Google Ads Copilot

Public Alpha — Builder Preview I built an OpenClaw agent that manages Google Ads accounts through intent analysis instead of bid automation. All three layers — Read → Draft → Apply — are live and tested on real accounts. This is an early release for builders and agencies who want to try the intent-first approach.

Most Google Ads problems are not bid problems. They are intent-mixing problems.

Your account is buying different kinds of searches — buyers, comparison shoppers, researchers, job seekers, freebie hunters — and mixing them together. When that happens, bids optimize against noise, RSAs get generic, good traffic subsidizes bad traffic, and every budget decision is distorted.

Google Ads Copilot is an AI agent that reads search behavior, builds a durable Intent Map, finds waste, isolates signal, and turns what the account is learning into structure, copy, and scale decisions. Every recommendation is staged as a reviewable draft. Nothing changes without your approval.

Not a dashboard. Not a bid optimizer. Not a generic audit checklist.

A strategist with memory.

---

How It Works

READ  →  DRAFT  →  APPLY

See ARCHITECTURE.md for the full design, safety model, and roadmap.

---

Two Ways to Run It

Connected Mode (recommended)

Hook up the MCP server with OAuth + developer token and the agent pulls live data automatically — search terms, keywords, conversions, budgets, everything via GAQL.

You get: live account data, automatic account discovery, date-range fallback for sparse accounts, cross-referencing search terms against targeted keywords, and health detection out of the box.

Setup: ./install.sh auto → /google-ads connect setup → you're live. See data/mcp-config.md for details. If your host uses a non-default skills path, override it with CLAUDE_TARGET=... or OPENCLAW_TARGET=....

Export Mode (zero setup)

Paste a CSV from Google Ads. The analytical engine is the same — you just feed it manually. Good for one-off audits or accounts where you don't have API access yet.

Both modes produce the same quality of analysis and drafts.

---

What It Can Do

15 skills across three layers. The full operator workflow is in OPERATOR-PLAYBOOK.md.

Run the Account

Analyze

Act

---

Why This Exists

Most Google Ads tools report metrics, enforce generic best practices, or automate bids. None of them think about what the account is actually buying.

Google Ads Copilot starts from search intent — what people typed, what they meant, and whether the account structure reflects that reality.

---

Getting Started

Try It in 5 Minutes (export mode)

1. ./install.sh auto
2. Export search terms from Google Ads (last 30 days, CSV)
3. /google-ads search-terms → paste your data
4. Check workspace/ads/drafts/ for what it found

Full Setup (connected mode)

1. Configure the MCP server: data/mcp-config.md
2. ./install.sh auto
3. Create local credential/env files from the committed templates: cp data/google-ads-adc-authorized-user.template.json data/google-ads-adc-authorized-user.json cp data/google-ads-mcp.test.env.example.sh data/google-ads-mcp.test.env.sh
4. Fill in your real values, then source data/google-ads-mcp.test.env.sh
5. ./scripts/test-mcp.sh to verify connectivity
6. /google-ads connect setup → discovers your accounts, picks one, writes workspace
7. /google-ads daily or /google-ads audit → live data flows automatically
8. For PMax-heavy accounts, use ./scripts/test-search-terms.sh <customer-id> to verify query visibility and fallback behavior
9. Review drafts → approve → apply → verify

The repo does not ship any real credentials or live test files. Only templates are committed.

Install from a Release Bundle

You can install from the repo as usual, or from a packaged bundle:

./install.sh auto
./install.sh /path/to/google-ads-copilot-0.2.0.tar.gz auto
./install.sh /path/to/google-ads-copilot-0.2.0 openclaw

Release bundles are built with ./scripts/package/build-release.sh <version>.

See DEMO-WORKFLOW.md for a guided walkthrough of the full cycle.

Environment variables:

---

Workspace Memory

Every session compounds. The agent writes to workspace/ads/ — a persistent knowledge base that survives between runs:

Draft filenames use YYYY-MM-DD-[account-slug]-[type].md. Multi-draft audit runs also produce _batch-YYYY-MM-DD-[account-slug].md as a durable audit packet, while _summary.md remains the live prioritized backlog snapshot.

---

Package Structure

google-ads-copilot/
├── README.md                    # This file
├── ARCHITECTURE.md              # Three-layer design doc
├── APPLY-LAYER.md               # Apply layer design (safety, reversibility, audit)
├── OPERATOR-PLAYBOOK.md         # Full operator workflow loop
├── CHANGELOG.md                 # Release history
├── LICENSE
├── install.sh
│
├── google-ads/                  # Orchestrator skill
│   ├── SKILL.md
│   └── references/              # Strategic playbooks
│
├── scripts/                     # Helper scripts + apply layer CLI
│   ├── apply-layer/             # Write-path CLI (bash + curl + jq)
│   └── list-customers.sh        # Account discovery
│
├── skills/                      # Analytical skills (15 skills)
│   ├── google-ads-apply/
│   ├── google-ads-audit/
│   ├── google-ads-budget/
│   ├── google-ads-connect/
│   ├── google-ads-daily/
│   ├── google-ads-draft-summary/
│   ├── google-ads-intent-map/
│   ├── google-ads-landing-review/
│   ├── google-ads-negatives/
│   ├── google-ads-plan/
│   ├── google-ads-pmax/
│   ├── google-ads-rsas/
│   ├── google-ads-search-terms/
│   ├── google-ads-structure/
│   └── google-ads-tracking/
│
├── data/                        # Data layer (MCP config, GAQL recipes)
├── drafts/                      # Draft templates per action type
├── examples/                    # Sanitized example outputs
├── evals/                       # Eval suite
└── workspace-template/          # Blank workspace for new accounts

---

Apply Scope

Budget applies are limited to campaign daily budgets and require an ## Apply Manifest JSON block plus hard guardrails: max 30% per action, min meaningful delta, 7-day cooldown, tracking confidence gate, pending-tracking-draft block, budget-neutral default, and confirm budgets confirmation.

Still excluded: shared budgets, bid strategy changes, campaign creation, RSA edits, enabling paused entities, deletions.

---

What's Next

This is a public alpha. The analytical engine and apply layer are solid — tested on real accounts with real money. Current next steps:

• Expanded write actions after budget apply proves out
• More example walkthroughs from real (anonymized) accounts
• Broader eval coverage for decision quality and write-path regressions
• Packaging and release automation for broader distribution

If you're running Google Ads for clients or your own business, try it. Break it. Tell me what's missing.

---

License

See LICENSE.