One bid.
Twenty-one steps, all shown.

The Strategist runs a state machine per ASIN — LAUNCH, GROWTH, PROFIT, DEFEND, HARVEST. PROFIT means the listing has enough conversion history to be bid for margin rather than for rank, which is what licenses the target-ACoS arithmetic further down. In LAUNCH the same keyword would be bid above break-even on purpose.

Where the ad can appear for this term. GENERIC is unbranded search — the buyer is describing a product, not naming a seller. Branded and competitor surfaces carry different conversion rates and different ceilings, so the surface is read before anything is computed rather than after.

Conversion rate for this hour of this weekday, trained on Amazon Marketing Cloud data. Amazon.de converts materially better on a Tuesday afternoon than at 03:00, and a bid that ignores the hour overpays overnight and underbids in the window that actually sells.

(2.2 + 14) ÷ (2.2 + 68.8 + 402) = 16.2 ÷ 473 = 3.42%

The posterior mean of a BetaBinomial: the category prior plus what this keyword has actually done. Raw observation says 14 ÷ 402 = 3.48%; the prior pulls it toward the category’s 3.10%. With 402 clicks the data dominates, which is the behaviour you want — at 12 clicks the prior would still be carrying most of the estimate rather than letting one lucky conversion set the bid.

402 ÷ (402 + 164) = 0.71

How much of the estimate comes from this keyword rather than from the hourly model. 164 is the half-weight point: at 164 clicks the two contribute equally. It is a deliberate distrust of small samples — a keyword with 20 clicks is bid mostly on the model, and earns its own weight as it accumulates evidence.

(0.71 × 3.42% + 0.29 × 4.81%) × €68.70 = 3.83% × €68.70 = €2.63

Expected revenue from one click: the blended conversion rate multiplied by the average order value. Everything above this line is about how often a click converts; everything below is about what fraction of that revenue we are willing to spend to get it.

1 ÷ (1 + e^(−1.6 × (38 ÷ 14 − 1))) = 0.94

Sentinel damps the bid as stock runs down, on a sigmoid over cover ratio — days of supply against effective lead time. At 38 days against a 14-day lead the multiplier is 0.94, so barely any damping. At parity it would be 0.50, and below it the bid falls away quickly. This is the gate that stops a tool spending hardest on the ASIN about to stock out.

0.70 × €0.59 median bid (14 days) = €0.41

An anti-decay floor. Several small legitimate decreases in a row can walk a keyword down to a bid that no longer wins impressions, at which point it stops gathering the data needed to correct itself. The envelope holds the floor at 70% of the recent median so a decline stays recoverable. Not binding here — €0.68 is well above it.

€0.58 × 0.8 = €0.46 · €0.58 × 1.2 = €0.70

No single cycle may move a bid more than ±20%. The proposal is €0.68, inside the band, so the cap does not bind. When it does bind the bid moves to the edge of the band and the rest of the change arrives next cycle — a target reached in steps is recoverable, a target reached in one jump is a live experiment on the seller’s money.

€0.68 ≤ 35% × €2.63 = €0.92

A hard ceiling on bid against expected revenue per click, independent of the ACoS target. It exists because the ACoS target is a setting a seller can mistype: this gate catches a bid that would be absurd whatever the target says.

€0.68 ÷ €2.63 = 25.9% ≤ 27.5%

The implied ACoS of the proposed bid against the seller’s target for this marketplace. 25.9% against a 27.5% ceiling. The two are close because the proposal was derived from the target — this gate is checking the arithmetic that produced it, not applying a second opinion.

€0.68 ≤ 5% × €68.70 = €3.44

A per-marketplace absolute ceiling at 5% of order value, applied regardless of what the model believes. The backstop for a corrupted input: if AMC returned a nonsense conversion rate, every proportional check above would agree with itself and still be wrong.

14 conversions in 402 clicks — not zero-conversion

Guard 3.5: a keyword that has spent $100 or taken 50 clicks with no conversion stops being bid up, whatever the model estimates. This keyword converts, so the gate is inert. It is the single cheapest waste control in the system.

Two switches, global and per-marketplace, either of which halts submissions immediately. Checked at decision time rather than at submission time so a switch thrown mid-cycle stops the decision as well as the write.

A per-keyword breaker that opens when return on ad spend collapses relative to its own recent baseline, and stays open until the keyword recovers. CLOSED is the healthy state — current flows. Per-keyword rather than per-campaign so one failing term cannot freeze a whole campaign, and cannot hide inside its average either.

A hash of (keyword, bid, campaign) checked against the last 24 hours. A retry after a timeout must not place the same bid twice — Amazon accepts it, and the audit log then shows two decisions where one was made.

€0.68 ÷ €0.58 = 1.17× ≤ 3×

The last gate before the write: proposal against the keyword’s own baseline, clamped at 3×. The cycle cap already bounds a normal move, so anything reaching this gate is a bug rather than a decision. It has fired in staging and never in production, which is the correct ratio for a check of this kind.

A Thompson draw decides whether this cycle exploits the current estimate or explores around it. In PROFIT state exploration is held to 5% of decisions, and this draw came back EXPLOIT — so no exploration multiplier is applied and the bid is exactly what the model asked for.

The bid currently live at Amazon for this keyword, read at decision time.

€2.63 × 27.5% × 0.94 = €0.68

Expected revenue per click, times the share of it the seller is willing to spend, times the supply multiplier. Every term is on this page and every term is checkable — which is the whole argument for writing the trace down.