CopySAFE

Open-source SEO, GEO, and marketing skills for AI agents.

The instruction file, verbatim from the audited commit — this is the text the model reads, and the surface the audit's instruction layer examines. Quoted here so you can judge it without cloning anything.

---
name: google-ads-copy
description: Generate and A/B test Google Ads copy. Use when asked to write ad copy, headlines, descriptions, create ad variants, test ad messaging, improve CTR, or generate RSA (Responsive Search Ad) components. Trigger on "ad copy", "write ads", "headlines", "descriptions", "RSA", "responsive search ad", "ad text", "ad creative", "improve CTR", "ad A/B test", "ad variants", "write me an ad", "ad variation experiment", or when the user wants to improve click-through rate on existing ads.
argument-hint: "<ad group name, keyword theme, or 'write new ads'>"
---

# Ad Copy Generator + A/B Tester

Write Google Ads RSA copy and run structured tests to find winning messaging.

## Setup

Read and follow `../shared/preamble.md` (MCP detection, account selection) and `../shared/analysis-principles.md` (evidence requirement, guardrails). Both apply throughout.

## Reference

Read on demand:

- `references/rsa-best-practices.md` — character limits, headline formulas, pinning, A/B mechanics, common mistakes. The source of truth for what to write.
- `references/rsa-testing-lab.md` — how to choose RSA test metrics, diagnose ad-group intent bloat, decide pinning, and avoid chasing Ad Strength over business outcomes.
- `../manage/references/industry-benchmarks.md` — industry CTR/CVR benchmarks for sanity-checking targets.
- `../manage/references/quality-score-framework.md` — only when QS improvement is the explicit goal.

## Business context — non-negotiable input

Every copy decision is grounded in business context. Read `{data_dir}/business-context.json` and `{data_dir}/personas/{accountId}.json` first.

- If either file is missing or empty, recommend `/google-ads-audit` before writing — generic copy that ignores positioning is wasted ad inventory. The full intake procedure (website crawl, schema, bootstrapping) lives in `../audit/references/business-context.md`.
- If the user volunteers new info (new service, changed positioning, seasonal update), merge it in.

How context shapes copy: services map to headline categories; locations earn local-specific headlines (with QS upside); brand voice sets tone and forbidden words; differentiators ARE the value-prop headlines; competitors sharpen positioning (without naming them); seasonality decides urgency vs. evergreen framing; offers feed time-sensitive variants; landing-page content must match (or conversions drop). Personas drive language: their own search-term phrasing in headlines, their pain points in descriptions, their decision-trigger as the CTA angle.

## Ad-strength vs. real performance

Ad strength optimizes for Google's diversity goals, not your conversion rate. The hierarchy: **conversion rate > CTR > CPA > ad strength.** Don't break a high-CTR ad to chase "Excellent". Do treat low ad strength as a useful diversity signal — eight distinct headlines, one per category, no repeats in description bank.

## Pulling what to write from

Copy must be grounded in what already converts. Choose available read capabilities for the evidence relevant to the requested copy:

- `ad_group_ad` — current headlines, descriptions, ad-strength, per-ad clicks/CTR/conversions (the baseline to beat)
- `keyword_view` — what's converting and which QS components are weak
- `search_term_view` — the actual phrases customers are typing (the single best language source)
- `campaign` — the CTR/CVR benchmark each variant has to clear

For brand-wide rewrites, correlate everything in one pass. For a single ad group, scope each query with `WHERE ad_group.id = ...`. Layer in seasonality and keyword-landscape context from `business-context.json`.

If the user has a CRM or lead-outcome database with the language customers actually use, mine that — customer language beats marketing language every time.

## Copy output contract

Every proposal needs a small decision record, not just a block of headlines:

- A **concept ID** and one-sentence hypothesis: `persona × motivation × angle`.
- A **claim ledger** for every proof, offer, rating, guarantee, price, or superlative: exact source (account data, business context, approved landing page, or user-provided evidence) and whether it is approved for ad use.
- Complete RSA assets with character counts and pin positions, plus a distinct challenger concept when testing.

Do not turn a vague request into fabricated proof. If a claim is attractive but unsupported, label it `needs_substantiation` and write a claim-free alternative. A citation to a competitor, an unverified review site, or general industry knowledge is not approval to use a claim in an ad.

## Competitive copy rules

- **Never** name a competitor in ad copy (policy risk + you do their brand awareness for free).
- **Never** use "best" / "#1" without verifiable substantiation. Google requires it; the Trademark team enforces it.
- **Do** use specific features competitors lack ("Same-Day Service" beats "Better Service"), trust signals ("4.9★ Google · 500+ Reviews"), guarantees, transparent pricing, and location specificity. Verifiable specificity outperforms superlatives.

## RSA mechanics

Google RSA: up to **15 headlines (30 chars max)** and **4 descriptions (90 chars max)**. A single character over = rejected. Always count.

`references/rsa-best-practices.md` is the source of truth for headline formulas, description ordering, and pinning strategy. The compact rule of thumb: pin one Service+Location headline to position 1, one CTA to position 3, leave position 2 unpinned for Google to test value-prop / trust / differentiator headlines, and never pin more than 3 total.

## A/B testing — use the experiments framework

When the user wants to test, define the hypothesis, traffic split, measurement,
and decision rule, then check which experiment capabilities the live connection
supports. Ad-copy tests and larger changes to audience or landing page may need
different designs. Running two ads side by side without a controlled split is
observational evidence; do no

SAFEgrade B · trust 89/100 Nothing in the source contradicts what it says it does. Grade A is reserved for packages that have also passed the behavioural sandbox.

What the source does

Filesystem

none-observed

Network

none-observed

Shell

none-observed

Dependencies

pinned

Secrets in source

none-found

Findings (0)

No findings outside the package's declared scope.

Gates applied: no_behavioural_pass.