paid-media-analysis

Paid-media analysis

Use this skill for performance questions, audits, comparisons, diagnosis, budget reasoning, and recommendations.

1. Read /skills/company-context/SKILL.md when present for this organization's goals, targets, conversions, and naming. Ask for missing facts the analysis needs; do not guess. Then read /skills/paid-media-wiki/decision-model.md and the page the question calls for: benchmarks.md for "is this good", anomaly-and-significance.md for spikes and drops, bidding-and-budget.md for pacing or budget changes, platform-playbooks.md for a platform's grains and caveats, and answer-style.md before the final answer.
2. Establish goal, account scope, entity grain, date window, comparison, timezone, and currency. Comparison windows must have the same day count; compare_periods rejects unequal windows. Resolve relative windows one way and say which: "last week" is the most recent complete Monday to Sunday week; "last N days" ends on the latest date the platform reports as complete (data_complete_through), not today; "this month" is the calendar month to date. When a platform's data ends inside the requested window, keep the requested window in the answer and name the missing days rather than silently shrinking it.
3. Call list_accounts for aliases, then discover_tools with keywords. Never invent a tool name. Platform tools are named <platform>__<tool> and take account_alias, never a provider id. Pipeboard loads all tools exposed by its eight configured MCP servers. Search the live catalog; availability depends on connected accounts and host policy. GA4 uses property aliases.
4. Pull the smallest complete data using the performance/report tool returned by discovery, for the union of both windows. get_campaign_performance is a fixture tool, not a universal live name. Go one grain lower only when the question needs it: get_ad_group_performance (ad sets, line items) or get_creative_performance where the platform exposes it; rows carry the parent campaign id. Run independent platform reads in parallel. Each read returns a compact read_result with an artifact_id, row count, actual window, missing fields, and flags. Native analytics and platforms without verified spend-unit mappings stay as provider_result artifacts. Do not pass them to spend comparisons or treat GA4 conversions as ad-attributed conversions.
5. Validate source coverage with references/validation-checklist.md.
6. For pacing, anomalies, top spenders, or per-entity efficiency inside one window, call summarize_window with the performance artifacts (and the list_campaigns artifacts for daily budgets); it returns per-entity totals, pacing, and a daily series with flagged days. For period-over-period change, call compare_periods with the artifact ids and both windows. List any failed read in unavailable_sources so it stays visible and suppresses the cross-platform total.
7. Read the analysis_summary. Quote its values verbatim; never recompute from previews or rows. unavailable means missing, not zero.
8. Explain observation, business meaning, likely drivers, confidence, and next action separately.
9. Include a measurement and reversal plan for any recommendation.
10. Create a proposal only when the user asks to change provider state (see paid-media-writes).

Do not use universal performance thresholds. Use configured goals or label the analysis as directional.