Signals the System Cares About
How the PhoenixScorer quantifies the value of your tweets.
Multi-Target Engagement Prediction
X's recommendation system doesn't rely on a single metric. PhoenixScorer predicts the Log Probabilities of a user performing various actions.
| Action Type | Description | Internal Name |
|---|---|---|
| Favorite | Liking a tweet | ServerTweetFav |
| Reply | Replying to a tweet | ServerTweetReply |
| Retweet | Reposting content | ServerTweetRetweet |
| Click | Expanding tweet details | ClientTweetClick |
| Dwell Time | Time spent viewing the tweet | DwellTime |
| Share | Sharing via DM or Copy Link | ClientTweetShare |
Negative Actions
These behaviors trigger immediate downranking and can place an author on a recommendation blacklist:
- Not Interested: User flags content as irrelevant.
- Block Author: User blocks the creator.
- Mute Author: User mutes the creator.
- Report: User reports the tweet for violations.
Weight Combination Logic
The final score is a sum of these probabilities multiplied by their respective weights. The platform frequently adjusts these weights to optimize user experience.
Fun Fact: In many iterations, a Repost (Retweet) is weighted several times higher than a Like (Favorite) because it signifies strong endorsement. Replies that spark further conversations are weighted even more heavily.