How Facebook's Ad Review System Actually Works in 2026 (Insider Breakdown)

1. The Three-Tier Review Architecture

Meta's ad review system is publicly described as "automated review with human oversight." Internally — and operationally observable from the outside — it is a three-tier funnel that downsamples ads through progressively more expensive checks.

• Tier 1 — Automated Crawler. Fast, stateless. Fetches the landing page, scans creative assets, parses HTML for prohibited markers. Runs on every single ad submitted, usually within seconds of campaign creation.
• Tier 2 — ML Content Classifier. Vision and language models score creative + landing page text against policy categories. Triggers selectively: most ads see it, but ad sets with low-risk creative (think enterprise SaaS, B2B leadgen) skip parts of it.
• Tier 3 — Human Reviewer. Outsourced moderation analysts (predominantly via Accenture, Cognizant, and Concentrix in 2026) handle the long tail of escalations: novel creative, complaint-driven re-review, high-spend campaigns, or anything Tier 2 scored as ambiguous.

Roughly 96% of ads never reach Tier 3. The other 4% are where most disapprovals — and almost all permanent bans — originate.

2. Tier 1 — The Automated Crawler

The crawler is the cloaker community's favourite target because it is the easiest tier to recognise. It fetches your landing page from servers inside Meta's autonomous system. The User-Agent strings observed in production traffic in Q1 2026 include:

facebookexternalhit/1.1
facebookexternalhit/1.1 (+http://www.facebook.com/externalhit_uatext.php)
meta-externalagent/1.1
meta-externalfetcher/1.1
facebookcatalog/1.0
FBAV/[version] (Facebook iOS App probing)
Mozilla/5.0 (compatible; FacebookBot/1.0; +http://developers.facebook.com)

These crawlers are generally honest about who they are — useful for legitimate Open Graph fetches, predictable for cloakers. They almost always originate from AS32934 (Facebook, Inc.), with the bulk of requests from these IPv4 ranges in our 2026 telemetry: 31.13.24.0/21, 66.220.144.0/20, 69.63.176.0/20, 69.171.224.0/19, 74.119.76.0/22, 102.132.96.0/20, 157.240.0.0/17, 173.252.64.0/18, 179.60.192.0/22, 185.60.216.0/22, 204.15.20.0/22.

The crawler does not execute JavaScript by default for plain link checks. It does execute JS for Open Graph image rendering and for landing-page integrity checks introduced post-2023. This bifurcation is important: a static HTML check is much cheaper than a full Chromium render, and Meta does not waste compute on every request.

3. Tier 2 — The ML Content Classifier

If Tier 1 returns the white page successfully, the workload passes to a vision-language model trained on Meta's policy taxonomy. This model — internally referred to in leaked documents as Adverse Risk Embedding (ARE) — operates on a multi-modal feature set: ad creative imagery, ad copy text, landing-page screenshot, landing-page DOM-text, and historical signals about the page's domain.

ARE produces a probability vector across roughly 40 policy categories: gambling, prescription drugs, weapons, financial scams, adult content, low-quality user experience, prohibited health claims, and so on. Categories above a confidence threshold (we estimate ~0.78 for hard-block decisions, lower for soft escalation) trigger one of three actions:

• Auto-disapprove — the ad never runs.
• Approve with monitoring — the ad runs but with elevated weight on user-report signals.
• Escalate to Tier 3 — the ad pauses pending human review (typically 4–48 hours).

The classifier explicitly rendered the landing page as a real Chromium browser starting in late 2024. This is the change that broke a generation of cloakers that filtered only on User-Agent. The renderer in 2026 still has tells — see our browser fingerprinting article — but it is not a static HTTP fetch any more.

4. Tier 3 — The Human Reviewer

When an ad lands in front of a human, the entire game changes. Reviewers operate from a custom moderation tool that overlays the live ad creative, the landing page rendered through a real browser, the campaign metadata, the BM history and historic risk score for the linked Page. They are paid per decision and have an average handle time of 26–60 seconds per ad.

Critically, in 2026 reviewers can — and increasingly do — request a render through a residential ISP via Meta's internal proxy mesh. This is the single biggest change in the past 18 months. The classic strategy of cloaking only AS32934 is dead. A request that arrives from Comcast, Spectrum, or any major consumer ISP can be a moderation analyst on a managed laptop, indistinguishable at the network layer from a real prospect.

This is why behavioural fingerprinting matters more in 2026 than IP filtering. The signal that leaks from a moderator-driven session is behavioural: deliberate cursor patterns, rapid sequential page-element inspection, headless-style timing curves on form-fill, and an absence of organic scroll behaviour. Detection at this tier is statistical, not rule-based.

5. AS32934 and the Network Footprint

AS32934 is Meta's primary autonomous system. It is documented in public BGP route data and corresponds to roughly 25 million IP addresses. Anything originating from this ASN should be considered Meta infrastructure: ad crawlers, Open Graph fetchers, link previews from Messenger, internal moderation tooling proxies.

Meta also publishes its own crawler IP list endpoint at developers.facebook.com/docs/sharing/webmasters/web-crawlers, which is updated several times a week. Production cloakers ingest this feed automatically.

Adjacent ASNs worth blocking by default:

• AS32934 — Facebook (primary)
• AS54115 — Facebook (overflow capacity)
• AS63293 — Facebook (edge-network ranges in APAC)
• AS149642 — Facebook (used for some iOS app probes)

This is necessary but no longer sufficient. The residential-proxy moderator workflow described in Section 4 sidesteps any ASN-level filter.

6. Ad-Level vs BM-Level Enforcement

Meta's enforcement system operates at four different scopes. Knowing which one fired is essential to understanding what to do next.

The catastrophic failure mode is BM-level enforcement. When Meta determines that a BM is operating in bad faith, it cascades sideways: every personal admin account is flagged, every linked employee is flagged, payment methods are blacklisted, and in extreme cases the same business entity is blocked from creating new BMs going forward. This is why experienced operators compartmentalise aggressively — see our Facebook cloaking matrix for vertical-specific BM segmentation patterns.

7. The Reality of Appeals

Meta's published appeal flow is straightforward: rejected ads include a "Request Review" link, BM bans surface a "Disagree with decision" form, and high-spend accounts have a Business Help Center contact channel. The reality is less rosy.

Based on appeal-resolution data we have collected from advertisers in the IPCloak.ai user base over Q1 2026:

• Ad-level disapproval appeal — ~58% reversal rate within 24 hours when the original decision was automated and the creative is genuinely policy-compliant.
• Ad-account suspension appeal — ~12% reversal rate, with most successful appeals occurring within 7 days. After 14 days the chance drops below 3%.
• BM-level ban appeal — under 2% reversal rate. Even with a paid Meta agency rep, success is essentially zero once a BM is flagged for cloaking specifically.

The system is not designed for restoration. It is designed for throughput. Operationally, the only sane stance is to treat enforcement as a one-way door and allocate budget to replacement capacity rather than negotiation.

8. What Triggers a Manual Review

Tier 3 is expensive, so Meta is selective. The triggers we have empirically reverse-engineered from incident telemetry include:

• Spend velocity. New ad accounts that ramp from $20/day to $2,000/day in under 72 hours almost always get sampled.
• Pixel anomalies. Conversion events at unusual ratios — for example, 30% conversion on a "View Content" pixel — flag a quality concern.
• User reports. Three or more "irrelevant," "scam," or "misleading" reports in a 6-hour window force a human re-review.
• Domain reputation. New domains under 30 days old are universally sampled. Domains shared with previously banned BMs are auto-escalated.
• Creative similarity. ARE flags creatives matching banned ad embeddings via perceptual hash. The same image with a 1px crop will still trigger.
• Vertical heuristics. Crypto, sweepstakes, weight loss, ED, and "make money online" categories are sampled at 10x the baseline rate regardless of creative quality.
• BM profile completeness. Empty business profiles, no two-factor on the admin user, and missing payment verification all elevate the sampling rate.

You cannot avoid all of these signals, but you can dampen them. Slow your spend ramp, complete your BM profile, run your white page on a domain at least 60 days old, and stay below the user-report threshold by writing creative that matches the offer's actual claim level.