Find inefficient search-term spend without killing good traffic.

1 · No-conversion evidence: JOOBIX calculates a binomial zero-result probability against an estimated leave-one-search-term-out campaign baseline. If campaign-level data is insufficient, an account-level fallback may support a Review signal, but it cannot create hard zero-conversion quantified excess. Because a scan can evaluate many zero-conversion queries at once, JOOBIX applies the Benjamini–Hochberg adjustment to those baseline-model probabilities. A query must pass the selected adjusted q-threshold before it can support a hard no-conversion verdict. The baseline rate is estimated rather than known, and search-term tests can share campaign context, so p/q values are conservative decision-support evidence—not a claim of formal exact inference or guaranteed FDR control.

2 · Target economics: CPA excess is estimated as spend minus conversions × target CPA. ROAS excess is estimated as spend minus conversion value ÷ target ROAS only when actual conversion value exists. JOOBIX does not assign an invented value to zero-conversion ROAS queries. These are target-based inefficient-spend estimates, not guaranteed recoverable savings.

3 · Conversion lag gate: If you do not confirm mature conversion data, performance-based exclusion verdicts are downgraded to Review and quantified excess is set to zero.

4 · Pattern safeguards: n-grams require multiple distinct queries. Hard pattern evidence uses campaign-stratified leave-pattern-out baselines with full click coverage; mixed or insufficient baseline coverage stays Review-only. Any pattern that collides with observed converting or protected queries cannot become an observed-safe negative hypothesis.

5 · Broad match: positive broad match is never treated as a waste signal by itself. Match type is context only.

6 · Negative matching simulation: Risk Guard models Google Search negative broad/phrase/exact behavior on literal observed tokens, preserves ampersand/asterisk distinctions, and applies Google’s documented 16-word query limit. It intentionally does not treat singular/plural or semantic variants as equivalent, and it downgrades candidates with platform-specific/invalid syntax to Review rather than claiming they are safe.