Google Is Building an AI Agent. The Real Victim May Be Google.

Google is building an AI agent. The company most dependent on human attention may have just invented the tool that makes human attention worth less.

Internal documents describe Remy — Google's codename for a 24/7 personal agent powered by its Gemini model — as one that "takes actions on your behalf," according to Business Insider. That language is word for word the pitch OpenClaw made when it went viral. The difference is that OpenClaw shipped and proved demand existed: it reached 247,000 GitHub stars and had 135,000 developers running it simultaneously at peak, before collapsing on April 4 when Anthropic cut off the third-party API access that had made it economically viable, per a detailed reconstruction of the rise and fall. Google's Remy is currently being tested by Google employees. Only by employees.

The business Google is building Remy to disrupt is, primarily, Google's own.

The ad-supported web runs on human eyeballs. Every impression, every click, every affiliate commission assumes a person on the other end making a decision. When an AI agent acts on a user's behalf — booking travel, buying products, comparing prices — it does so as a machine consumer that is faster, more methodical, and immune to the dark patterns that human-oriented sites use to extract attention. The affiliate commission structure that funds thousands of content sites assumes a human who can be shown an ad before they buy. An agent that goes directly to the merchant with a purchasing instruction does not generate that impression. No impression. No payment. No ad-supported content.

The shift is not theoretical. AI bot activity — traffic from automated systems harvesting content or completing transactions — surged 300% in 2025, with media and publishing absorbing a disproportionate share, according to Akamai's bot management platform data. The more damaging variant is the fetcher bot: systems that extract content in real time to answer user queries directly, capturing value at the moment it's created without sending any traffic back to the source site. AI chatbot referrals now drive approximately 96% less traffic than traditional search, and users click cited sources in AI-generated answers only about 1% of the time.

The structural fragility runs deeper than traffic numbers. The US digital advertising market booked $259 billion in 2025, per industry figures. That revenue is predicated on the assumption that digital real estate can be continuously monetized through human impressions. AI agents do not view display ads. They are immune to emotional or impulsive marketing appeals. They cannot be swayed by visual persuasion. An agent comparing product specs across five merchants is not generating an impression on any of them. It is executing a transaction.

"The proliferation of AI agents operating on personal devices severs the critical psychological link upon which the digital advertising revenue model depends," Klover.ai's analysis of ad-market fragility puts it directly. The economic foundation of the modern internet — built on the assumption that human attention can be aggregated and sold — faces a structural challenge it has never faced before.

The infrastructure layer is scrambling to respond. Cloudflare launched an Agent Readiness score in April, a diagnostic tool for site owners to assess how well their properties support AI agents. The framing is revealing: "The web has always had to adapt to new standards. It learned to speak to web browsers, and then it learned to speak to search engines. Now, it needs to speak to AI agents." The adoption of that framing — by an infrastructure company whose business depends on sites remaining accessible — signals that the agent-readable web is not a future scenario. It is a present infrastructure requirement.

The second-order consolidation is where the startup problem sharpens. When the agent-to-merchant transaction layer is controlled by Google, OpenAI, and Apple — the three platforms with consumer hardware and distribution — independent startups face a structural trap: they can build agentic products, but the platform layer that mediates every transaction is owned by someone else. OpenClaw demonstrated this vulnerability directly. Its viral growth was not driven by superior capability; it was driven by access to Anthropic's models at a price that made independent development economically viable, according to analysis of the collapse. When that access closed, the ecosystem evaporated overnight. The capability was real. The structural dependency was fatal.

Google has no equivalent vulnerability — but it also has no equivalent proof of demand. Agent Mode, Google's existing multi-step task feature, is live with access varying by subscription tier and region. It has limited uptake. Remy is being positioned as the consumer-facing answer. The question at Google I/O, which begins May 19, is whether that answer ships publicly, goes enterprise-only, or gets quietly delayed again.

The kill-if-false condition for this story is concrete: if Google announces at I/O that Remy is enterprise-only or indefinitely delayed, the competitive urgency narrative collapses. Google's structural incentive to move slowly is also real. The ad business books tens of billions annually from search and content. An agent that completes transactions without generating impressions is a direct threat to the page that business runs on. But so is ceding that layer to OpenAI, which already has a shipped consumer agent and recently hired Peter Steinberger — OpenClaw's original builder — to drive next-generation personal agents. Google is caught between the market it needs to create and the revenue it needs to protect. That tension is the story. Remy is the symptom.