Ad Campaign Analyzer

Ad Campaign Analyzer

Composite: ingest/normalize → diagnostics → significance → funnel → reallocation → report. This is fundamentally a reasoning skill: the scripts normalize the data and compute statistical significance deterministically; you (the agent) do the diagnostics, verdicts, funnel diagnosis, and the reallocation plan.

When to use

• "Analyze my Google/Meta Ads performance" / "Which ads should I kill?"
• "Where am I wasting ad spend?" / "Is this campaign working?"
• "How should I split my budget across channels?" / "Google or Meta?"

How to run

1 — Normalize the data + build the channel rollup

If the input is dashboard screenshots, first OCR/transcribe them into a CSV or JSON (the agent does this), then:

python3 ${SKILL_DIR}/scripts/normalize_campaigns.py --csv ${WORKSPACE}/perf.csv \
  --funnel ${WORKSPACE}/funnel.json \
  --output ${WORKSPACE}/normalized.json

Maps the many possible column names to a standard schema, derives CTR/CPC/conv-rate/CPA/ ROAS where computable, rolls up per channel, and (when --funnel is given with lead_to_mql / mql_to_sql / sql_to_close / avg_deal_size) computes funnel-adjusted CAC per channel. --funnel is optional.

2 — Diagnostics (you, the agent)

Over normalized.json: health-check each item vs. benchmarks (CTR, CPC, conv rate, CPA, ROAS, impression share); detect waste (zero-conversion items, >3× CPA outliers, low-CTR ads, broad-match bleed); identify winners (lowest CPA / highest conv-rate keywords, ads, audiences, times). Label each Scale / Optimize / Pause.

3 — A/B significance (deterministic — do NOT eyeball it)

For any A/B pair, call the test so verdicts rest on real math, not a guessed p-value:

# CTR test (two-proportion z-test)
python3 ${SKILL_DIR}/scripts/ab_significance.py --metric ctr \
  --a-clicks 320 --a-impr 12000 --b-clicks 410 --b-impr 12500

# CPA / conversion-rate test
python3 ${SKILL_DIR}/scripts/ab_significance.py --metric cpa \
  --a-conv 45 --a-clicks 900 --b-conv 60 --b-clicks 880

It enforces minimum samples (100 clicks/variant for CTR, 30 conv/variant for CPA) and returns enough_data:false when below them. If enough_data is false, report "not enough data" — never force a verdict. You only narrate the returned result.

4 — Funnel + reallocation (you)

Diagnose the biggest impression→click→conversion→revenue drop-off. Then build a budget-neutral cross-channel reallocation: efficiency index per channel, marginal- return / saturation read, funnel-stage coverage gaps, and a dollar-shift table. Keep it budget-neutral unless the user explicitly asks for a budget-increase scenario. Where target CPA/ROAS weren't supplied, fetch category benchmarks via web search (Robomotion Proxy): if DATAFORSEO_LOGIN/DATAFORSEO_PASSWORD (or SERPER_API_KEY) is set → structured SERP API; else → the agent's keyless web search (default).

5 — Scenarios + render

Model conservative (±20%), aggressive (±40%), and (if requested) budget-increase scenarios with projected conversions/CPA. Write campaign-analysis-<YYYY-MM-DD>.md to ${WORKSPACE} and attach to the Agent Teams channel.

Outputs

campaign-analysis-<YYYY-MM-DD>.md — performance dashboard, budget-waste report ($ + items), winners to scale, A/B significance verdicts, funnel drop-off diagnosis, channel efficiency ranking, current-vs-recommended allocation table, 3 scenarios, phased action plan.

Credentials / env

• Required: none — the core path is data-in → analysis-out; normalize_campaigns.py and ab_significance.py are stdlib and keyless. The verdicts/reallocation are the agent's reasoning (no LLM key in scripts).
• Optional (each with a keyless default fallback):
  • CRM keys (HUBSPOT_API_KEY / SALESFORCE_ACCESS_TOKEN / PIPEDRIVE_API_TOKEN) — if set → pull live conversion/close data from the CRM; else → use the supplied performance file/CSV (default).
  • DATAFORSEO_LOGIN/DATAFORSEO_PASSWORD (or SERPER_API_KEY) — if set → structured SERP API for missing-benchmark lookups; else → the agent's keyless web search (default).

Notes & edge cases

• No scraping in the core path; a web search is used only to fetch missing benchmarks.
• Significance is computed deterministically (z-test via stdlib math.erf), never by the LLM — this avoids hallucinated p-values.
• Refuse to over-claim on thin data: surface "not enough data" rather than a forced verdict.
• Keep reallocation budget-neutral unless a budget-increase scenario is explicitly requested.