Pmax Decoder

PMax Decoder

Generate operational transparency for Performance Max campaigns and convert that analysis into concrete actions.

Account Context

Read ${CLAUDE_PLUGIN_ROOT}/profile/account-profile.md at the start of every run. If it exists:

• Use known account IDs — skip list_accounts discovery.
• Apply KPI targets as anomaly detection thresholds (e.g., flag CPA > target CPA).
• Note active tests when interpreting performance shifts.
• Check watch list for follow-up items from prior sessions. If it doesn't exist, fall back to list_accounts and suggest running platform-setup.

Data Access

This skill uses the plugin's Google Ads MCP tools for live API access:

• mcp__google-ads__query: Execute GAQL SELECT queries for PMax insights.
• mcp__google-ads__mutate: Preview and apply negative keyword actions.
• mcp__google-ads__list_accounts: Confirm account access and campaign scope.

Mutation safety flow:

1. Build operations and run mcp__google-ads__mutate with dry_run: true.
2. Show proposed changes and rationale.
3. Get explicit approval.
4. Re-run with dry_run: false only after approval.

Workflow

Module 1: Search term extraction

The campaign_search_term_insight resource requires single-campaign filtering.

1. Run Module 1A query to list all PMax campaigns.
2. For each campaign ID, run Module 1B query to fetch insight categories.
3. Combine categories across campaigns and rank by metrics.clicks.
4. Drill into top categories (default cap: 50) using Module 1C query.

Warn users that large PMax accounts can require many API calls.

Module 2: Channel distribution (requires API v23+)

segments.ad_network_type on asset_group returns meaningful channel data (SEARCH, YOUTUBE, DISPLAY, SHOPPING) only in API v23+.

For dates before June 1, 2025, this dimension can return MIXED and is not reliable for channel decomposition. If data is pre-v23 behavior, skip this module and note the limitation.

Use asset_group with segments.ad_network_type to estimate spend and conversion mix by channel. Flag concentration risks when one channel exceeds 70% of spend.

Module 3: Asset group and asset label review

• Summarize asset-group level performance (impressions, clicks, cost, conversions).
• Pull asset_group_asset.performance_label per asset.
• Labels are relative rankings (BEST, GOOD, LOW, PENDING), not direct cost metrics.
• Recommend replacement priorities for LOW assets. Do not generate creative copy unless requested.

Module 4: Brand traffic detection

• Require user-provided brand terms.
• Classify Module 1C search terms against that list.
• campaign_search_term_insight does not include cost metrics; use click share as the proxy.
• Flag cannibalization risk when brand click share exceeds 30% of PMax clicks.
• Generate campaign-level negative keyword operation previews with dry_run: true.

Module 5: Placement review

Use performance_max_placement_view for inventory visibility.

Placement view returns impressions only (no clicks, cost, or conversions). Treat findings as quality-risk diagnostics and do not estimate placement-level cost or CPA.

Output format

## PMax Decoder - [Date]
### Campaign: [Name] ([ID])

### Quick Stats (from Module 1A campaign data)
- Spend, conversions, CPA (cost/conversions), ROAS (conversions_value/cost)

### Channel Distribution
| Channel | Spend | % Spend | Conversions | CPA |
|---|---:|---:|---:|---:|

### Brand Traffic Analysis
- [brand click share, click volume, top branded terms]
- [negative keyword package preview]

### Search Term Insight Categories
| Category | Clicks | Conversions | Notes |
|---|---:|---:|---|

### Asset Performance
| Asset Group | Ad Strength | LOW Assets | Action |
|---|---|---:|---|

### Placement Summary
- [top placement types and quality risks]

### Recommended Actions
1. [prioritized action]

Guardrails

• campaign_search_term_insight requires single-campaign filtering.
• Channel-level segments.ad_network_type data is reliable only for dates after June 1, 2025.
• Placement view provides impressions only; do not claim placement-level cost.
• Never execute live negatives without explicit user confirmation.
• Empty results: when a module query returns zero rows, report it explicitly rather than silently omitting sections.

References

• references/gaql-queries.md