Somewhere between the fifth suspicious loan ad and that oddly persuasive “guaranteed returns” banner, the modern internet user has quietly become a full-time skeptic. And yet, the ads keep getting smarter, almost as if they’re learning us as quickly as we’re learning to distrust them. It’s in this uneasy middle ground that Google’s latest Ads Safety Report lands, not as a quiet data dump, but as a rather loud assertion: the system is watching, learning, and, more importantly, intervening before things go wrong.
In a interaction during a Google roundtable, Keerat Sharma, Vice President, Ads Privacy and Safety at Google, unpacked how 2025 has already reshaped the very infrastructure of ad safety, less as a reactive clean-up job and more as a predictive, AI-first defence system.
“We have been able to stop 99% of policy-violating ads before they were served,” he said, calling it the company’s “north star.” He added, “This is effectively about stopping threats before they start,” underlining a shift from moderation to preemption.
The scale of that ambition is reflected in the numbers. Google has reported blocking or removing over 8.3 billion ads globally in 2025, while suspending nearly 24.9 million advertiser accounts and restricting 4.8 billion ads. In India alone, 483.7 million ads have been removed and 1.7 million advertiser accounts suspended, figures that Sharma suggested are less about volume spikes and more about sharper detection.
“What you’re really seeing is the evolution of us being able to deploy things with much greater precision at the creative level,” he said. “Over the course of time, as we implement new types of defences, those numbers will fluctuate.”
At the core of this transformation is Gemini, which Sharma described not just as a tool, but as “an infrastructural construct in how we are operating.” He said, “Instead of looking at discrete parts of an ad or the advertiser account, we combine all of that into one larger piece of analysis that allows us to really get to the intent.”
That idea of “intent” becomes particularly crucial in an era where AI-powered scams have surged dramatically. Responding to a question on how Gemini differentiates between legitimate financial offers and socially engineered scams, Sharma said, “There’s no one silver bullet or one key thing here. There are lots of different signals that we use, attributes about the account, the creative, the landing page, when the domain was registered.”
He added, “An advertiser that has been with us for a long period of time…is extraordinarily different from an advertiser who has just showed up with a brand new domain and a bunch of content that may look dubious.” Sharma noted that such signals are evaluated collectively to determine risk, after which “we either filter what we want to do or move that advertiser through another series of defences.”
Speed, in this system, is as critical as intelligence. Sharma highlighted that one of the biggest shifts in 2025 has been real-time enforcement. “Something that would run asynchronously to now something that happens in real time,” he said. “For responsive search ads, we scan in milliseconds and provide policy feedback instantaneously.”
This acceleration has had tangible outcomes. “This also resulted in roughly 4 million accounts that were suspended for scam-type activity, the majority of which happened before any ads were run,” Sharma added.
But with aggressive enforcement comes an equally pressing concern: overcorrection. Particularly in sensitive sectors like finance, healthcare, or crypto, where legitimate advertisers often walk a fine compliance line. Addressing this, Sharma said, “We take both sides of this very seriously, protecting users and providing a great experience for well-intentioned advertisers.”
He added, “That 80% reduction in incorrect suspensions is good progress, and we will continue to focus on improving that experience.” The report corroborates this, noting a significant drop in false positives alongside improved detection accuracy.
Zooming out, Sharma positioned Google’s approach as layered rather than singular. “It’s important to have a defensive strategy that is layered and not reliant on one silver bullet,” he said. He pointed to advertiser verification as a key pillar, adding, “Over 90% of the ads seen on Google are from verified advertisers.”
On whether rising enforcement has come at the cost of revenue, Sharma was unequivocal. “We’ve never sought to make money off of any policy-violated ads. There’s no revenue equation. We just don’t want that in our system at all,” he said.
The report also underscores the breadth of enforcement beyond ads themselves, with action taken on over 480 million web pages and 245,000 publisher sites. AI systems have driven over 467 million of these actions, contributing to detection rates exceeding 97% in sensitive categories like sexual content.
When asked where Google still falls short, Sharma didn’t deflect. “That 1% is important for us to keep making progress on,” he said, referring back to the 99% pre-emptive blocking metric. “We know there’s still room to grow.”
And perhaps that’s the most telling takeaway from this interaction. Despite billions of ads blocked, millions of accounts suspended, and near-perfect detection rates, the system still measures itself against what slips through, not what it catches.
Because in a landscape where scams are evolving as fast as the safeguards, perfection isn’t the benchmark. Progress is.
