How AI recommendations work

The full flow

1. Data pull: scheduled daily (also manually triggerable) from your connected Google / Meta / TikTok accounts — 7-day and 30-day metric snapshots
2. Anomaly detection: against historical baselines, flag CPA > target by 30%, ROAS < 1.5, CTR < 50% of industry average, etc.
3. LLM analysis: anomalies + account metadata + prior optimization history → Claude (Sonnet 4.6) or OpenAI GPT-4o
4. Structured output: the model writes 3–8 specific recommendations via our JSON schema — each with type / target / reason / expected impact / confidence
5. Human approval → execution: in default approval mode, recommendations enter the pending queue; only after you click "Apply" do we call the platform APIs to mutate your account

Recommendation types

• pause_keyword — pause high-CPA keywords
• pause_campaign — pause an entire losing campaign
• increase_budget — push more budget to high-ROAS ad groups
• decrease_budget — cut budget on inefficient ads
• add_keyword — add keywords surfaced from search-term reports
• change_bid — adjust bid strategy or target CPA / ROAS
• creative_refresh — creative fatigue → swap assets

Three confidence tiers

• high: clear data outliers (e.g. CPA $45 vs target $8) — safe to apply
• medium: directionally right but uncertain (e.g. seasonal effects) — small-step recommended
• low: advisory only — human review recommended