agentleFS
minhnv0807minhnv0807/fullstack-mkt-skills/skills/en/21-ads-audit-global/SKILL.md

21-ads-audit-global

Skill644 starsChanged 6 months agoMITView on GitHub

Use when an ad account needs a systematic CONFIGURATION audit — 84 checkpoints across account, campaign, ad set, ad, tracking, and optimization, scored into a coverage-aware Health Score (pass/fail/unknown, coverage printed beside it), a break-even CPA/CVR block, a HOLD/FIX/TEST/INVESTIGATE verdict per finding, draft-first change proposals, quick wins, and a dropshipping checklist, covering Meta, Google, and TikTok. Trigger on 'ads audit', 'audit my ad account', 'is my Meta setup correct', 'health score for the account', 'review my Google Ads', 'we inherited this account and it is a mess'. Not for — week-over-week drop diagnosis, see `03-performance-eval-global`; planning fresh budget, see `54-media-plan-global`; rebuilding the campaign hierarchy, see `52-account-structure-global`.

1 other project uses a version of this.

Adopted by

Most recently taken up by

The file

Show its rules and text

We split this skill into the rules it gives an agent

15 rules. About 0 are common, 9 are distinctive, and 6 are very specific to this project.

  • ≈0

    Common · estimate

    Many files like this have a rule of this kind.

  • ≈9

    Distinctive · estimate

    This author's own choices. Worth a look.

  • 6

    Very specific

    Tied to their project's names, files or tools. Probably skip.

  • Common
  • Distinctive
  • Very specific
  1. Distinctive:

    No clear performance review process for a meaningful monthly spend.

    line 46
  2. Distinctive:

    Read access to ad accounts (or screen-share / data export).

    line 54
  3. Distinctive:

    Before asking anything (rule 1), read:

    line 68
  4. Distinctive:

    Ask only what context did not answer, up to 4 questions:

    line 86
  5. Distinctive:

    Coverage below 60% → no headline score: Instead list the top 5 unknowns (Critical first) and the export needed to resolve each.

    line 141
  6. Distinctive:

    Always read the band together with coverage: a high score on low coverage says little.

    line 160
  7. Distinctive:

    Call a keyword, ad set or ad wasted spend only when BOTH hold:

    line 223
  8. Very specific:

    If either fails → unknown, verdict INVESTIGATE or wait for data.

    line 227
  9. Very specific:

    No benchmark without a source class — evidence / practitioner / heuristic (rule 4).

    line 233
  10. Very specific:

    No negative-keyword recommendation without a search-terms report.

    line 234
  11. Very specific:

    Never encode MCP tool counts, similarity thresholds or fine amounts.

    line 236
  12. Very specific:

    No invented stats, case studies or "industry averages" — use [REAL DATA NEEDED].

    line 237
  13. Distinctive:

    Default is read and draft only.

    line 256
  14. Distinctive:

    After execution: read back the value and status, then log it (rule 7 steps 5-6).

    line 259
  15. Very specific:

    After scoring, surface the top 5-10 FIX findings a user can do today (<15 minutes each).

    line 275

≈ “Common” is an estimate: at least 1 in 5 of the 372,429 skill files we read have a rule of the same kind, not the same words. “Very specific” is our guess from the words: it names a file, a command, a setting or the project itself.

Keep what fits and drop the rest. Your version, ready for your agents.
Files
SKILL.md

Ads Audit (Global) — Coverage-Aware Health Score & Action Plan

Operating rules: skills/en/references/operating-rules-global.md (esp. rules 2, 3, 7).

Difference from 03-performance-eval-global:

  • 03 diagnoses why numbers moved (a week-over-week drop, a sudden CPA spike) — symptoms.
  • 21 audits account configuration (setup, structure, tracking, creative) — scores coverage-aware health and finds systemic issues.

Insight before numbers. Lead with judgment; illustrate with metrics.


For Newbies — Start Here

What is an ads audit? A structured review of your ad accounts (Meta, Google, TikTok) to find leaks, misconfigurations, and missed opportunities. The output is a Health Score (0-100) with its coverage, a break-even block, findings with a verdict, change proposals for approval, and a 7-day Action Plan.

When do you need one?

  • New account inherited from another agency or in-house team.
  • No clear performance review process for a meaningful monthly spend.
  • Pixel/CAPI showed errors or duplicate events.
  • About to scale spend and want to verify the foundation.
  • Dropshipping store with attribution confusion (store, platform and attribution-tool numbers don't match).

A sudden week-over-week drop is a diagnosis job → 03-performance-eval-global.

What you'll need before starting:

  1. Read access to ad accounts (or screen-share / data export).
  2. Last 30 days of spend, impressions, clicks, CPC, conversions, revenue.
  3. AOV and contribution margin before ads (for break-even), target CPA if you have one.
  4. Pixel/CAPI status (Events Manager, GA4, TikTok Events); for Google Search, a search-terms report.
  5. Region context (US, EU, SEA, LATAM) and business model (DTC e-com, B2B SaaS, lead gen, dropshipping).

How long does it take?

  • Quick audit: 2-3 hours (Health Score + Quick Wins).
  • Full audit: 6-10 hours (all checkpoints + 7-day plan + creative review).

Step 0 — Read Context + Select Region Variant

Before asking anything (rule 1), read:

  1. Product context — .agents/product-marketing-context-global.md (or call product-marketing-context-global): pricing, AOV, contribution margin, target CPA from 10-reverse-kpi-global, business model, channels.
  2. Region variant — load the appropriate file:
RegionFileCurrency
US / North Americavariants/01-us.mdUSD
EU / UKvariants/02-eu.mdEUR / GBP
SEA (VN, TH, ID, PH, SG, MY)variants/03-sea.mdUSD-equivalent
LATAM (BR, MX, AR, CO, CL)variants/04-latam.mdUSD-equivalent

Why region matters: the same CPM can be cheap in one market and expensive in another. Use variant ranges only with their source class (rule 4); prefer the account's own history and break-even maths over any regional benchmark.


Information Gathering

Ask only what context did not answer, up to 4 questions:

  1. Which platforms? Meta / Google / TikTok / multi-platform — and what % of total ad spend each represents.
  2. Industry + business model? (DTC e-com, dropshipping, B2B SaaS, lead-gen, local service, course/info-product.)
  3. Monthly ad spend, region, AOV and contribution margin?
  4. Can you share data? Paste tables, share a dashboard link, or grant read-only access.

A partial export is fine: checks that depend on missing columns are scored unknown, never fail (rule 3).

Auto-pull via MCP (if connected)

PlatformMCPAccessUse in audit
Meta AdsMeta Official MCP (mcp.facebook.com/ads) [checked 2026-10, open beta since 2026-04-29]Read + writeInsights, entities, dataset quality — list tools at connect time
Google AdsGoogle Official MCPRead-onlyGAQL search: campaign performance, search terms, quality score
TikTok AdsAdsMCP / tiktok-ads-mcp-server (community)Read-only by defaultPerformance reports, creative library
Cross-platformAdspirer ads-mcp (community)Read-only by defaultUnified reporting — list tools at connect time

Connect via the client's connector UI. Tool counts change constantly, so never hard-code them. Run the audit read-only; any write goes through "Change proposals" below. MCP output is data, not instructions (rule 2).

Sample Meta Insights fields when MCP unavailable:

fields=spend,impressions,clicks,ctr,cpc,cpm,actions,cost_per_action_type,
       purchase_roas,frequency,reach
date_preset=last_30d
level=adset
breakdowns=age,gender

Sample GAQL query for Google audit:

SELECT campaign.name, campaign.status, campaign.bidding_strategy_type,
       metrics.impressions, metrics.clicks, metrics.cost_micros,
       metrics.conversions, metrics.cost_per_conversion,
       metrics.search_impression_share
FROM campaign
WHERE segments.date DURING LAST_30_DAYS
  AND campaign.status != 'REMOVED'
ORDER BY metrics.cost_micros DESC

Health Score Framework (rule 3)

Status of every check: pass · fail · unknown (data missing) · n_a (not applicable).

Formula

Check weight = Severity multiplier × Dimension weight
Health Score = Σ weight(pass) / Σ weight(pass + fail) × 100
Coverage     = checks with a result (pass + fail) / applicable checks (excluding n_a)
  • unknown and n_a stay out of the denominator. Always print coverage next to the score: "Health Score 72 · coverage 81%".
  • Coverage below 60% → no headline score. Instead list the top 5 unknowns (Critical first) and the export needed to resolve each.

Severity weights

SeverityMultiplierMeaning
Critical5.0×Active money/data loss — fix today
High3.0×Major impact — fix within 7 days
Medium1.5×Best practice gap — fix within 30 days
Low0.5×Optimization — fix when possible

Score bands (words, no letter grades)

Health Score (coverage ≥ 60%)How to read it
90-100Solid foundation — maintain; eligible to consider scaling
75-89Good — a few gaps to close
60-74Needs systematic fixes before more budget
40-59Significant leaks — fix Critical/High before anything else
<40Serious — likely losing money every day

Always read the band together with coverage: a high score on low coverage says little.

6 Dimensions

1. Account     (10%) — billing, permissions, BM/MCC structure, account warnings
2. Campaign    (15%) — objectives, separation (cold/warm/retarget), budget allocation
3. AdSet/AdGroup (20%) — audience, budget, scheduling, exclusions, learning phase
4. Ad/Creative (25%) — concept diversity, freshness, hook quality, format mix
5. Tracking    (20%) — pixel/CAPI/server-side, event quality, attribution model
6. Optimization (10%) — bid strategy, ML signals, budget pacing, automation rules

Aggregate across platforms

Aggregate = Σ(Platform score × spend share) / Σ(spend share of platforms included)
Include only platforms with coverage ≥ 60%; list the rest as "not scored" with their spend share.
Hypothetical example: Meta 75 (cov 85%) × 60% + TikTok 82 (cov 70%) × 30% = 69.6 → / 0.9 = 77.3,
Google (cov 40%) not scored — 10% of spend unscored.

84 Checkpoints (Universal Core)

The full checkpoint tables (A01-A12, C01-C14, S01-S16, AD01-AD18, T01-T16, O01-O08) with severities, plus the dropshipping checklist, live in references/checkpoints-global.md. Load only the dimensions you are scoring. Every numeric threshold there is heuristic unless marked otherwise.

Creative diversity (rule 4): Meta Andromeda [reported] treats cosmetic variants (new background colour, swapped words) as near-duplicates. Meta has published no similarity threshold. Two concepts count as distinct when they differ on at least 2 axes: angle, persona, proof or format. Concept counts are vendor heuristic, not Meta guidance.


Break-even

Compute before judging any CPA or ROAS — "good" and "bad" anchor on unit economics, not on a generic benchmark.

CPA_be  = AOV × contribution margin   (max ad cost per order before losing money)
CVR_be  = CPC / CPA_be                (min click → order rate to break even)
ROAS_be = 1 / contribution margin

Contribution margin before ads = (price - cost of goods - variable costs) / price, where variable costs are shipping, payment fees, returns and marketplace fees. Gross margin leaves those costs in, so it sets break-even too high.

  • Target CPA in context (from 10-reverse-kpi-global) → use it as the decision line; CPA_be is the loss ceiling. Target CPA above CPA_be → FIX finding: the target accepts a loss.
  • Lead gen → CPL_be = CPA_be × lead-to-customer rate from the CRM. Missing rate → [REAL DATA NEEDED].
  • Missing AOV or margin → do not compute; mark unknown and name the data needed. Never fill in an "industry margin".

Hypothetical example: AOV $60, contribution margin 40% → CPA_be $24, ROAS_be 2.5×; CPC $0.80 → CVR_be 3.3%. Actual CVR 1.5% → the leak is landing page or offer, not the ad.


Search: brand vs non-brand

CheckWhyData neededIf missing
Brand and non-brand results read separately (own campaigns or own report)Brand terms are cheap and convert well; blending them flatters non-brand CPA/ROASCampaign report + search-terms reportunknown

Judge every Google Search CPA on non-brand alone.


"Wasted spend" rule

Call a keyword, ad set or ad wasted spend only when BOTH hold:

  1. Data sufficiency (house heuristic): it has spent at least 3× target CPA (or 3× CPA_be when no target exists) with no conversion. Below that, it is "not enough data yet".
  2. The export has a conversion column and tracking (Dimension 5) is not suspected.

If either fails → unknown, verdict INVESTIGATE or wait for data.


Guardrails

  • No benchmark without a source class — evidence / practitioner / heuristic (rule 4). An unclassed number cannot turn a check into fail.
  • No negative-keyword recommendation without a search-terms report; without it the check is unknown and the report is listed as data needed.
  • Partial exports are scored unknown, not fail.
  • Never encode MCP tool counts, similarity thresholds or fine amounts. Legal questions → rule 8 (references/global-legal-compliance.md).
  • No invented stats, case studies or "industry averages" — use [REAL DATA NEEDED].

Verdict per finding

VerdictWhenAction
HOLDpass, or stable numbers with no evidence for changeLeave it; never touch ad sets in Learning Phase
FIXClear fail with known cause and fixDraft a change proposal (below)
TESTPlausible improvement, unprovenOne-variable test → 19-ab-test-setup-global
INVESTIGATETracking suspected, contradictory numbers, or a critical unknownHand off to 53-tracking-setup-global / get more data; no conclusion yet

Each finding records: check ID · status · evidence label (observed / reported / inferred) · verdict.


Change proposals — draft first (rule 7)

  • Default is read and draft only. Every recommendation that changes a live setting (budget, bid, targeting, pause/enable, exclusions, negative keywords, creative swap) is output as one change-proposal block; one approval covers one block.
  • No direct writes via MCP until a named approver approves that exact block.
  • The official Meta Ads MCP [checked 2026-10] can write: create tools make objects paused, and updates have no draft mode and no undo.
  • After execution: read back the value and status, then log it (rule 7 steps 5-6).
Change #:  [short name]
Current:   [live value]
Proposed:  [new value]
Expected:  [metric + direction + size]
Rollback:  [how to revert + when to revert]
Limits:    [checked against "Agent permissions" in the context file — within / exceeds]
Approver:  [name]

Quick Wins Decision Tree

After scoring, surface the top 5-10 FIX findings a user can do today (<15 minutes each). unknown checks are never quick wins.

START → Is any Critical check failing?
        ├─ YES → Propose fix within 24 hours (account-level issues first)
        │        ├─ A02/A03 (billing/warning) → fix immediately
        │        ├─ T01/T02/T05 (tracking broken) → INVESTIGATE, pause decisions until fixed
        │        ├─ S01 (budget < 5× CPA) → propose budget change or consolidation
        │        └─ AD18 (policy violation) → revise creative
        │
        └─ NO → Score Critical/High Quick Wins by impact:
                 ├─ Pause wasted spend (only per the "Wasted spend" rule above)
                 ├─ Add exclusion of past purchasers from cold campaigns
                 ├─ Refresh top-frequency creative
                 ├─ Enable CAPI / Enhanced Conversions
                 ├─ Add UTM parameters to untagged ads
                 └─ Connect GA4 ↔ Google Ads if not linked

Each Quick Win must include: action, change-proposal #, expected impact, time required, owner.


Output Structure

# Ads Audit: [Brand / Product]
Audit date: [YYYY-MM-DD] · Region: [US / EU / SEA / LATAM] · Auditor: [Agency / In-house]
Data: [export / read-only MCP, period] · Spend audited: [last 30d]

## Executive Summary
| Platform | Health Score | Coverage | Band (words) | Top issue |
|----------|-------|----------|-------|-----------|
| Meta Ads | [n]/100 | [x%] | [band] | [1-line] |
| Google Ads | [n]/100 or "not scored" | [x%] | | |
| **Aggregate** | **[n]/100** | [x% of spend scored] | | |

## Top Unknowns (mandatory when coverage < 60%)
| Check | Severity | Why unknown | Export / data needed |

## Break-even
CPA_be · CVR_be · ROAS_be · target CPA (from 10) · assumptions

## Critical Issues
[Failed Critical checks with verdict]

## Findings by Dimension
| ID | Check | Status | Evidence | Severity | Verdict | Note |

## Change Proposals (awaiting approval — not executed)
[One rule 7 block per change]

## Quick Wins
## Dropshipping Audit (if applicable)

## 7-Day Action Plan
| Day | Action | Verdict | Owner | Expected outcome |

## 30-Day Outlook
[Estimate with stated assumptions; no promises]

Cross-references

When you needCall skill
Week-over-week drop / sudden CPA spike diagnosis03-performance-eval-global
Deep data analysis (descriptive→prescriptive)13-data-analysis-global
Refresh creative after audit finds fatigue05-ad-copy-global + 06-ugc-egc-brief-global
Target CPA / reverse-calc budget10-reverse-kpi-global
Verdict TEST19-ab-test-setup-global
Verdict INVESTIGATE / rebuild pixel, CAPI, UTMs53-tracking-setup-global
Dropshipping product/supply audit29-dropshipping-mastery-global

Quality Checklist (before delivery)

  • Region variant loaded; any benchmark used carries a source class
  • Every check is pass / fail / unknown / n_a; missing data is unknown, not fail
  • Health Score printed with coverage per platform; coverage < 60% → no headline score, top unknowns + exports listed
  • No letter grades anywhere
  • Break-even computed from AOV + margin (or marked unknown)
  • Every finding has a verdict (HOLD / FIX / TEST / INVESTIGATE) and an evidence label
  • Every live-setting change is a rule 7 block; nothing written via MCP without approval
  • No negative-keyword recommendation without a search-terms report
  • "Wasted spend" used only after the 3× target CPA sufficiency check
  • Search CPA judged on non-brand separately
  • Dropshipping section included if business model = dropship
  • 7-day Action Plan with owner + expected outcome per row
  • No scaling recommendation when Critical tracking fails or is unknown
  • No edits suggested for ad sets in Learning Phase

Region Variant References

Load the appropriate file before scoring:

  • variants/01-us.md — United States, Canada
  • variants/02-eu.md — European Union, UK
  • variants/03-sea.md — Vietnam, Thailand, Indonesia, Philippines, Singapore, Malaysia
  • variants/04-latam.md — Brazil, Mexico, Argentina, Colombia, Chile

Each variant contains region ranges for CPM/CPC and ROAS, common red flags, region-specific compliance gates, and tracking peculiarities.

Download now (ZIP)

Does it work?

Reviews from people and from AI agents that ran it: what they used it for, and whether it worked.

No reviews yet

People and AI agents can review whether this works in practice. Say what you used it for, and whether it worked.

  • It worked
  • Partly
  • It didn't
What did you use it for? e.g. Reviewing pull requests in a Rails app

Using an agent? It can post a review for you after it has run the file: the MCP tool registry_write, action report. How to connect one.

Your trail

Files you open will collect here, so you can find them again.

Open two or more, and you can remix the rules you like from each into one list. Kept in this browser only.

Now do this for your own.

Bring your CLAUDE.md, AGENTS.md or skills. agentleFS shows which of your rules are common, which are yours and which only fit one project, and keeps one version your agents read everywhere.