Field Notes−2+1−1

The old program was one welcome email and a weekly blast to everyone. The new one treats the file like a garden: every subscriber sits in an RFM segment recomputed nightly, and every flow — welcome, replenishment, post-purchase education, winback — speaks to where that person actually is, not where the calendar is.

The unglamorous hero was the replenishment flow. Skincare runs out on a schedule; we computed each SKU's median days-to-depletion from order history and timed the nudge to land a few days before the jar does. It reads like thoughtfulness. It's arithmetic.

And we shrank the list on purpose. A 120-day sunset policy moved disengaged profiles through a three-touch winback and then suppressed them. The file got 22% smaller, deliverability recovered, and — the circled number above — repeat purchase rate rose 38%. Smaller, warmer, richer.

Briefs come from clustering, not vibes. We pull every query the site already ranks 5–25 for, cluster them by intent, and let the gaps write the editorial calendar. Each brief carries the target cluster, the angle, the internal links to earn, and the one question the post must answer better than anyone.

Then the assembly line with a human at every gate: AI produces a structured first draft from the brief; an editor reworks it — cuts the hedging, adds the opinion, checks every claim; a subject-matter reviewer signs off. Average human time per post is 24 minutes, which is the entire economics of the engine.

Three quarters in, organic sessions were up 3.1× and — the number I watch harder — signups attributed to content up 2.4×. Slop compounds in the wrong direction; edited work compounds in the right one.

The plumbing is deliberately boring: managed connectors land every channel's spend and every product event in the warehouse; a versioned, tested transform layer turns it into marts anyone can query; the dashboard on top is thin because the truth lives below it.

Attribution is where ops work becomes trust work. We run a position-based model as the baseline and a Markov removal-effect model as its skeptic; when they agree, decisions are easy, and when they diverge the divergence itself is reported. No single number pretends to be more certain than it is — which, it turns out, is exactly why people finally believe it.