google-ads-audit

Google Ads — Audit

Read-only diagnostic. Score the account, surface issues by severity, and — critically — never fire a finding on an unverified assumption. Reuse the live MCP detection proven in setup.

STEP 0 — Load context (MANDATORY, before any scoring)

1. Read account-context.yaml from the working directory. If missing, run setup first.
2. Pull from it: google_ads.customer_id, brand_terms, competitor_terms, margin_tiers, data_source (for landing-page URL checks), measurement, and guardrails.
3. Output is written to the working directory (GOOGLE-ADS-AUDIT.md), never into the plugin.

⚠️ The six guards (apply to EVERY check — they prevent the most common wrong findings)

These are generic; guardrails in the context may strengthen or add to them. Honor both.

• GUARD-1 Conversions — check the WHOLE picture, not the action list. Inspect campaign-level conversion goals + which actions each active campaign actually fires (segment by segments.conversion_action_name), and each action's source. A primary action no campaign uses is inert; a second purchase action kept secondary (e.g. GA4 purchase alongside the Shopify channel "App Purchase") is the CORRECT anti-double-count setup, not a bug. Only flag a real micro-conversion goal or two value-carrying purchase sources both primary and both firing. Before flagging ANY conversion-action contamination (out-of-brand / micro / duplicate), check primary_for_goal. A non-primary action does not enter Smart Bidding and does not affect performance — it's harmless; surface it as a NOTE (tidy-up), never a scored finding. Only a PRIMARY out-of-brand/micro action that an active campaign optimizes toward is a real issue. See ${CLAUDE_PLUGIN_ROOT}/references/conversion-tracking-logic.md.
• GUARD-2 Negatives — FOUR sources, merge ALL. (1) campaign_criterion (negative=true) — campaign-level; (2) campaign_shared_set → shared_set → shared_criterion — SHARED LISTS (the bulk: e.g. "Account Level Negative", "OPI Negative", brand blocks); (3) customer_negative_criterion — ACCOUNT-LEVEL negatives applied account-wide; (4) campaign_criterion.brand_list.shared_set — brand exclusions. Only flag a gap if terms are absent across ALL FOUR; if a shared/account list covers them, cite it, don't flag. A campaign_criterion-only pull is the #1 false-positive in this audit — it looks "thin" when coverage is actually 100+ terms/campaign. After merging all four, declare pulled:[...,"negatives_complete"] so the report renders a real verdict; without that assertion a campaign-only bundle renders VERIFY (re-pull), never a false "weak negatives".
• GUARD-3 Budget scaling — check change_event first. Before recommending a budget/bid increase, pull recent change history (~14d). A recent budget/bid/asset/structure change → recommend cooldown, not scale. (Strengthen with any change-event-cooldown guardrail in context.)
• GUARD-4 Ad strength — confirm ENABLED + impressions. Before flagging POOR ad strength, confirm ad_group.status AND ad_group_ad.status are ENABLED and the group has impressions. Paused = intentional; do not flag.
• GUARD-5 ROAS — apply margin tiers from context. Read margin_tiers; a line with a tier accepts its min_roas (house brands run lower ROAS by design). Never apply a generic ROAS threshold to such a line or a strategic campaign. If unsure, ASK rather than flag.
• GUARD-6 No false "absence". MCP results can be truncated (token cap or your limit). Before claiming "no video / no signal / no negative / no X", re-query filtered on the enum NAME (e.g. field_type = 'YOUTUBE_VIDEO') or pull with returned-rows < limit (proof of completeness). NEVER infer presence/absence by decoding enum integers from memory. Track failed fetches as a G-SYS1 diagnostic; never silently skip a check — explain why data is unavailable.