A shield that is too lax lets farms through; one that is too harsh excludes press, blogs and legitimate apps until the campaign has nowhere left to appear. The difference lies in calibration: known safe sites, aggressiveness levels with coherent thresholds, and an honest measure of what the shield contributes.
Safe Sites: the shield's whitelist
Domains and apps that are never excluded on score: well-known publishers, sites in your sector you want to be on, relevant apps. Sources:
- Your own knowledge (press, sector portals).
- The strongest trust signals (lesson 2): CMP + company number + longevity + schema
- social profiles → a candidate for automatic safe status.
- The harvest campaign: a Display campaign with a small budget and broad targeting whose sole purpose is to discover new placements, score them and classify the good ones as safe (and the bad ones as excluded) before the main campaigns ever reach them. Controlled exploration: pay a little to learn a lot.
Safe Sites get reviewed: a publisher can degrade (start selling resold inventory) and an app can change hands.
Aggressiveness levels
Instead of tuning twenty thresholds, a five-level selector that sets them all coherently in one go:
| Level | Exclusion score | Trust signals | Apps | Clustering | Who it is for |
|---|---|---|---|---|---|
| 1 · Conservative | > 85 | Maximum discount | Only the obvious categories | Identical pub-ID only | Brands that prioritise reach; sectors where press is key inventory |
| 2 | > 75 | High | Filler categories | pub-ID + NS | — |
| 3 · Standard | > 65 | Normal | All but the relevant ones | pub-ID + NS + IP | Most accounts |
| 4 | > 55 | Low | All | + RDAP | Acquisition on a tight CPA |
| 5 · Maximum | > 45 | Minimum | All | Everything | Accounts battered by fraud; budget that would rather lose reach |
Confirmed farms (an extreme score or a known network) are excluded at every level: the level never relaxes the obvious.
The retroactive effect when you relax
Moving up a level (more aggressive) excludes more from that night onwards. Moving down raises a decision: do you reinstate the exclusions made under the previous level that would no longer meet the threshold? With retroactive on, the shield releases those placements (and puts them back under watch); with retroactive off, what was excluded stays excluded and only the future criterion changes. Rule of thumb: retroactive on when the campaign has lost reach and the CPA did not improve; off when the high level was set because of a clear fraud problem.
An impact forecast before changing
Before moving the level, simulate over the last 30-90 days: how many placements would be excluded (or reinstated), what spend and impressions they represent, how many click-through conversions they had. If level 4 would exclude placements holding 15% of the spend and 2% of the conversions, the change is clearly a good one; if it would exclude 15% of the spend and 12% of the conversions, you are cutting legitimate reach.
Measuring what the shield contributes
The honest metric is not "how many placements have I excluded" but:
- The click-through CPA of Display/PMax before and after (comparable windows), with IS and spend stable.
- The conversion rate per click: it rises when the traffic that was not converting leaves.
- The share of spend on low-score (legitimate) placements: it should rise.
- An experiment (a campaign with the shield's exclusion list versus one without it) for large accounts: the definitive proof.
- Reinstated false positives: few, and explainable.
If the CPA does not improve and reach falls, the shield is badly calibrated (or the campaign's problem was never fraud).
Standard scenarios
Beyond the level, a catalogue of validated scenarios (complete rules with their thresholds, tested on real accounts) — "B2B acquisition Display", "Ecommerce with PMax", "Local with remarketing" — which set aggressiveness, apps, IPs and lists together. Automation that adjusts values within sensible ranges (Smart Mode) is useful; automation that invents scenarios is not: the catalogue bounds what it is allowed to touch.
💡 Ninja trick: in Ninja Shield, Safe Sites live in their own tab with the harvest campaign as their feeder, aggressiveness is a selector from 1 to 5 with retroactive Yes/No and an impact forecast per level, and Smart Mode tunes thresholds within validated scenarios — with TEST mode and a false positives tab that the Agent keeps an eye on. Calibration is your decision, taken with data; the execution is nightly.
What you should remember
- Safe Sites (a whitelist) fed by trust signals and by a harvest campaign.
- Five aggressiveness levels that set the thresholds in one go; confirmed farms fall at every level.
- Retroactive when you relax: yes if you lost reach without improving the CPA.
- An impact forecast before changing; measure the shield by click-through CPA and conversion rate, not by the number of exclusions.