| name | conversion-cpa |
| description | Compute the cost per INDIVIDUAL conversion action — cost per phone call, cost per form submit, cost per booking — which Google Ads does not give natively (it only shows one blended cost/conversion across all primary actions). Splits conversion counts by campaign × action, divides by spend, and flags when a campaign's real driver doesn't match its name/goal. Use this skill whenever the user asks "what's my cost per phone call / per lead form / per [action]", per-conversion-action CPA, cost by conversion type, or which action a campaign actually drives — even if they don't say "CPA". Reports a FULLY-LOADED per-action CPA (never additive); the conversion setup verdict belongs to `conversion-tracking` and budget moves to `spend-allocation`. |
Per-Conversion-Action CPA (cost per phone call, per form, per booking)
Goal (job-to-be-done)
Answer the question clients ask constantly that Google Ads refuses to answer directly: "what does
ONE phone call / ONE form submit / ONE booking cost me?" Google only reports a single blended
cost / conversion across all counted actions. This skill rebuilds the per-action number by pulling
conversion counts split by campaign × action, dividing by spend, and — critically — labelling it
honestly so nobody adds up numbers that can't be added.
- Who: media buyer / PPC manager / the account owner. When: any time someone asks "cost per X",
or during the audit to see which action each campaign really buys.
- Decision it drives: which conversion type is cheap vs expensive, and whether a campaign is
actually producing the action it was built for.
- The differentiator: the honesty layer — a per-action CPA is fully-loaded (the same spend
produced several actions), so it is never additive, and it is a true CPA only when a campaign
optimizes to that single action. This skill states which kind every number is.
Scope
- ✅ Per-action fully-loaded CPA (campaign + account), the global CPA, the real-driver-vs-name
mismatch, and actions driven but not counted ("All conversions" only).
- ❌ Is the conversion setup correct (offline import / depth / values / how many primaries) →
conversion-tracking.
- ❌ Which budget to raise or cut →
spend-allocation.
- ❌ Per-segment (match-type / age / device) efficiency — conversion counts do not split by
those views; that's the blend guardrail in
conversion-tracking.
Components (read these references as needed)
Operate
Input: per campaign × action — conversions + all_conversions (counts); per campaign —
cost_micros + total conversions; and campaign_selective_optimization_conversion_actions to know
whether a campaign optimizes to a single action. Join by campaign_name.
Process: apply references/framework.md. For each action compute the
fully-loaded CPA = spend ÷ that action's count (campaign-level and account-level). Compute the
global CPA = total spend ÷ total conversions (the only additive figure). Mark a CPA as a true
CPA only when the campaign optimizes to that one action (single-goal). Flag a driver mismatch when
a campaign's biggest action by count isn't the one its name/goal implies, and flag actions with
all_conversions > 0 but conversions = 0 (driven but not optimized). Gate on a thin-volume floor
— don't quote a CPA off 1–2 conversions.
Emit the JSON in references/output.md:
synthesis — three strings: headline (the cheapest/most expensive action + its cost), diagnosis
(where the spend actually goes by action type), action (the one move — e.g. give the phone-call
action its own single-goal campaign to get a true CPA).actions[] (account-level) + campaigns[] (with per-action CPAs) + rollup, plus the
nonAdditiveWarning.
A renderer (the orchestrator's formats/*) turns this JSON into the human document. Emit pure data
— no emojis, tables, markdown, or colors in the output.
Voice (don't copy the rules, link them): write every line per
../../../_framework/writing.md — the question the data answers;
the metric carried as data; plain language for a non-technical owner ("each phone call costs about
$79"), the caveat ("fully-loaded — the same spend also brought in form fills") stated once, not buried.
Example (illustrative — NOT rules)
- Fully-loaded: account spend $2,830 ÷ 36 phone clicks ≈ $79 per call; ÷ 16 catering forms ≈ $177
each — but these don't add up to the spend; the same dollars bought all of them.
- Driver mismatch: a campaign named "Form Submits" actually drives 22 phone clicks vs 4 form
submits → its real product is calls, not forms — rename/retarget or split.
- Driven but not optimized: "Calls from ads" shows 19 in All conversions but 0 in Conversions →
it's happening but bidding ignores it; decide whether to count it.
