# ROAS Benchmark — Paid Ads Evaluation Standard

ROAS evaluates **incremental paid-media contribution and operating quality under declared business constraints**. The mnemonic remains Return · Offer · Audience · Spend Efficiency, but the objective is not to maximize a platform-reported ROAS ratio. A high reported ratio can coexist with weak incrementality, under-spending, attribution inflation, or poor profit.

The framework is advisory. Executable items, profiles, context, and vetoes live in [`framework-catalog.json`](framework-catalog.json); common evidence, missingness, score, status, and verdict rules live in [`scoring-semantics.md`](scoring-semantics.md).

**Keyless by design:** Tier 1 accepts the user's own exports from ad managers, analytics, ecommerce/CRM, and placement reports. Keyed platform APIs are optional conveniences, never a baseline requirement.

## Unit and Required Context

Score one account or campaign portfolio for one currency, normalized attribution window, observation period, conversion-lag assumption, goal, and business constraint. A change in those fields creates a different run. Platform self-report is evidence, not the outcome truth set.

## The 20 Items

| ID | Dimension | Criterion |
|---|---|---|
| `R1` | Return | Conversion instrumentation is verified against an own-data truth set. |
| `R2` | Return | Cross-platform attribution is deduplicated; windows and currency are normalized. |
| `R3` | Return | Incremental contribution or profit is measured against the declared target/control. |
| `R4` | Return | CAC/CPA and payback satisfy the declared business constraint. |
| `R5` | Return | Marginal return is read after conversion lag with uncertainty stated. |
| `O1` | Offer | Claims and required disclosures are substantiated. |
| `O2` | Offer | Platform policy and restricted-category requirements are satisfied. |
| `O3` | Offer | Economics, eligibility, terms, and availability are explicit. |
| `O4` | Offer | Ad-to-landing message and intent match. |
| `O5` | Offer | Hook, format, accessibility, and fatigue state fit the placement. |
| `A1` | Audience | Brand and placement safety are verified from placement evidence. |
| `A2` | Audience | Targeting and query/audience intent fit. |
| `A3` | Audience | Negative keywords, exclusions, and suppression controls are maintained. |
| `A4` | Audience | Account structure supports the objective without avoidable overlap. |
| `A5` | Audience | Reach, overlap, and audience saturation are measured. |
| `S1` | Spend Efficiency | Budget pacing remains within the declared plan and constraints. |
| `S2` | Spend Efficiency | Bid strategy and learning-state changes are governed. |
| `S3` | Spend Efficiency | Marginal CPC/CPM/CTR/CVR efficiency uses a normalized window. |
| `S4` | Spend Efficiency | Frequency and creative decay are separated from audience saturation. |
| `S5` | Spend Efficiency | Paid/organic and cross-campaign cannibalization are assessed. |

## Profiles and Scoring

| Profile | R | O | A | S | Use |
|---|---:|---:|---:|---:|---|
| `direct-response` | .40 | .20 | .15 | .25 | Conversion program under an explicit CAC/payback constraint |
| `prospecting` | .15 | .30 | .30 | .25 | Reach/consideration program with a declared downstream outcome |
| `incremental-profit` | .50 | .15 | .10 | .25 | Holdout or causal design centered on contribution/profit |

Per item: Pass = 10, Partial = 5, Fail = 0. The Paid Ads Index (`RQS`) is the floor-rounded profile-weighted mean after 100% applicable evidence coverage. `RQS` is not the literal ROAS ratio and profiles are not interchangeable.

For `R=75 O=80 A=85 S=78`, direct response is `78`; prospecting is `80`; incremental profit is `77`. These are arithmetic fixtures, not outcome predictions.

## Vetoes and Unknowns

| ID | Verified failure |
|---|---|
| `ROAS-R1` | Instrumentation is demonstrably broken or does not reconcile to the named truth set. |
| `ROAS-R2` | The same outcomes are demonstrably double-counted or materially inflated. |
| `ROAS-O1` | A material claim/disclosure is false, unsubstantiated, or missing. |
| `ROAS-O2` | The offer violates an applicable platform/restricted-category rule. |
| `ROAS-A1` | Placement evidence shows a material brand-safety breach. |

No data or no access is `unknown`, producing `NEEDS_INPUT/UNDECIDED`; it is not a veto. iOS/ATT modeled data may be partial or weak evidence but is not automatically a failure. Only an `ad-account-auditor` invocation qualifies veto evidence and applies the truth table: one verified veto caps the final score at 59; two or more produce `verdict: BLOCK` and no final score. Every other skill reports potential ROAS control evidence and hands it off without calling it a verified veto or emitting a gate verdict.

Premature scaling and learning-phase disruption are high-severity `S2` findings, not automatic vetoes.

## Evidence Contract

| Need | Preferred evidence |
|---|---|
| Spend, pacing, queries, placements | Campaign, search-term, audience, and placement exports |
| Outcomes | GA4/ecommerce/CRM order or lead truth set with stable IDs |
| Attribution integrity | Deduplicated IDs/timestamps, normalized windows/currency, documented lag |
| Incrementality | Holdout, geo/market split, causal test, or explicitly labeled proxy |
| Profit constraint | Margin, fulfillment, CAC/payback, and contribution assumptions with provenance |

Use `~~web analytics`, `~~ecommerce`, and `~~ad platform` connector categories from [`CONNECTORS.md`](../CONNECTORS.md). Keep raw platform metrics, calculated reconciliations, and estimates visibly distinct.

## Skill Ownership

- **Research** — `campaign-architect`, `search-term-miner`, and `audience-segment-builder` own structure and audience inputs.
- **Orchestrate** — `ad-creative-builder`, `ad-test-designer`, `landing-optimizer`, and `bid-strategy-planner` own offer/experience and test design.
- **Activate** — [`ad-account-auditor`](../ad/activate/ad-account-auditor/SKILL.md) produces the gate; `conversion-signal-qa` fixes signal prerequisites but does not score them.
- **Scale** — `paid-measurement-loop`, `attribution-reconciler`, `budget-pacing-monitor`, and `fatigue-frequency-manager` supply normalized return and efficiency evidence.

Experiment helpers return statistical facts; the calling skill owns the precommitted business decision under [`measurement-protocol.md`](measurement-protocol.md). ROAS remains advisory until its versioned profiles pass the calibration protocol.