Google Gemini Blocked 8.3 Billion Harmful Ads in 2025 and Is Expanding in 2026

Google Releases Its 2025 Ad Safety Report — The Gemini Numbers Are Striking

On April 16–17, 2026, Google published its annual ad safety report alongside a detailed update on how it is expanding its use of Gemini AI models for threat detection across the advertising ecosystem. The top-line figure: 8.3 billion harmful ads blocked and 24.9 million advertiser accounts suspended in 2025. Among those, 602 million ads were tied specifically to scams.

These are detection numbers at a scale that no human review process could match. The interesting part is not just the scale — it is how Gemini's analytical approach differs from what came before, and why that difference matters for defenders everywhere.

From Keywords to Behavioral Analysis

Traditional ad safety systems operated on keyword matching and pattern recognition: flag certain phrases, URL structures, or advertiser behaviors that correlate with policy violations. This approach worked until it did not — adversarial actors have spent years learning to phrase around keyword filters and rotate accounts fast enough to stay ahead of static rules.

Gemini changes the signal set entirely. Instead of asking "does this ad contain suspicious keywords?", the system now evaluates a broader set of signals simultaneously:

- **Advertiser history**: Full account behavioral history across time, not just the current campaign

- **Campaign structure**: Pattern matching against known fraud campaign architectures

- **Network relationships**: Connections between this account and other accounts in the broader ecosystem

- **Intent signals**: Landing page content, ad copy, and bidding behavior analyzed together for coherent intent

Multi-signal behavioral analysis is substantially harder to game than keyword filtering. Changing a phrase is trivial. Constructing a plausible account history, normal campaign structure, and legitimate-looking bidding behavior simultaneously is costly and technically demanding for adversarial actors.

The Real-Time Catch Rate

Google reports that the majority of Responsive Search Ads created in Google Ads are now reviewed instantly, with harmful content blocked at submission — before the ad ever reaches a user. The catch rate for policy-violating ads before serving exceeds 99%.

Why This Matters Beyond Google's Platform

The ad fraud problem is not unique to Google. Any platform running user-generated advertising at scale faces the same adversarial pressure. What Google is demonstrating with Gemini's deployment is a model for how large language models can serve as force multipliers for content safety systems — not just on ads, but on any high-volume content moderation challenge.

The arms race dynamic here is real. Google explicitly notes that bad actors are using generative AI to create deceptive ads at scale. The defensive deployment of Gemini is a direct response to the offensive use of comparable technology. As generative AI tools become more accessible, every content platform will face this challenge.

What Expands in 2026

Google plans to extend Gemini's real-time review capability to additional ad formats beyond Responsive Search Ads throughout 2026. Each format that moves from delayed review to instant at-submission blocking narrows the window for harmful ads to appear even momentarily — a meaningful improvement for the hundreds of millions of users who encounter Google ads daily.

The defensive use of AI to protect users at advertising scale is one of the clearest illustrations of where AI delivers unambiguous value: doing something necessary, at a volume no human team could handle.

Sources: BleepingComputer (April 16, 2026), Help Net Security (April 17, 2026), TechRadar (April 2026), Cyberpress (April 2026)