What were the famous 2023 engagement weights?
The 2023 system scored tweets by predicting several engagement types with a neural "heavy ranker" and combining those predictions with weights — and the relative magnitudes became infamous: a reply you replied to was worth far more than a like, video views and dwell mattered, and negative signals (reports, "show less often") carried large negative weights. The exact multipliers came from X's 2023 engineering blog rather than a single clean line of open code; the heavy-ranker model that produced the predictions is in the open the-algorithm-ml repo. Two important caveats: these are 2023 values, and the 2026 system no longer uses a fixed published table.
"A reply is worth 13.5x a like" is the single most-repeated number in X-algorithm folklore. It comes from the 2023 release era — and getting it precisely right means separating what was in the open code from what X stated in its blog.
The mechanism: predict, then weight
The 2023 Heavy Ranker (in the-algorithm-ml) was a neural model that predicted the
probability of multiple engagement outcomes for each candidate tweet — likes, replies, replies the
author engages with, retweets, video views, dwell, profile clicks, and negative actions like reports
and "show less often." The final score combined these predicted probabilities with weights.
The 2023 'For You' Heavy Ranker (open-sourced in twitter/the-algorithm-ml, projects/home/recap) was a neural model that predicted probabilities of multiple engagement outcomes (favorite, reply, reply-engaged-by-author, retweet, video view, dwell, profile click, and negative actions); the final tweet score combined these predictions with weights.
The weights came from the blog, not a clean code constant
Here's the precise part: the famous multipliers — the relative weights that made "replies dominate
likes" true — were published in X's 2023 engineering blog post about the release, not as a single
readable constant table in the open code. The model that produced the predictions is open; the exact
combining weights are best cited to X's own stated description. We mark the specific numbers as
officially-stated rather than claiming a code line we can't point to.
The specific 2023 engagement-weight multipliers (e.g. replies weighted far above likes, large negative weights for reports/'show less often') were published in X's 2023 engineering blog post about the release, rather than as a single readable constant table in the open code. The model producing the predictions is open; the exact combining weights are best cited as officially stated.
The shape of it (2023)
| signal class (2023) | direction |
|---|---|
| Reply, and reply the author engages back on | Large positive — the heaviest engagement signals |
| Retweet, like | Positive, with like notably lighter than reply |
| Video view, dwell, profile click | Positive attention signals |
| Report, "show less often", block, mute | Large negative |
Why this is history, not strategy
The 2026 system no longer exposes a fixed weight table. It uses a Grok-based learned ranker predicting engagement contextually per post and viewer. So "optimize for the 13.5x reply weight" is advice about a system that no longer runs that way. The positive-engagement principle survives; the specific gameable numbers don't.
What the code doesn't say
The exact 2023 multipliers as open code, and whether any survive into 2026. The numbers are from
X's 2023 blog (officially-stated, historical); the 2026 weights are withheld and applied to learned
predictions. Anyone quoting "13.5x" as a current fact is citing a 2023 blog about a replaced system.
The 2023 twitter/the-algorithm release describes a system that has been substantially replaced by the 2026 Grok-based release. Claims about 2023 specifics (weights, TweepCred, visibilitylib) are historical and are not evidence about how X ranks content in 2026.
What to do with this
Know the number's provenance so you can correct the folklore: it's a real 2023 figure from X's blog, not a current code constant. Optimize for genuine engagement the current learned ranker predicts — which xDoctor measures — rather than a multiplier from a system X has retired.