The Intermediate level introduced the judge; here is its complete specification — the one that lets a person (or a script) take, every night, the same decision an experienced manager would take with all the data in front of them, and leaves every decision justified.
Inputs per campaign (or portfolio)
- Business target: mode (CPA/ROAS/new customers) and figure.
- Reference conversions: the list of actions (and, if the mode is ROAS, their values).
- Conversion latency (bid strategy report): days to 90%.
- Strategy status (learning, limited, active).
- Limits: minimum/maximum budget, minimum/maximum target per campaign (the plan's guard rails).
- History of moves (date of the last one, size).
Windows and volume
- Judgement window: 14-28 complete days, excluding the last N days (N = latency to 90%). Long-cycle campaigns get a longer window.
- Minimum volume: around 10 reference conversions in the window (or 30 clicks per expected conversion) before a verdict is issued; fewer = "no data".
- Status: while "learning", nothing moves unless there is a serious alarm.
- Wait: do not move a campaign if its last move is less than 7-14 days old (depending on the cycle).
Compliance bands
| Band | Criterion (CPA) | Criterion (ROAS) |
|---|---|---|
| Overperforming | actual CPA < target × 0.85 | actual ROAS > target × 1.15 |
| On target | between 0.85 and 1.15 | between 0.85 and 1.15 |
| Missing | actual CPA > target × 1.15 | actual ROAS < target × 0.85 |
| Missing badly | actual CPA > target × 2 (two windows running) | actual ROAS < target × 0.5 |
The bands are tuned per account (wider when volume is low).
The decision matrix
| Verdict | Lost IS (budget) | CPA trend (recent steps) | Action |
|---|---|---|---|
| Overperforming | High (> 20%) | Stable or improving | Budget +15-20% |
| Overperforming | High | Getting worse | Budget +10%; watch the marginal cost |
| Overperforming | Low | — | Loosen the target by 10% (more volume, slightly dearer) |
| On target | High | — | Budget +5-10% if the plan allows it |
| On target | Low | — | Nothing |
| Missing | — | Improving | Nothing (it is correcting itself); review in the next window |
| Missing | — | Stable/getting worse | Review search terms, ads, landing page, measurement (diagnosis); if all is fine, tighten the target by 10% |
| Missing badly | — | — | Budget −20% to limit the damage + diagnosis; do not touch the target until you understand it |
| No data | — | — | Nothing; if it persists for two months, propose a portfolio/merger |
Rules that cut across everything: one move per campaign per cycle; maximum steps (budget ±20%, target ±15%); within the plan's limits; never budget and target at the same time.
Loose campaigns versus portfolios
- Campaign with its own strategy: the judge moves its target and its budget.
- Campaign in a portfolio: the judge aggregates the portfolio's campaigns for the target verdict (it is the portfolio's target that moves) and judges each campaign separately for the budget.
- Campaign on target impression share (brand): budget only, based on IS and on the reference CPA (cross-mode).
Anomalies and safeguards
Before deciding, the judge throws out broken data:
- Conversions at zero with normal clicks → broken measurement: do not move, raise a warning.
- An absurdly high CPA on one day (20× the target) → anomaly (an expensive click with no conversion yet because of the delay): it is excluded from the calculation.
- A recent budget or target change made some other way (change history) → wait.
- Declared seasonality (an adjustment active) → recalibrated window or the judge pauses.
The record
Every move, one row: date · campaign (ID) · verdict · window metrics · IS · trend · action · from → to · rule applied · result three weeks later (filled in on the next cycle). It is what lets you audit the judge, calibrate its bands and prove to management why every euro moved.
💡 Ninja trick: this matrix is, literally, the configuration of Smart Bidding (SBNS): bands, maximum steps, waiting periods, per-campaign limits, anomaly exclusion, TEST mode (it writes the verdict and the action without applying it) and an automatic record. What changes compared with doing it by hand is not the rule — any good manager knows the rule — but that it is applied every night, across every campaign, without skipping a single one.
What you should remember
- Inputs: target, reference conversions, latency, status, limits, history.
- A window that excludes the latency, minimum volume, a wait between moves.
- Bands of ±15% (badly = ×2); the verdict × IS × trend → action matrix; one move, capped steps, never budget and target at the same time.
- Anomalies thrown out; a record of every move with its result.