Margin math first. Then a daily engine, inside guardrails, measured against a holdout.

1. Margin math first

Before a single bid moves, we build your margin model: landed cost + real Amazon fees per SKU, rolled into product families with individual profit targets. Loss leaders are declared strategy, not discovered accidents.

2. The daily engine

Every morning: reports pulled, history banked permanently (Amazon only keeps ~94 days; we keep everything), bids/budgets/negatives/harvests proposed or applied inside guardrails, and a plain-English report with a TLDR pulse posted to your Slack: yesterday / 7-day / 14-day / 30-day, top movers, flags.

3. The guardrails

• ±20%/day max bid move. Nothing dramatic can happen to your account in one day.
• 7-day cooldown per keyword. A keyword that just moved waits before it can move again.
• Daily change caps. A hard ceiling on how many changes any single day can apply.
• Attribution-aware. Amazon credits sales for 14 days, so recent days never panic the rules.
• Deal/coupon calendar. Freezes affected bids and quarantines deal-day data so it never teaches the algorithm bad lessons.
• Anomaly brake. An abnormal day applies nothing and pings a human.

4. The trust ratchet

Month one is propose-only: you see every recommendation in the daily report before anything is touched. Then you arm capabilities one at a time: bids, then harvesting, then negatives, then budgets. Every write is in a permanent ledger; any day is revertible with one command.

5. The holdout

At kickoff we freeze a stratified random ~10% of your keywords for the season. They get zero optimization. The gap that opens between the managed account and the holdout is our measured, causal lift, reported to you, not marketed at you.