# Meta Ads Audit

> Use this skill when someone wants a Meta Ads (Facebook and Instagram) account audited or optimized at senior-media-buyer depth: account health, which conversion metric to run on, campaign structure, ad set bid and audience splits, top-performer analysis, or video creative diagnosis. Runs a health check on event match quality and audience segments, sets the primary conversion metric by account type, detects the active campaign structure, applies Pareto analysis to ads, and reads hook rate and hold rate to diagnose creative. Read-and-recommend only: it never changes the account, the user applies every recommendation in Ads Manager. Trigger phrases: Meta Ads audit, creative fatigue, which ads to pause, ad set analysis, hook rate, pixel issues, why did ROAS drop.

- Skill: `swan-gtm/meta-ads-audit` (Agent Skill)
- Install (CLI): `npx get-skill add swan-gtm/meta-ads-audit`
- Raw SKILL.md: https://api.skillmd.com/api/skills/swan-gtm/meta-ads-audit/raw
- Safety review: pending
- Works with: Claude Code, Claude.ai, OpenAI Codex
- Category: Marketing & Growth
- Author: swan-gtm (https://skillmd.com/u/swan-gtm)
- Updated: 2026-11-14
- Page: https://skillmd.com/skills/swan-gtm/meta-ads-audit

---


> Adapted from the GoMarble Paid Media Claude Skills repository (https://github.com/gomarble-ai/gomarble-ai-paid-media-claude-skills) by its author, shared under GoMarble's attribution license. Works with any data source: a connected ads MCP such as GoMarble, CSV exports, screenshots, or pasted tables.

A consolidated set of skills for analyzing and optimizing Meta Ads (Facebook + Instagram) accounts. These skills are **framework-agnostic** - the analytical logic (Pareto, Hook/Hold scenarios, decision matrices, guardrails) applies regardless of where the data comes from.

The skills define **what to know, what data to look for, how to interpret it, and how to recommend action**. They do not depend on any specific tool integration.

> **Read-and-recommend only.** These skills never make changes to the user's Meta Ads account. Claude reads data and produces recommendations in plain language; the user applies the recommended changes themselves in Meta Ads Manager.

---

## Data Sources Are Not All Equal

Data can come from four kinds of sources, but they have **very different completeness**. Some skill rules silently break on weaker sources. Always identify the source first and consult the **Data Source Compatibility Matrix** below before applying any rule.

| Priority | Source | Notes |
|---|---|---|
| 1 (preferred) | A connected MCP server exposing Meta Ads data (e.g. GoMarble, Meta official) | Returns raw API fields, action arrays, full targeting, breakdowns. All rules work. |
| 2 | A CSV/Excel export from Ads Manager | Most rules work, but a few are inert - see compatibility matrix. Several signals require *separate* exports (breakdowns, video metrics, ad-level vs ad-set-level views). |
| 3 (narrow questions only) | Screenshots of report tables | Use only for spot questions on metrics that are visibly captured in the screenshot. Skip deep analysis. |
| 4 (last resort) | Direct copy-paste of metric values | Treat as user-asserted; ask for the underlying export when stakes are real. |

---

## How to Use These Skills

1. **Identify the user's intent** - are they asking for an audit, a performance review, a creative deep-dive, or a targeted question? Use the routing table below.
2. **Run the Data Inventory step (Step 0 below)** - confirm what source is being used and what's actually available before invoking any skill.
3. **Gather the required data** - ask the user to share the data the skill needs in their preferred source. The "Minimum required data" section of each skill lists what's needed.
4. **Apply the analytical framework** - classify, diagnose, decide based on the rules in the relevant skill.
5. **Always check guardrails** before recommending any action.
6. **Output recommendations** in the structured format defined per skill.

If required data is missing or unverifiable, say so explicitly and ask the user for it. **Never fabricate metrics or invent thresholds.**

**Always work from current data** - campaigns change frequently (budget edits, paused ads, new creatives). If the user shared data more than a day or two ago, ask for a fresh pull before making recommendations.

---

## Step 0: Data Inventory (Mandatory Before Any Skill)

Run this once at the start of every analysis. It takes one short turn and prevents wasted effort.

1. **Identify the data source** - MCP / CSV / screenshot / copy-paste. If MCP is connected, use it. If the user attached a file, inspect its columns/fields before asking for more.
2. **State what's available and what's missing** for the requested skill, based on the Data Source Compatibility Matrix below.
3. **For each missing signal, decide:**
   - **Blocking** - cannot proceed without it; ask user to fetch.
   - **Downgrade** - can proceed with caveats; flag as low-confidence in the output.
   - **Skip** - rule doesn't apply to this source (e.g. Purchase De-Duplication on CSV); note it and move on.
4. **Confirm the date range** in the data matches what the skill expects (typically 30 days; 7d for ad set analysis; 60–90d for trend or Hold Rate baseline).

Output one short paragraph before applying the skill. Example:
> *"Using your GoMarble MCP. I have ad-level insights with `actions` and `purchase_roas` for the last 30 days - sufficient for Performance Analysis (Skill 5). The EMQ check (Skill 1) requires you to look in Events Manager directly; I'll flag it as a manual step."*

---

## Data Source Compatibility Matrix

Use to plan which signals you can get from which source before running any skill.

| Signal | MCP (GoMarble / Meta / etc.) | CSV (Ads Manager export) | Screenshot |
|---|---|---|---|
| Spend, impressions, clicks, CTR, CPM, reach, frequency | ✅ direct | ✅ default columns | ✅ if shown |
| Purchase count (deduplicated) | ✅ via `actions` array - apply Purchase De-Duplication rule | ✅ "Purchases" column (Meta already deduplicates - **skip the dedup rule**) | ⚠️ if shown |
| Purchase ROAS | ✅ `purchase_roas` field | ✅ "Purchase ROAS" column (must be enabled) | ⚠️ |
| Revenue / conversion value | ✅ `action_values` | ✅ "Purchases Conversion Value" column | ⚠️ |
| Video Plays | ✅ `video_play_actions` | ✅ "Video Plays" column | ❌ rarely shown |
| 3-Second Video Plays | ✅ via `actions` (`video_view` action) | ✅ "3-Second Video Plays" column (already pre-computed - **skip the action-name distinction**) | ❌ rarely shown |
| ThruPlays | ✅ via `actions` | ✅ "ThruPlays" column | ❌ rarely shown |
| Hook Rate / Hold Rate | ❌ must compute from the above | ❌ must compute from the above | ❌ |
| Promoted object / `custom_event_type` | ✅ from ad set object | ❌ NOT in default CSV exports - **ask user to confirm per ad set / campaign** | ❌ |
| Campaign objective, ad set bid strategy, bid amount | ✅ direct | ⚠️ requires "Bid Strategy" + "Bid Amount" columns enabled (often not default) | ⚠️ |
| Full targeting object (audiences, interests, custom audiences) | ✅ via ad set object | ❌ NOT in CSV - **must be described by user or pulled via MCP** | ⚠️ if visible |
| EMQ Score | ❌ not in standard insights APIs - **manual check in Events Manager** | ❌ same | ❌ same |
| Audience Segment breakdown (New / Engaged / Existing) | ✅ insights with `breakdowns=['user_segment_key']` | ⚠️ requires a *separate* breakdown export | ❌ |
| Placement breakdown (Feed / Stories / Reels / etc.) | ✅ `breakdowns=['publisher_platform','platform_position']` | ⚠️ requires separate breakdown export | ❌ |
| Age / gender breakdown | ✅ `breakdowns=['age','gender']` | ⚠️ separate export | ❌ |
| Geographic / device breakdown | ✅ via breakdowns | ⚠️ separate export | ❌ |
| Ad creative content (video file, image, copy) | ✅ via creative endpoint (e.g. `facebook_get_ad_creative_details`) | ❌ never in CSV - **must be shared separately** | ⚠️ |
| Comments / sentiment | ❌ requires manual check in Ads Manager or the Page | ❌ same | ⚠️ if shown |
| Attribution window settings | ✅ via ad set `attribution_spec` | ⚠️ requires "Attribution Setting" column | ❌ |
| Daily time-series (for trend / fatigue detection) | ✅ insights with `time_increment=1` | ⚠️ requires day-by-day export | ❌ |
| 3rd-party tracker config (Profitmetrics / Triple Whale etc.) | ✅ inferred from `custom_event_type` on PURCHASE campaign objective | ❌ **must ask user explicitly** - see upgraded check in Skill 5 | ❌ |

✅ = available and reliable. ⚠️ = available but requires extra export / setup. ❌ = not available from this source.

---

## Meta CSV Column Translation

When the source is a CSV export from Ads Manager, the column names rarely match the API field names used throughout the skills. Use this table to translate. (Column names vary slightly by language and export template - match by meaning, not by exact string.)

| Skill / API reference | Typical Ads Manager CSV column |
|---|---|
| `spend` | "Amount spent (USD)" / "Amount Spent" / "Spend" |
| `impressions` | "Impressions" |
| `reach` | "Reach" |
| `frequency` | "Frequency" |
| `cpm` | "CPM (cost per 1,000 impressions)" |
| `ctr` | "CTR (all)" or "CTR (link click-through rate)" |
| `clicks` | "Link clicks" or "Clicks (all)" |
| `purchase` / `omni_purchase` (deduped) | "Purchases" (Meta has already deduplicated) |
| `purchase_roas` | "Purchase ROAS (return on ad spend)" |
| Revenue (conversion value) | "Purchases Conversion Value" |
| `video_view` action (3-second views) | "3-Second Video Plays" (already pre-computed) |
| ThruPlays | "ThruPlays" |
| Video Plays | "Video Plays" |
| Leads | "Leads" or "On-Facebook Leads" + "Off-Facebook Leads" - confirm with user which is the conversion event |
| Campaign objective | "Campaign Objective" (often not in default export - must enable) |
| Bid strategy | "Bid Strategy" (often not in default export - must enable) |
| Daily / lifetime budget | "Budget" or "Campaign Budget" / "Ad Set Budget" |
| Attribution window | "Attribution Setting" |
| `custom_event_type` / promoted_object | ❌ NOT in CSV - ask user to confirm |

**Rule for Claude:** if a CSV column maps to a pre-computed metric (Purchases, 3-Second Video Plays, ThruPlays), use it directly and skip the upstream dedup / action-array rules. Those rules exist to clean raw API data.

---

## Routing - Which Skill to Use

| User asks about | Use Skill |
|---|---|
| "Audit my account" / full optimization request | Run **Initial Health Check** → **Conversion Metric Focus** → **Performance Analysis** → **Creative Analysis** in that order |
| Pixel / tracking setup, audience segments setup | **Initial Health Check** |
| "What's my account type?" / which KPI to use | **Conversion Metric Focus** |
| Campaign structure, primary vs testing, bid splits, audience splits | **Campaign Structure Analysis** |
| Ad set targeting, budget utilization, audience comparison | **Ad Set Analysis** |
| Top performers, Pareto analysis, which ads to keep/pause | **Performance Analysis** |
| Hook rate, hold rate, video diagnostics, why creative isn't working | **Creative Analysis** |
| Change requests (pause campaign, increase budget, change CTA, etc.) | Apply the relevant analysis skill first, then **recommend** the change in plain language with rationale. Claude does not execute changes on the user's ad account - the user applies recommendations manually in Ads Manager. |

If the user's intent is ambiguous, ask which to focus on first.

---

## Universal Guardrails - Apply Before Every Recommendation

**Read these before generating any recommendation.**

### Budget Reallocation Prohibition

You **cannot** allocate, shift, increase, reduce, or scale budget for: individual ads (only campaigns and ad sets have budgets), placements (Meta controls distribution), demographics (age, gender, location breakdowns), user segments (new, existing, engaged).

**Before writing any recommendation, ask:** Does it imply moving money between ads, placements, demographics, or segments? If yes → rewrite using the valid controls below.

**Valid budget controls (only three):**

1. Campaign budget (CBO) or ad set budget (ABO)
2. Bid caps / cost caps
3. Turning ads / ad sets on or off

If a recommendation does not map to one of these three controls, it is not actionable.

**Mandatory rewrites:**

- ❌ "Reallocate $X from [ad A] to [ad B]"
- ✅ "Pause [ad A]. Increase the campaign/ad set budget by 15–20% to give [ad B] more spend opportunity."

- ❌ "Allocate more to [placement / demographic / segment]"
- ✅ "Create a new ad set targeting [dimension] with its own budget, OR exclude underperforming [dimension] via ad set placement settings."

**Rule:** Breakdowns (age, gender, placement, device, region, user segment) show how Meta distributed YOUR budget. They are not levers the user controls. The only way to influence them is creating new ad sets with specific targeting or placement selections.

### Metric Selection by Conversion Type

- **Ecommerce** (purchases / revenue): Use ROAS, revenue, AOV, cost per purchase. Never use ROAS for leadgen.
- **Lead gen** (form submits, leads): Use CPL, cost per qualified lead. Never use ROAS unless offline revenue is mapped.
- **App** (installs, in-app events): Use CPI, cost per activation. Never use ROAS unless in-app revenue tracking is confirmed.
- **Custom conversion**: Use cost per result for that specific event. Do not mix in other conversion types.
- **Unclear**: Ask the user. Do not assume.

### User Segment Targeting Notes

Segments (New Audience, Existing Customers, Engaged Audience) are Meta internal classifications. **They cannot be targeted without custom audiences.**

- Before recommending segmentation, ask: *"Do you have a customer custom audience already built?"*
- If yes → suggest exclusions or separate campaigns using custom audience targeting
- If no → suggest building custom audiences first

### Top Performer Protection (CRITICAL)

**Never recommend pausing the top-converting ad in an ad set.** If Meta's algorithm allocates 60–80% of an ad set's budget to one ad and that ad produces the majority of conversions, that's Meta's optimization working correctly - not a problem.

**Pause decision matrix:**

| Condition | Action |
|---|---|
| Ad has highest conversion volume in ad set | **KEEP** - regardless of budget share |
| Ad has 0 conversions AND spend > 2× ad set CPA target | PAUSE |
| Ad has conversions but CPA > 3× ad set average AND ≥ 3 conversions (statistically significant) | PAUSE |
| Ad has 1–2 conversions with high CPA | WATCH - insufficient data, do not pause yet |
| Ad is < 7 days old | WATCH - let it exit learning phase |

**Self-check before any pause recommendation:** *"Am I about to recommend pausing the ad that drives the most conversions in this ad set?"* If yes → DO NOT recommend pausing. Instead recommend testing new creatives alongside it.

### Methodology Consistency

Once you establish a methodology for evaluating performance in a conversation, do NOT change it unless the user explicitly asks. Changing criteria mid-analysis destroys trust. If you realize your methodology was wrong, say so clearly, explain what you're changing and why, and apply the new methodology consistently from that point.

### STOP Rules

**Pre-check before any recommendation:** Can the user point to the exact Ads Manager UI field that this change maps to? If no → do not recommend.

- No ad-level budget moves
- No demographic or placement budget allocation
- No frequency caps for conversion objectives (Sales, Leads, App install)
- No scaling individual ads (only campaign / ad set level)
- No scaling > 20% at once
- Do not pause ads or ad sets younger than 7 days without enough signals
- Do not increase budget on entities below ROAS/CPL/CPA targets - diagnose first
- Do not call something "scalable" with < 10% revenue share of total account or statistically weak conversion volume
- Minimum signal: 3–5 conversions per ad set per week before performance judgments
- Avoid suggestions that restart the learning phase unless absolutely required
- Do not rely on the first 24–48 hours of data as reliable

### Required Diagnostics Before Action

Before any scale / pause / duplicate / structural recommendation:

1. **User segments**: New vs Existing vs Engaged distribution
2. **Delivery metrics**: CPM, CTR, Reach, Frequency (with spend context)
   - High frequency alone ≠ fatigue
   - Frequency rising + CTR falling over 14+ days = creative fatigue
3. **Engagement relevance**: offer-aligned comments vs spam / backlash
4. **Attribution windows**: 1-day click vs 7-day click vs incremental, if platform / external data mismatch
5. **Confirm results not dominated by warm / small pools** before scaling

### Scaling Guardrails

When recommending a budget increase:

- Only campaign or ad set budget fields (never ad-level)
- Max increase: **20% at once**
- Require before scaling: performance target hit, ad set revenue share ≥ 10% of account, not dominated by 80%+ warm/customer pocket
- If data is noisy or contradictory → say **insufficient data** and shift to a structured testing recommendation instead
- All scaling recommendations must include: deterioration-risk warning + review window + spend cap

### Change-Recommendation Safety Language

When recommending changes the user will apply manually, always include these warnings where relevant:

- **Budget increase > 100%**: warn this is a major increase and ask the user to confirm intent
- **Budget decrease > 50%**: warn this significantly reduces reach and performance
- **Daily vs lifetime budget**: always distinguish which one you're talking about
- **Pausing**: traffic and spending stop immediately
- **Deletion**: permanent, cannot be undone - require explicit confirmation
- **Archiving**: reversible - can be unarchived back to PAUSED
- **Bid strategy change**: resets the learning phase - always warn
- **Targeting change**: if non-trivial, may reset learning phase

---

## Metric Glossary - Canonical Definitions

Used across every skill below.

### Currency Note

Meta API budget values are typically in **cents** (integer). $50 = 5000 cents. If you see raw integer budget values from an API or export, divide by 100 to display. **Confirm account currency before showing monetary values - never assume USD.**

### Direct Metrics

| Metric | Notes |
|---|---|
| Spend | Total spend in account currency |
| CPM | Cost per 1,000 impressions |
| CTR | Click-through rate (decimal or %) |
| Impressions | |
| Clicks | |
| Reach | Unique users reached |
| Frequency | Average impressions per user |
| Video Plays | Initial impressions of the video |
| 3-Second Views | "video_view" actions - hook engagement (this is inside the `actions` array, not a separate field) |
| ThruPlays | 15-second or completed video watches |
| Purchases | Use ONE purchase action type only - see Purchase De-Duplication below |
| Leads | Lead events |
| Revenue | Conversion value for purchases |
| Purchase ROAS | Returned directly by Meta |

### Calculated Metrics

| Metric | Formula |
|---|---|
| Hook Rate | (3-second views / video plays) × 100 - **API/MCP source: use the `video_view` action inside the actions array, NOT `video_p25_watched_actions`.** **CSV source: use the "3-Second Video Plays" column directly - it's already pre-computed; the action-name distinction does not apply.** |
| Hold Rate | (ThruPlays / 3-second views) × 100 - for CSV, use "ThruPlays" ÷ "3-Second Video Plays" columns directly. |
| Cost Per Purchase | spend / purchases |
| Cost Per Lead (CPL) | spend / leads |
| Cost Per Result (CPR) | spend / specific custom-event count |
| Conversion Rate (CVR) | conversions / clicks × 100 |
| Budget Utilization | spend / (daily budget × days in period) |
| Campaign Spend Share | entity spend / parent total spend (require ≥ 25% before comparing ad sets) |
| Pareto Set | Sort ads by spend desc; keep ads where cumulative spend ≤ 90% |

### Purchase De-Duplication (CRITICAL - API/MCP only)

**This rule applies only when the source is an MCP / API that returns the raw `actions` array.** If the source is a CSV export, the "Purchases" column has already been deduplicated by Meta - **skip this rule entirely and use the column directly.**

When working with raw API data: Meta returns overlapping purchase action types in the `actions` array: `omni_purchase`, `purchase`, `offsite_conversion.fb_pixel_purchase`, `onsite_web_purchase`, `web_in_store_purchase`, `app_custom_event.fb_mobile_purchase`.

**Rule:** Use ONLY ONE. Check `omni_purchase` first; if absent, fall back to `purchase`. **Never sum multiple types** - this double-counts revenue and inflates ROAS.

**Source check before applying:**
- MCP returning an `actions` array → apply the rule.
- CSV with a "Purchases" column → skip the rule; the column is already deduplicated.
- Screenshot showing a "Purchases" number → skip the rule; trust the figure but flag the source.

### Account Type → Primary Conversion Metric (PCM)

Read the **promoted object / conversion event** from each ad set. **Never infer from campaign or ad set names** - names frequently don't match the actual config.

| Custom Event Type | Account Type | PCM |
|---|---|---|
| PURCHASE | Ecommerce | Purchase ROAS |
| LEAD | Lead generation | Cost per Lead |
| COMPLETE_REGISTRATION | Registration | Cost per Registration |
| OTHER (custom event) | Custom conversion | Cost per Result for that specific event |

If the promoted object is missing → **ask the user**. Never assume.

**Mixed accounts**: group ad sets by their custom event type. Each group gets its own PCM. Do not mix conversion types across groups.

### Special Case: COMPLETE_REGISTRATION on PURCHASE Objective (Cross-Source Safety Check)

If an ad set has campaign objective `OUTCOME_SALES` (or legacy `CONVERSIONS` / `PRODUCT_CATALOG_SALES`) AND the optimization event is `COMPLETE_REGISTRATION`, this is **almost always a 3rd-party purchase tracker** (Profitmetrics, Cosmise, Triple Whale, Northbeam) posting purchase events through Meta's CompleteRegistration endpoint - **not a misconfiguration**.

Especially common on:
- Nordic-currency accounts (DKK / SEK / NOK / EUR)
- Shopify + Profitmetrics stacks

**This is the highest-stakes safety check in the skill - getting it wrong can kill a working campaign.**

#### Source-aware detection

| Source | How to detect |
|---|---|
| MCP / API | Read `campaign.objective` + `ad_set.promoted_object.custom_event_type` directly. If the pair matches the trigger, apply the check below. |
| CSV with "Campaign Objective" + "Performance Goal" / "Optimization Goal" columns | Match by column values. If columns are missing, fall through to the manual check. |
| CSV without those columns, screenshot, or copy-paste | **Cannot detect automatically.** Use the proactive question below before any pause / scale / budget-cut recommendation on Nordic-currency or Shopify accounts. |

#### Proactive question (use whenever auto-detection isn't possible OR the account profile fits)

> *"Before I make recommendations on this account: do you use a 3rd-party purchase tracker like Profitmetrics, Triple Whale, Northbeam, or Cosmise? If yes, the conversion event reported as 'Registrations' or `COMPLETE_REGISTRATION` is real purchase volume and the campaign is correctly configured - I should treat it as purchase data."*

Ask this **once per account / conversation**, store the answer, and apply it to all subsequent recommendations.

#### Action rules (apply after detection or confirmation)

Before any optimization recommendation on such ad sets:

1. State the ambiguity in your response, naming the 3rd-party tracker explicitly.
2. **Do NOT** recommend any of the following without explicit user confirmation first:
   - "Change the optimization event to PURCHASE"
   - "Verify and switch the optimization goal"
   - Pause / archive / delete the ad set or campaign
   - Reduce budget by more than 20%

Only after confirmation may you treat the events as registrations (and recommend an optimization-event change) OR as real purchases (apply ROAS / CPA against revenue).

### Performance Benchmarks

| Metric | Good | Average | Poor |
|---|---|---|---|
| Hook Rate | ≥ 40% | 26–39% | < 25% |
| Hold Rate | > Account Avg + 25% | ≈ Account Avg | < Account Avg |
| CTR | ≥ 1.25% | 0.65 – 1.24% | < 0.65% |
| CPM variance vs Pareto avg | ≤ 20% | 20 – 60% | > 60% |
| PCM variance vs Pareto avg | Above avg | ± 20% | > 30% below |
| Budget utilization | ≥ 85% | - | < 85% |
| Min ad-set spend share for comparison | ≥ 25% of campaign | - | - |

For Hold Rate, compute the account average from a 90-day account-level pull (video fields aggregated) before applying the comparison.

### Statistical Significance Floors

| Decision | Minimum signal |
|---|---|
| Per-ad-set performance judgment | 3–5 conversions / week |
| Per-segment recommendation (placement, age, device) | ≥ 100 impressions AND ≥ 3 conversions |
| Pause ad on high CPA | ≥ 3 conversions if non-zero, OR spend ≥ 2× ad-set CPA target if zero |
| Ad set learning-phase exit | Spend ≥ 2× AOV OR 4× CPA, whichever is higher |

### Diagnostic Signals (used in skills below)

| Signal | Meaning |
|---|---|
| Frequency rising AND CTR falling over 14+ days | Creative fatigue |
| CBO with one ad set holding 80%+ of budget | Algorithm working as intended - do NOT pause the dominant ad set on small-sample noise |
| High frequency alone (no CTR drop) | Not fatigue |
| Cost cap + rising CPM | Either delivery squeeze OR audience saturation - separate before acting |

### Video Funnel

```
Video Plays → Hook Rate → 3-Second Views → Hold Rate → ThruPlays
```

---

## Depth-of-Analysis Framework - Apply on Audits & Optimization Reviews

Apply on: audits, performance reviews, optimization requests, strategy questions, funnel analysis.
Skip for: simple data fetches, campaign listings, single-metric lookups.

### 1. Hierarchical Drill-Down

Never recommend on campaign-level or ad-set-level aggregates alone. Drill one level deeper.

- Asked about a campaign → check ad set performance within it.
- Asked about an ad set → check individual ad performance within it.
- Asked about ads → check creative-level patterns (same creative across multiple ads).

Quantify the contribution: *"Ad Set X accounts for 72% of campaign spend but only 40% of conversions."*

### 2. Objective-Contextual Evaluation

Do not apply universal KPIs. Evaluate each campaign against its actual objective.

**Check first:**

- The ad-set-level optimization event (the actual `custom_event_type` - not the name)
- The campaign objective
- Bid strategy (cost cap, bid cap, lowest cost)

**Apply correct KPIs:**

| Objective | Evaluate on | NOT on |
|---|---|---|
| Sales / Purchase | ROAS, CPA, revenue | Reach, CPM alone |
| Lead gen | CPL, lead volume | ROAS |
| Traffic | CPC, CTR, landing page views | Conversions |
| Video views | Cost per ThruPlay, hook rate | ROAS, CPA |
| Awareness | CPM, reach, frequency | CPA, ROAS |

**Flag bid-strategy issues:** if a cost cap is set 3× above actual CPA, or tROAS target is 0.5× when achieving 3×, the algorithm isn't actually constrained. Call it out.

### 3. Breakdown & Segmentation Analysis

For audits, performance reviews, and optimization work, check **at least 2** of these dimensions. (This overrides the general guidance to only use breakdowns when specifically requested - deep analysis requires segmentation.)

- Placement (Feed vs Stories vs Reels vs Audience Network) - quantify CPA gaps
- Device (mobile vs desktop)
- Age / gender - flag Advantage+ expansion leakage if delivery is going outside the target
- Out-of-target segments consuming budget with poor conversion rates

**Statistical significance check:** do not make segment-level recommendations from < 100 impressions or < 3 conversions in a segment. Call out low-confidence segments explicitly.

### 4. Temporal & Trend Analysis

Never treat the date range as a single static number. Always compare time windows.

- Last 7 days vs previous 7 days
- Identify direction for each key metric: improving, declining, stable
- Flag inflection points: *"CTR started declining on [date], 3 days before CPA started rising"*

**Connect the chain:** frequency rising → CTR falling → CPA rising → this is creative fatigue, not audience exhaustion. Correlate timing across metrics to establish cause, not just observe symptoms.

### 5. Cross-Entity Dependency Mapping

Don't analyze campaigns in isolation. Map the funnel.

- Which campaigns are **prospecting** (broad / lookalike audiences)?
- Which campaigns are **retargeting** (custom audiences, website visitors)?
- Does pausing a prospecting campaign starve the retargeting audience pool?

**How to identify:**

- Prospecting: low frequency, lower CPM, lower CTR, broader audience
- Retargeting: higher frequency, higher CPM, higher CTR, smaller audience
- Check the targeting object on each ad set - presence of custom audiences = retargeting / warm

**Flag dependencies:** *"Pausing Campaign X will reduce the Website Visitors 7-day pool that feeds Campaign Y's retargeting. Expect retargeting CPA to rise within 1–2 weeks."*

### 6. Platform Mechanic Awareness

Explain how Meta's mechanics affect the specific situation - explain the compound interaction, don't just mention the mechanic.

| Signal | Mechanic | Consequence |
|---|---|---|
| Rising frequency + falling CTR | Creative fatigue | CPA will spike - need new creatives, not budget changes |
| CBO + one ad set getting 80% budget | CBO optimization | Usually working as intended. Do not recommend pausing the dominant ad set on small-sample noise. Only intervene if the dominant ad set itself is underperforming on ROAS/CPA. |
| Cost cap + rising CPM | Delivery squeeze or audience shift | Two possible causes - separate before acting: (a) algorithm can't find conversions at the cap (raise cap or loosen targeting); (b) original audience is saturated and delivery is leaking to worse segments. Check frequency, reach trend, and audience size to decide which. |
| Advantage+ audience expansion ON | Targeting leakage | Check breakdowns for out-of-target delivery eating budget |
| Insufficient spend on new ad set | Learning phase | No optimization decisions until the ad set has spent at least **2× AOV or 4× CPA, whichever is higher**. The 50-conversion rule only applies to brand-new accounts with no prior history. |

### 7. Proactive Signal Detection

After answering the user's question, surface 2–3 additional concerns or opportunities they didn't ask about.

Scan for:

- Any entity with frequency > 3 and rising - creative fatigue incoming
- Any ad set spending > 30% of campaign budget with 0 conversions - budget waste
- Any campaign where the top ad has declining CTR over 7+ days - degradation ahead
- Any retargeting audience pool shrinking (declining reach WoW)
- Any metric that looks fine now but the trend predicts problems in 7–14 days

Format: *"Outside your question, I noticed [signal] in [entity] - this suggests [predicted consequence] within [timeframe]."*

### 8. Attribution Rigor

Don't take reported conversion numbers at face value.

Always check:

- What attribution window is active? (1-day click, 7-day click, 1-day view?)
- What share of conversions are view-through vs click-through?
- Is the reported ROAS inflated by view-through on high-impression campaigns?

Flag these situations:

- Retargeting campaign with very high ROAS + mostly view-through conversions → likely over-attributed
- Two campaigns targeting the same audience → conversions may be double-counted
- Advantage+ Shopping targeting existing customers → attributed ROAS includes conversions that would have happened organically

Frame as caveat: *"Before scaling Campaign X based on its reported 4× ROAS, note that [specific attribution concern] means the true incremental ROAS is likely lower."*

---

# Skill 1: Initial Health Check

Validate account setup before any optimization work. **Run this first on any audit.** If checks fail, fixing setup is the top priority - optimization on a broken setup wastes effort.

## When to use this skill

- First step of any account audit
- Before recommending any structural change
- When CPA or ROAS suddenly degrades and tracking issues are plausible

## Step 1: Event Match Quality (EMQ) Score Check

EMQ score is **not directly available via standard insights APIs.** The user must check it manually in Events Manager.

**What to do:**

1. First determine the primary conversion event (from the ad set's optimization goal - covered in Skill 2).
2. Ask the user to manually check: **Events Manager → Data Sources → [Pixel] → Event Match Quality** for that event.
3. Apply the threshold:

| EMQ Score | Status | Action |
|---|---|---|
| ≥ 8.5 | Good | Proceed with analysis |
| < 8.5 | ALERT | Stop and notify the user before optimizing |

**Alert wording when EMQ < 8.5:**

> ⚠️ **CRITICAL:** Your EMQ Score for [Purchase / Lead] is below 8.5.
>
> This means your Pixel is not set up correctly. Likely consequences:
> - Lower conversion rate than possible
> - Higher cost per result than necessary
>
> **Recommendation:** Reach out to a technical expert and fix this as TOP PRIORITY before any other optimization work.

## Step 2: Audience Segments Configuration Check

Audience segment breakdown (New / Engaged / Existing) is **available via insights breakdowns**, but the underlying engaged/existing audience setup is managed in Audience Manager and isn't directly readable via standard APIs.

**What to do:**

1. Pull account-level insights for the last 30 days, broken down by **user segment** (the breakdown that splits New / Engaged / Existing).
2. If the breakdown returns empty or is missing segments, alert the user:

> ⚠️ **Setup needed:** Audience Segments are not configured.
>
> Why this matters: you can see spend breakdown between existing customers and new audience, at campaign / ad set / ad level. Without this, you can't tell where money is actually going.
>
> Note: this does NOT affect targeting - only improves reporting.
>
> Setup guide: Facebook Business Help → Audience Segments configuration.

## Step 3: Decide whether to proceed

| Result | Action |
|---|---|
| EMQ ≥ 8.5 AND segments configured | Proceed to Conversion Metric Focus |
| EMQ < 8.5 | Recommend fixing pixel before any optimization |
| Segments not configured | Flag for setup; can proceed with limited reporting |

## Minimum required data (by source)

| Source | What you need |
|---|---|
| MCP | Account-level insights for last 30 days with `breakdowns=['user_segment_key']` to confirm segments are configured. EMQ requires manual user check. |
| CSV | A breakdown export by user segment (if available). EMQ requires manual user check. |
| Screenshot / copy-paste | Not recommended for this skill - too many missing pieces. Ask for at minimum a screenshot of Events Manager → EMQ for the primary event. |

## Tool hints

EMQ Score is **never** in standard insights data - it always requires a manual check in Events Manager → Data Sources → [Pixel]. Audience Segment configuration is observable from any source via the user-segment breakdown; if the breakdown returns empty, segments are unset.

---

# Skill 2: Conversion Metric Focus

Determine account type and set the **Primary Conversion Metric (PCM)** that all subsequent analysis will use. **Run this second on any audit, immediately after the Health Check.**

## When to use this skill

- Once per account, at the start of analysis
- Whenever account-level KPIs are ambiguous
- When new campaigns are added that may use a different conversion event

## Data to gather

For each active campaign in the account:

- Campaign objective (e.g., OUTCOME_SALES, OUTCOME_LEADS, OUTCOME_TRAFFIC)
- The optimization event at the ad set level - the actual `custom_event_type` of the promoted object. **Never infer from campaign or ad set names.**
- The list of conversion actions used

## Decision Logic

For each live campaign:

```
IF majority conversion event = PURCHASE         → ECOMMERCE
IF majority conversion event = LEAD             → LEAD GENERATION
IF custom conversion event (OTHER)              → CUSTOM CONVERSION (treat standalone; do not group with other types)
```

**Mixed accounts:** group ad sets by their event type and apply each group's PCM separately. Do not mix conversion types into one aggregate.

## Set PCM by Account Type

### Ecommerce

| Metric | Type | Source |
|---|---|---|
| Purchase ROAS | Primary (PCM) | `purchase_roas` field, deduplicated per the Purchase De-Duplication rule |
| Cost per Purchase | Secondary | spend / purchase count |
| Conversion Rate | Secondary | purchases / clicks |
| Revenue | Secondary | conversion value for purchase events |

### Lead Generation

| Metric | Type | Source |
|---|---|---|
| Cost per Lead | Primary (PCM) | spend / lead count |
| Conversion Rate | Secondary | leads / clicks |

### Custom Conversion

| Metric | Type | Source |
|---|---|---|
| Cost per Result | Primary (PCM) | spend / count of the specific custom event |
| Conversion Rate | Secondary | custom events / clicks |

> **Note:** For custom conversion campaigns, metrics are scoped to that specific conversion event only. No other conversion events should be included in the analysis for that campaign.

## Output

Set internally:

- `account_type`: `ecommerce`, `lead_gen`, or `custom_conversion`
- `primary_metric`: `purchase_roas`, `cost_per_lead`, or `cost_per_result_<custom_event>`

These flow into every later skill.

## Minimum required data (by source)

| Source | What you need |
|---|---|
| MCP | Per active campaign: campaign objective + per-ad-set `promoted_object.custom_event_type`. Directly available. |
| CSV | "Campaign Objective" + "Performance Goal" / "Optimization Event" columns (often not in default exports - must be enabled). If unavailable, **ask the user to confirm the conversion event for each campaign** before proceeding. |
| Screenshot / copy-paste | Ask user to confirm directly per campaign. |

## Tool hints

The most reliable signal is `promoted_object.custom_event_type` on the ad set object. Campaign / ad set *names* lie often - never infer the conversion event from a name. If using MCP, fetch the ad set object explicitly. If using CSV without the optimization-event column, this entire skill blocks until the user confirms.

---

# Skill 3: Campaign Structure Analysis

Identify the strategic pattern behind how the account splits spend across campaigns. Different structures need different evaluation rules.

## When to use this skill

- Account audits
- When the user asks "is my campaign structure right?"
- Before doing ad-set-level analysis (the structure type changes what comparisons are valid)

## Data to gather

For each active (non-retargeting) campaign:

- Objective, bid strategy, daily/lifetime budget, status
- Ad-set level: bid strategy, optimization goal, promoted object (the conversion event), budget, attribution window, full targeting object
- Spend distribution at the campaign level over the last 30 days, sorted by spend desc

**Targeting fields to inspect on each ad set:**

| Field | Purpose |
|---|---|
| Age min / max | Demographics |
| Genders | Demographics |
| Geo locations | Geo targeting |
| Custom audiences | Retargeting detection |
| Detailed / flexible targeting (interests, behaviors) | Interest targeting |
| Publisher platforms / placement positions | Placement strategy |
| Languages | Language targeting |

## Step 1: Identify Retargeting Campaigns

Check the targeting custom audiences for: website visitors, past purchasers, video viewers, page engagers.

```
IF targeting.custom_audiences contains retargeting audiences
→ LABEL as "Retargeting Campaign"
→ EXCLUDE from structure analysis
```

## Step 2: Match the Active Structure to One of These Patterns

### Pattern A: Primary vs Testing

**Criteria:**

```
campaign_budget_share = campaign_budget / total_non_retargeting_budget

IF Campaign A:
   - budget_share < 30%
   - ad_count ≥ 5
AND Campaign B:
   - budget_share ≥ 40%
   - ad_count < Campaign A's ad_count
   - 70%+ of ads also exist in Campaign A
→ CONFIRMED: Primary vs Testing structure
```

### Pattern B: Bid Limit Split

**Criteria:**

For campaigns/ad sets (excluding retargeting), compare the targeting objects and bid configurations.

```
IF targeting is IDENTICAL across splits
AND bid_strategy OR bid_amount DIFFERS
→ CONFIRMED: Bid Limit Split
```

### Pattern C: Audience Split

**Criteria:**

For campaigns/ad sets (excluding retargeting), compare interest spec, custom audiences, demographics.

```
IF audiences DIFFER between splits
→ CONFIRMED: Audience Split

NOTE: Different bids/budgets across audience splits is OK.
```

### Pattern D: Too Many Campaigns (Custom Pattern)

**When:** more than one campaign and the structure doesn't fit A, B, or C.

**Investigate using creative inspection** - review the actual creative content of ads in each campaign and look for:

- Video vs Image split
- Catalog vs Static split
- UGC page vs Brand page
- Different products
- Different personas / angles
- Part

…(truncated)
