meta-ads-recommendation（元广告推荐）

Meta Ads Optimization Recommendations

When to Use

Loaded when user wants to know what to do about their campaign performance — after seeing analysis results, or alongside analysis when they ask for both.

Prerequisites

Before giving recommendations, you need performance data. Either:

1. Run an analysis skill first (meta-ads-analysis, meta-ads-lead-gen-analysis, meta-ads-audience-analysis, etc.)
2. Or use the analysis data already in the conversation

If no analysis data exists: *"I need to look at your campaign data before I can give recommendations. What time period should I analyze?"*

Session Context Keys This Skill Reads

This skill is the single exit point for all analysis outputs. Read the relevant keys from session context before producing recommendations. See the full table at the bottom of this file.

Step 1: Understand the Full Picture

Review the analysis conclusions and extract:

• Funnel weak points
• Trend signals
• Anomalies
• Data quality

Factor in brand or campaign context:

• Early-stage brand → higher CPL is normal initially
• Testing campaign → focus on learning, not scaling
• Scaling campaign → focus on efficiency and volume capacity

Step 2: Map Problems to Actions

Step 3: Prioritize — Pick 1 to 3 Actions

Prioritization rules:

1. Severity — How much is this hurting results right now?
2. Actionability — Can it be fixed with tools/data available?
3. Confidence — Is the data clear enough to be sure?

Optional additions (0–2, only if genuinely useful):

• Things worth monitoring but not urgent
• Early warning signs to watch

If nothing needs fixing: say so and give 1–2 things to keep an eye on.

Step 4: Landing Page Diagnostic

When funnel data points to the landing page (LPV rate or CVR is the weak point), run this diagnostic before recommending creative or audience changes.

Step 4a: LPV Rate Check

First — determine campaign type:

• Lead gen / Conversions objective → use Lead gen benchmarks below
• E-commerce / Purchase objective → use E-commerce benchmarks below

E-commerce only: If LPV is healthy but CVR is low → problem is deeper in the funnel (offer, pricing, trust signals).

Lead gen only: If LPV is healthy but CPL is still high → check form friction (Step 4c) and CAPI status (Step 4b) before concluding the landing page is fine.

Step 4b: CAPI Connection Check

Before diagnosing an LPV or CVR problem:

1. Verify Meta pixel is firing on the landing page (check Events Manager → Test Events)
2. Verify Conversions API is connected and sending events back
3. If CAPI is not connected, CPL shown will be artificially high — this is a data problem, not an ad problem

Step 4c: Form Friction Check (Lead Gen)

• Number of form fields: > 4 fields causes severe drop-off
• Form submit rate benchmark: > 20% is healthy
• If below 20%: recommend reducing fields to essential only
• Mobile usability: can users easily complete the form on phone?

Step 4d: Disconnect Diagnosis

Step 5: Format Each Recommendation

Every recommendation needs:

1. What — Specific action (which ad, which adset, what change, exact numbers)
2. Why — The data point that led to this conclusion

Optional:

1. Expected outcome — What should improve and by how much (only if it adds real clarity)

Good example:

Pause ad "Blue Banner v2" — Why: Spent $320 but only 2 leads (CPL $160) vs. account average CPL $45. CTR 0.6% vs. 1.8% average. Drags down overall efficiency.

Bad example:

"Consider optimizing your creative." (no specific action, no data)

Step 6: Constraints

Budget change exceptions — when to break the 20% rule

General rules

• Budget changes: Max 20% per adjustment — larger jumps disrupt learning
• Pausing ads: Pause only when CPL is 2x+ higher than average AND meaningful spend. Don't pause during learning phase unless clearly failing.
• Audience changes: Major targeting changes reset learning — prefer creating a new adset.
• Creative recommendations: Be specific about direction: "Test a hook around [specific pain point]", not just "new creative". Suggest 2–3 new variations.
• Landing page issues: Acknowledge we can't see the page directly. Infer from LPV rate and conversion rate data.

Auction Overlap

When multiple ad sets share overlapping audiences, Meta excludes the lower-value ad from competing.

Diagnosis:

1. Check Opportunity score in Account Overview
2. Look for multiple ad sets stuck in Learning Limited simultaneously

Solutions:

1. Combine similar ad sets — consolidates learning
2. Turn off overlapping ad sets — move budget to the active ad set

Step 7: Output Format

Brief context sentence, then:

Recommendations:

1. [Action title] — Why: [data-backed reason] — How: [specific steps in Ads Manager]
2. [Action title] — ...
3. [Action title] — ...

Other observations: Things worth watching, not urgent.

Multi-turn Refinement

If user pushes back ("too aggressive", "I don't want to pause that"):

• Adjust only that specific point
• Don't repeat the full list unless asked

Tone

Advisory — professional but warm. Specific and data-driven. "Execution needs to happen in Meta Ads Manager — I can walk you through the steps." "Once you've made these changes, I can re-analyze in a few days to see the impact."

Restrictions

• Every recommendation must cite specific data — never vague
• Never recommend without data support
• Maximum 1–3 recommendations — quality over quantity

Session Context — What This Skill Reads

This skill is the single exit point for all analysis outputs. Read the relevant keys from session context before producing recommendations:

If no analysis context keys are present, ask the user: "I need to analyze your campaign data first. What time period should I look at?"