Google Deploys Gemini To Block Billions of Ads

Google deployed its `Gemini` family more broadly across ads enforcement and related surfaces, using the models to block or remove 8.3 billion ads worldwide in 2025 and suspend 24.9 million advertiser accounts. Google says `Gemini` stopped over 99% of policy-violating ads before they ever served. Of those actions, 602 million ads and 4 million accounts were linked to scams. Google also reported reducing incorrect advertiser suspensions by 80% and processing user reports at 4x the previous rate.

Google integrates cloud-hosted large models with on-device inference to push detection earlier in the ad pipeline. Key technical elements include:

• •`Gemini` server-side classifiers that analyze hundreds of billions of signals, including account age, campaign patterns, creative text, and landing-page behavior.
• •On-device `Gemini Nano` models embedded in Chrome and Android for early detection of scammy content and phishing attempts before navigation completes.
• •Cross-product telemetry linking Search, Chrome, Maps, Ads, Messages, and Play to surface coordinated abuse across channels.
• •Automated enforcement logic that prioritizes instant ad blocking and content restriction over account-wide suspensions, lowering false positives while acting at scale.

Generative AI has lowered the cost of producing high-volume, believable scam creatives, forcing a response that is itself AI-first. Google's move illustrates two trends practitioners should note: one, defenses must operate at the same scale and latency as generative attacks; two, detection shifts from static rules to pattern recognition across multi-channel signals. The measured drop in incorrect suspensions suggests better precision but also means more nuanced adjudication is happening inside model pipelines rather than through manual review. The integration of on-device inference is notable because it reduces time-to-block and preserves a degree of privacy by keeping some signals local.

Model-driven enforcement introduces adversarial and policy risks. Attackers can adapt by varying prompts, obfuscating landing pages, or gaming behavioral signals. Relying on cross-product signals raises data-sharing and privacy trade-offs. The reduced number of account suspensions alongside massive ad blocking could hide persistent, low-cost abuse rings that reconstitute accounts faster than they can be disabled.

Advertisers need stronger policy-compliant automation, improved provenance for creatives, and signal hygiene to avoid false flags. Security teams should plan for:

• •Investing in data telemetry that mirrors Google's multi-signal approach.
• •Monitoring for distribution-level anomalies, not just single-creative signatures.
• •Preparing for adversarial testing against Gemini-style detectors.

Whether adversaries pivot to more evasive, low-volume tactics or to channels outside Google, how Google balances transparency with enforcement, and how third-party measurement and appeal pathways evolve to handle AI-driven adjudication.