Reading the Machine

Spent the evening reverse-engineering why my X impressions fell 41% in three months. 60K followers, consistent posting, same content quality. Something structural shifted and I wanted to know what.

Two things happened almost simultaneously. In February, the algorithmic Following feed changed how originals surface. In April, the repost scoring system got rewritten — 90% deduction on repost impressions, new diversity slot mechanics, session-level dedup replacing global. My entire distribution strategy was built for the old physics.

I pushed back on my own AI’s analysis twice. Once on quote tweets — they do cost a diversity slot, not free as initially stated. Once on self-reposts — they do provide real reach through session-level dedup, which my own testing had already proved. The corrections mattered. Getting the model right means the adjustments actually work.

The root cause wasn’t engagement decay or content drift. It was platform infrastructure moving underneath a strategy that hadn’t been re-examined. Now I have a two-week experiment queued: protect the first 30 minutes per post, cut self-reposts to best performer only, replace repost volume with community engagement, and push harder on video. Clean data in, clean signal out.

Algorithms aren’t mysterious. They’re just code someone else wrote. You can read them if you’re willing to sit with the details long enough.