5, the same arithmetic on the current numbers. It is the exact calculation the source code carries a comment to warn against.
- "Retweets ×20" — we cannot find this in either repository. A repost is weighted 1.0 in both. If you have seen that figure, it did not come from the code.
None of which makes the 2023 release fake — it was real, and accurate then. But a feed ranked by a model trained in 2023 is not the feed you are posting into now, and the giveaway is easy to spot: if a breakdown does not mention sharing by copy link, it is describing the old one.
What this means if you post videos
Almost every creator ends on some version of "like and follow" — and a like is the cheapest action in the list. The action the system values most is the one where somebody copies your link because they want to keep it, or send it to a friend.
X is the only large platform that has published its actual numbers, so it is the only one that can be quoted from source. It is not alone in the direction, though: Instagram's head, Adam Mosseri, has said publicly that "sends per reach" is one of the most important signals in ranking, and asked creators to "think about creating something people want to send to a friend". On TikTok, shares are widely reported to matter more than likes — but TikTok has never published its weights, so nobody can honestly put a number on it.
Where we come in
We build MemRa, a place to keep the videos you save so you can find them again by asking what they were about. We noticed all this because copying a link is exactly what saving something to MemRa is — so the action our app depends on turns out to be the one the ranker values most.
That is the whole pitch, and we would rather be useful than pushy: the table above is the point of this page, and it is right whether or not you ever install anything.
FAQ
Which action is weighted highest in X's algorithm?
Sharing a post by copying its link. In the published default weights it carries 20.0, against 5.0 for a reply, 1.0 for a repost and 0.5 for a like. It is the highest-weighted positive action in the blend.
Does one report really cancel 468 likes?
No. The source code contains a comment written specifically to correct that claim. The weights multiply the model's predicted probability of each action, not raw counts, so they cannot be read as count equivalences.
Do engagement pods work on X?
Not by this mechanism. The same code comment states that engagement only counts when the post was served in someone's Home Timeline, and that navigating directly to a post has no ranking impact.
About · Privacy Policy · Terms of Use · Help · Contact · Delete account
MemRa is a service of Dostum Amigo SL (CIF B76188499), C/ Princesa Teguise 27, Puerto del Carmen, 35510 Tías, Spain.