Published recommendation code can explain how a feed system is organized. It does not reveal every rule or the exact weights running on X today. Treat a historical implementation as evidence about that implementation, not a guaranteed growth formula.
Understand the main stages
- Candidate sourcing:find posts that could be relevant to a reader.
- Ranking:estimate which candidates to prioritize using available signals.
- Filtering and mixing:apply constraints and combine material into a feed.
The public recommendation repository describes components for these jobs, including relationship and interaction signals. Product surfaces such as search, notifications and the home feed need not use identical ranking.
Separate signal categories from live weights
A system can consider interactions, relationships, content and safety without one permanent multiplier for a reply or like. Code constants can change, and a public repository is not a live production configuration.
No fixed Premium 2-4 xboost, catastrophic negative-action penalty or first 30-minute cutoff is established here. An account’s observed results also depend on audience, topic, post age and available distribution.
Read technical claims carefully
- Check publication date, commit and scope.
- Find the actual component supporting the claim.
- Distinguish model inputs, predicted outcomes and business rules.
- Check whether a number is a default, experiment or historical value.
- State what the source cannot establish about today’s service.
Choose actions supported by your own work
Publish clear, useful material for intended readers, answer relevant questions and measure comparable examples. If your feed is the problem, use its available controls rather than treating it as a creator-ranking issue.
For practical tests, see the algorithm experiment guide. For a broken or unhelpful timeline, see feed troubleshooting.
Sources and checks
Reviewed October 16, 2026. Twitter recommendation algorithm repository.