Google CEO Sundar Pichai says an agentic layer, not a product merger, will bring Search and Gemini into one box

• Gemini Spark: built on Gemini 3.5 and the Antigravity agent harness, on dedicated Google Cloud VMs. It accepts tasks at its own Gmail address, browses through Chrome and supports MCP (Model Context Protocol) connectors. It launched for US AI Ultra subscribers (TechCrunch; The Next Web).
• AI Mode: Gemini 3.5 Flash is now the default model worldwide. Information agents were due for US AI Pro and Ultra users in summer 2026, along with agentic calls to local service businesses (Google).
• Next: Gemini 4 Argon models, with wider access expected "over the next few months" (SEJ).

• Shared stack: Pichai said enterprise and consumer agents use the same model, harness and runtime; only the context differs. Each task gets its own micro VM running securely around the clock (SEJ).
• Two ways agents reach websites: they can browse pages like a person, which works now, or use protocols such as Google's Universal Commerce Protocol, which Pichai called more scalable but slower to build. He expects both to develop (SEJ).
• Crawler controls: browser agents triggered by a user act as user fetchers, which can bypass robots.txt rules (SEJ).
• No published evaluations: Google has released no benchmarks for task success or for how agents choose sources.

• Usage figures: AI Mode passed 1 billion monthly users, and its queries more than doubled every quarter. These are Google-reported numbers (Google). The video also cites 2.5 billion AI Overviews users (04:11).
• "The drop almost doubled in a year" (04:20): the two Ahrefs studies used different baselines and time windows (Ahrefs), so the 34.5% and 58% figures are not directly comparable.
• Agentic query share: Pichai said the share of more agentic queries is "steadily increasing." He gave no figure (SEJ).
• Rankings hold while clicks fall: this comes from the video's agency client work only (03:26). No independent source corroborates it.

• AI Overviews had about 2 billion monthly users and AI Mode about 100 million in July 2025 (TechCrunch).
• The video presents information agents as the successor to Google Alerts (launched 2003) and argues the agent now acts on answers instead of just showing them (06:25).
• Industry analysts describe Pichai's framing as "Search as agent manager": Search runs tasks across parallel threads instead of returning a list of answers (iloveseo.net).

• TechCrunch's Spark coverage gives no detail on confirmation steps or safety controls for autonomous actions (TechCrunch).
• Pichai sidestepped questions about how Google's ad business holds up when agents do the browsing (SEJ).
• Pichai also said compute is limited: Gemini 4 Argon processed about 25 trillion tokens in its first 72 hours (Sources). That could slow wider rollout.
• SEJ's Roger Montti questions whether human referral traffic has any place in an agentic layer (SEJ).
• The video's own forecast of a search with no human users is speculation; the presenter says "we don't know" (08:49).

• Make facts easy for agents to read: what you do, for whom, at what price, where, plus reviews. Third-party mentions may decide whether an agent picks you (07:52–08:42). This advice is from the video; it is not backed by Google.
• Prepare for both access paths: browser agents now, commerce protocols such as Universal Commerce Protocol later (SEJ).
• Track AI citations alongside rankings, since top-ranking pages lose clicks when AI Overviews appear (Ahrefs).
• Discount the pitch: the video promotes paid coaching and consulting (04:41, 09:26).

The quotes come from automatic captions, which may contain errors.