Campaign Analysis

Campaign Analysis & Knowledge Accumulation

Analyze campaign performance across Google Ads, Meta, GA4, and Search Console. Build accumulated knowledge over time. Investigate root causes before recommending actions. Track whether past actions worked.

Two Modes

Monthly review: Full flow through all steps — context, data, analysis, investigation, action plan, knowledge.

Ad-hoc investigation: Jump to Step 3 with a focused question ("why is CPL high?", "investigate these search terms"). Still must pass through the investigation gate before recommending actions.

CLI Quick Reference

Global flags: --format json, --brand <name>, --month YYYY-MM, --days N

For full output schema, load references/output-schema.md.

Analysis Workflow

Step 1: Load History and Context

Before looking at any data, load what is already known:

campaign --format json check <client_id> --brand <brand>

The data package includes knowledge (knowledge.md contents), memory (past actions), and context (client config). Read these FIRST.

For each prior finding in knowledge.md:

1. Does the current data confirm or contradict it?
2. Has the situation evolved?

Frame the analysis around these questions — not as a fresh discovery exercise.

Step 2: Evaluate Past Actions (MANDATORY if actions exist)

If memory contains executed actions, invoke the campaign-reviewer skill before proceeding. This is a hard gate — do not skip to new analysis without first measuring what past actions achieved.

The campaign-reviewer evaluates each action: what was done → expected impact → enough data? → current vs expected → verdict (WORKING / PARTIAL / NOT WORKING / TOO EARLY).

Incorporate the verdicts into the analysis:

• WORKING actions → validated approach, consider scaling
• PARTIAL actions → investigate what's missing, look for problem shifting
• NOT WORKING actions → revise root cause hypothesis, do not repeat the same approach
• TOO EARLY actions → note review date, do not take additional actions on the same metric

Step 3: Analyze Current Data

With history and outcomes established, analyze current data. Prioritize:

1. Findings related to past action outcomes — If Scout was PARTIAL, what are the new junk clusters?
2. Prior finding validation — Do known patterns still hold? (CONFIRMED / CONTRADICTED / EVOLVED)
3. New discoveries — What appears in the data that wasn't known before?

For cross-source reasoning patterns (CPL diagnosis, CVR diagnosis, Search Terms + Search Console connections), load references/cross-source-playbook.md.

Step 4: Apply Analytical Judgment

The data package contains raw metrics without pre-filtering. Apply judgment to determine what matters. Key principles:

Search terms: Intent > metrics. Cluster by theme. Consider WHY terms matched (match type problem vs missing negatives). Junk ratio > 40% is systemic, not individual bad terms.

Budget: Portfolio decision. Only increase budget on efficient (low CPL) AND budget-constrained campaigns. IS lost to rank = QS/bid problem, not budget.

Quality Score: Read all 3 components together. Cross-reference with GA4 engagement. QS < 5 on high-spend = urgent.

Trends: R-squared < 0.5 = directional only. Check anomaly dates against known events.

For detailed analytical frameworks, benchmarks, and the band-aid vs structural fix matrix, load references/analytical-principles.md.

Step 5: Investigation Gate (MANDATORY)

STOP. Before recommending any action, investigate root cause for every significant finding.

This is not optional. Spawn a subagent to investigate independently:

Subagent prompt: "Given this data [attach relevant metrics], investigate
why [specific anomaly]. Consider: campaign structure, match type settings,
keyword selection, audience targeting, landing page alignment.
Return: root cause hypothesis with supporting evidence."

While the subagent investigates, continue investigating from a different angle. When the subagent returns:

• If both analyses agree → proceed to action plan with confirmed root cause
• If they disagree → state both hypotheses, identify what data would resolve the disagreement
• If data is insufficient → recommend investigation actions (e.g., "audit match types in Google Ads UI"), not optimization actions

Every recommended action must trace back to a root cause. Format:

• Root cause: [confirmed or hypothesis]
• Structural fix: Change that prevents recurrence
• Immediate action: Band-aid for symptom relief now

Step 6: Action Plan

Present findings as a prioritized action plan:

### [URGENT|WATCH|OPPORTUNITY] Title
- **Evidence:** Data with source citation
- **Root cause:** From Step 4 (confirmed or hypothesis + what data needed)
- **Structural fix:** Prevents recurrence
- **Immediate action:** Addresses symptom now
- **Expected impact:** Quantified where possible

Severity: URGENT (>20% KPI change + confirmed root cause), WATCH (10-20% or unconfirmed), OPPORTUNITY (improvement possible, no regression).

Step 7: Write Knowledge

Append to data/<client_id>/knowledge.md:

## YYYY-MM-DD
- CONFIRMED [prior date]: "[Finding] still holds. [Current data]."
- CONTRADICTED [prior date]: "[Finding] no longer holds. [New data]."
- OUTCOME [action date]: "[Action taken]. Result: [WORKING|PARTIAL|NOT WORKING|TOO EARLY]. [Evidence]."
- NEW: "[Finding]. [Evidence from which sources]."

Write: cross-source patterns, action outcomes, demand signals, revised understanding. Do not write: raw data points, one-time observations.

Step 8: Generate Reports

campaign brief <client_id>

Additional Resources

Reference Files

• references/cross-source-playbook.md — Multi-brand analysis, cross-platform patterns, GA4/SC benchmarks, CPL/CVR/volume investigation reasoning
• references/output-schema.md — JSON output structure for campaign check and campaign investigate
• references/analytical-principles.md — Deep reasoning frameworks: search term intent classification, budget portfolio optimization, root cause investigation, band-aid vs structural fix matrix, outcome tracking patterns