Ad Review: What the ML Classifier Looks at First Explained

Quick Answer

Meta's ad review pipeline runs new ads through an ordered automated classifier stack in roughly 60-180 seconds for most ads. The stack checks brand safety first, then prohibited content categories, then restricted categories, then claims/promises classifiers, then landing page checks. Failing any stage halts review and either rejects or escalates to human review.

The Mechanism Explained

The review pipeline runs in this order (faster classifiers first, expensive ones last):

1. Hash check — has this exact creative been reviewed before? If yes, copy the old verdict.
2. OCR pass — extract text from the image/video for downstream classifiers.
3. Brand safety classifier — explicit content, hate speech, violence, drugs, weapons. Fast and aggressive.
4. Prohibited categories — financial scams, fake degrees, miracle cures, MLM pyramid signals. Each category has its own model.
5. Restricted categories — alcohol, gambling, supplements, dating, political ads. These need extra fields filled in correctly.
6. Claims classifier — exaggerated promises, before/after manipulation, miracle weight loss claims, income guarantees.
7. Personal attribute classifier — does the ad imply personal characteristics about the viewer (race, religion, sexuality, health condition)? This is one of the most common rejections.
8. Landing page scan — does the destination URL pass the same classifier checks?
9. Creative quality threshold — too low resolution, broken aspect ratio, missing primary content.
10. Final policy aggregator — combine all signal scores into Approved / Rejected / Manual Review.

The whole stack runs in under 3 minutes for ~95% of ads. Failures at any stage either get auto-rejected (with a reason code) or escalated to human review (typically resolved within 24 hours).

The most common rejection causes for advanced advertisers:

• Personal attribute language — using "you" in a way that implies a characteristic ("Are you struggling with debt?")
• Before/after imagery — particularly fitness and beauty
• Implied claims — even if the actual claim isn't made
• Low-resolution creative — falls below the quality threshold automatically

Practical Implication

Pre-clean your ads against the personal attribute and claims classifiers because they're the most common rejection causes. Use neutral language instead of "you" framing. Use comparison imagery instead of literal before/after. And always upload high-resolution creative — sub-resolution is an instant fail.

Real Numbers

• Ad review completion: 60-180 seconds for ~95% of ads
• Auto-rejection rate for new advertisers: ~12-15% of submitted ads
• Human review escalation rate: ~4-7% of submitted ads
• Personal attribute rejections account for ~25% of all rejections in lead gen verticals

FAQs

Q: Does Meta human-review every ad?
No — only ~5-7% of ads escalate to human.

Q: How do I avoid the personal attribute rejection?
Use neutral language; avoid "you" framing implying characteristics.

Q: Does the review classifier learn from appeals?
Yes — successful appeals retrain the classifier on edge cases.

Q: Is review faster for established accounts?
Yes — high-trust accounts skip some early stages.

Q: Can I see which classifier rejected my ad?
Only the high-level reason code, not the specific classifier.

Pix-Vu

Low-resolution and miscroped images are auto-rejected. Pix-Vu produces high-resolution product images in the correct aspect ratios so you never lose ads to the resolution classifier — at https://pix-vu.com.