Google CEO Sundar Pichai told podcaster Alex Heath that Search and Gemini will not be merged into one product. A layer of cloud-based agents will tie them together instead. Search Engine Journal covered the interview in two reports on October 22, one by Roger Montti and one by Matt G. Southern. The stakes are practical: a growing share of site visitors may be software acting on behalf of a person.
Heath noted that Gemini and Search each serve enormous audiences with overlapping AI features, then asked whether they would ever fully combine. Pichai said he expects agents to supply the shared experience, not a product-level convergence. His reasoning was that agents run around the clock in the cloud and carry a user’s context, so a task started in one place can resume in another. He summed up the near-term plan as “one box.”
Pichai said Google’s own usage data shows the agentic share of Gemini queries climbing steadily. He gave no figure. The capabilities currently sit mainly with Ultra and Pro subscribers, and he said Google will widen access over the next few months. He described the push into Search as happening “in a pretty deep way,” and added that consumer and enterprise agents share the “same model, same harness, same runtime.”
The second report covers how those agents will reach websites. Heath described the current arrangement, where agents operate browsers built for people, as “like trying to fit a triangle into a square.” Pichai called it “a complex moment” and predicted the technology would outrun the ecosystem around it. He pointed to momentum behind MCP, agent protocols, and payment protocols, and expects “a chaotic evolution” before standards converge.
Pichai does not see one path replacing the other. Browser-based computer use is “a very elegant way to work too,” he said, because it mirrors what a user would do and keeps people in control. Protocols are “a far more scalable approach” that would work anywhere, but he believes they will need more time to mature. “I think you have both ways and they’re going to develop,” he said.
Google already runs infrastructure on both sides. The Universal Commerce Protocol (UCP) is an open standard Google built with partner companies, and Google’s UCP documentation says retailers can switch on agent-driven actions inside AI Mode and Gemini, starting with buying. On the browsing side, Google lists Google-Agent among its user-triggered fetchers, the user agent used when agents hosted on Google’s servers visit pages for someone who asked. Google says such fetchers generally ignore robots.txt, since a person initiated the request. It also warns that user-agent strings can be spoofed and publishes the IP ranges these fetchers use.
Heath pressed on advertising, noting that Google’s large ads business depends on engagement and citing Amazon’s September decision to block Meta’s Muse agent for not identifying itself. Pichai answered that “the right answer longer term is you put users first and say what would users want.” He compared the moment to the early internet, when quick change was predicted for home buying and selling that never fully arrived, and said agents will produce “micro equilibriums.” Southern noted that the answer left open how Google’s ads would work if an agent is the one visiting pages.
That gap matters, because the interview contains claims and no measurements. No percentage for agentic queries, no timeline beyond “the next few months,” and no independent data on how much agent traffic reaches publishers. Montti raised the open question of whether the agentic layer leaves room for human referral traffic at all. Pichai’s comments do not settle it.
The measurable piece is already in your logs. Because Google-Agent visits are user-triggered and generally bypass robots.txt, blocking them there is not a reliable control. Search teams should identify that user agent in server logs this quarter, verify it against Google’s published IP ranges, and separate it from human sessions before agent traffic distorts conversion and engagement reporting.
Search Engine Journal, reporting by Roger Montti and Matt G. Southern, published October 22, 2026.
