Performance Max is no longer a black box: pipe PMax into Claude via MCP

Performance Max has long been called a "black box" — Google controls asset combos, bidding, and audience targeting, and you only see aggregated data. In 2025 Google finally opened up finer-grained PMax API data. This post shows how to wire those data points into Claude via MCP, so AI can help you make PMax decisions.

Why not just use the Google Ads UI?

• PMax asset-group insights, Top Search Terms, and spend curves are scattered across 4-5 different screens
• Cross-account comparisons require manual CSV exports + Excel
• Google's in-UI AI suggestions are generic ("add assets", "adjust bid") and ignore your business context (inventory, season, margin)

5-minute MCP setup

Prereqs: you've connected Google Ads in DEXUN AdWhiz and generated an API key (dashboard → Settings → API Keys).

Step 1: Edit Claude Desktop config

macOS: `~/Library/Application Support/Claude/claude_desktop_config.json`. Windows: `%APPDATA%\Claude\claude_desktop_config.json`.

{
  "mcpServers": {
    "dexun-adwhiz": {
      "command": "npx",
      "args": ["-y", "@dexun/adwhiz-mcp@latest"],
      "env": {
        "DEXUN_API_KEY": "dxk_live_..."
      }
    }
  }
}

Step 2: Restart Claude Desktop

You should see a new 🔌 icon in the bottom-right. Hover to confirm "dexun-adwhiz: 100+ tools available".

Real example: let Claude audit your PMax

Just ask Claude:

Claude automatically chains these tool calls:

1. `campaign.list(filter={type: "performance_max"})` — list all PMax campaigns
2. `metrics.by_asset_group(campaign_id=..., window="30d")` — pull per-asset-group metrics
3. `metrics.search_terms(campaign_id=...)` — pull PMax Top Search Terms (you can't see these in the UI)
4. Generate recommendations + concrete `campaign.update_asset_group_status` calls

Real case: ROAS from 1.9x to 3.4x

A home goods client had PMax stuck at 1.9x ROAS for 3 months. Claude (via MCP) discovered:

• PMax was spending 35% of budget on "home cleaning" search terms — but the client sells high-end furniture (intent mismatch)
• The "Summer promo" asset group used outdated images, CTR 0.8%
• A customer-match audience signal had expired 6 months earlier

Claude proposed 3 specific adjustments + ready-to-execute tool calls (after user approval). Two weeks later, ROAS hit 3.4x.

Why MCP + PMax is the killer combo

PMax's optimization model gives 60% of control to Google. But the assets, audience signals, and negative signals you feed it determine the ceiling Google's algorithm can reach. MCP lets the LLM see everything you can see (and more), and propose strategic adjustments in milliseconds — exactly the "human oversight layer" PMax needs most.

Security notes

• API key is used only in your local Claude Desktop process — never sent to any third party other than Anthropic
• Default approval mode: AI suggestions don't auto-execute. You must confirm before any Google Ads account is changed
• All actions are written to the audit log, visible in DEXUN dashboard → Recent changes