How Google AI Mode works and how it is evolving in 2026

AI Overviews and AI Mode are two different things. An AI Overview is the AI answer box that sits at the top of an otherwise normal results page, with the usual ten blue links below it. AI Mode is a separate, dedicated, conversational, multi-turn full-page search experience where you ask a question, get a synthesized cited answer, and keep following up. At Google I/O 2026 the two were unified into one flow.

AI Mode vs AI Overviews, and how I/O 2026 unified them

For most of 2025 these were distinct surfaces. AI Overviews reached over 2.5 billion monthly users and stayed anchored to the classic SERP. AI Mode launched in Search Labs as a US opt-in in March 2025, got its full US rollout announcement at I/O in May 2025, and completed that US rollout around mid-June 2025 (blog.google).

At Google I/O 2026, Google unified them so that users can, in Google's phrasing, "flow effortlessly from your question, to a search results page with an AI Overview, to a follow-up in AI Mode." In practice that means the boundary between the two is now a continuum rather than two separate products.

The strategic read: Google is making an AI-first surface the default and letting a user slide from a quick answer into a deeper conversational session without leaving the flow.

Query fan-out, and why it matters for visibility

Query fan-out is Google's own official term. Instead of running your one query, AI Mode "breaks your question into subtopics and issues a multitude of queries simultaneously," then synthesizes the results into a single cited answer (blog.google, I/O 2025). Analysts observing the feature have reported on the order of eight to twelve or more parallel sub-queries per question (aleydasolis.com). Deep Search extends the same mechanism much further, running what Google describes as "hundreds of searches" to produce an "expert-level fully-cited report."

Here is why this reshapes visibility work. If a user asks "what is the best CIAM platform for a mid-market SaaS company," AI Mode does not just search that string. It fans out into the sub-questions behind it: what CIAM is, what mid-market buyers need, which vendors serve that segment, pricing models, compliance requirements, and so on. The synthesized answer cites whatever pages best answered each sub-question.

The unit of competition is no longer the head term. It is the set of sub-questions that fan out from it. You can rank well for the head query and still be absent from every sub-query the model actually issued.

So the practical goal shifts from "rank for the keyword" to "be citeable for the constituent sub-questions." That means content that cleanly answers the discrete pieces of a topic, not just the umbrella phrase. See AEO vs GEO explained for how this splits into two related disciplines, and measuring AI visibility for tracking whether you show up.

The Gemini models behind AI Mode over time

AI Mode has not run on one static model. The lineage matters because capability and default behavior shifted several times in about a year.

Gemini 3.1 Pro has been referenced as a benchmark baseline, but I could not confirm an exact release date, so treat that as reported and unverified. The takeaway for practitioners is that the model powering the default AI Mode experience changes on the order of every few months, and each change can move how answers are synthesized and which sources get cited.

Agentic features: from answering to acting

The larger shift in AI Mode through 2025 and into 2026 is from answering questions toward completing tasks.

• Project Mariner task completion. Introduced in AI Mode at I/O 2025, it can handle tickets, reservations, and appointments through partners including Ticketmaster, StubHub, Resy, and Vagaro. By the August 21 2025 180-country launch, restaurant reservations were live.
• Agentic checkout. AI Mode can buy on your behalf via Google Pay.
• Search Live (Project Astra). Real-time voice and camera search, announced at I/O 2025, launched in the US in July 2025 and in India on October 8 2025.
• Information Agents. Always-on agents that do 24/7 monitoring, announced at I/O 2026 for a summer 2026 arrival.
• Universal Cart. A shopping cart that persists across the experience, part of the I/O 2026 shopping announcements.

I/O 2026 also billed a "biggest Search box upgrade in 25+ years," accepting text, image, file, video, and Chrome-tab input, plus Generative UI, custom mini apps, and SynthID's "Is this AI generated?" checks.

Scale and reach

AI Mode crossed 1 billion monthly users by I/O 2026, and Google said it had been "doubling quarterly since launch" (blog.google, I/O 2026). The geographic expansion moved fast:

• US: March to June 2025
• India: June 24 2025
• UK: August 2025
• 180 countries, English-only: August 21 2025 (searchengineland.com)
• 35+ languages across 40+ countries, then 200+ countries by around October 2025

Personalization deepened in parallel. An optional Gmail connection arrived at I/O 2025. By I/O 2026, "Personal Intelligence" was extending to roughly 200 countries and 98 languages, connecting Gmail, Photos, and soon Calendar to tailor answers.

What it does to clicks and citations

This is where practitioners should be careful with numbers, because the strongest figures are contested.

A Pew Research Center study published July 22 2025 (about 68,000 queries from roughly 900 US adults) reported that users clicked a result on 8% of searches that showed an AI summary, versus 15% of searches without one; that only 1% clicked a link inside the summary; and that 26% of sessions ended on the results page with a summary present versus 16% without (pewresearch.org). Google publicly disputed the study's methodology and conclusions (ppc.land). Report the Pew numbers as Pew's, and note the dispute.

Separately, Ahrefs reported in February 2026 that AI Overviews correlate with roughly a 58% click-through-rate reduction for top pages, up from about 34.5% in April 2025 (searchenginejournal.com). Treat that 58% as reported and unverified on the exact figure. Importantly, do not conflate this Ahrefs CTR "58%" with the unrelated "58%" number that circulated in antitrust coverage; they measure different things.

The honest summary: multiple independent measurements point the same direction, fewer clicks reaching publisher pages when an AI answer is present, even where the exact magnitudes are disputed. For a deeper playbook on the traffic side, see surviving the AI Overviews traffic drop.

What this means for AEO and GEO practitioners

The direction is clear even where individual figures are not. Search is moving from returning answers toward taking agentic action, and the default surface is becoming AI-first and conversational.

For AEO and GEO teams, that reframes the work:

• Optimize for fan-out sub-queries, not just the head term. Break a target topic into its component questions and make sure each has content that answers it cleanly and directly.
• Invest in entity clarity. The synthesizer has to understand what you are, what you cover, and how you relate to the entities in a query before it will cite you.
• *Be safe to cite.* Accurate, current, well-sourced, and unambiguous content is what a cited-answer engine reaches for. Ambiguity and unsupported claims get skipped.
• Track presence in synthesized answers, not just rank. Rankings on the classic SERP tell you less than whether you appear in AI Mode's cited output. See measuring AI visibility.
• Plan for the agentic turn. As checkout, booking, and always-on agents move to the center, structured, machine-actionable information about your product, availability, and pricing starts to matter as much as prose.

What to do now

• Map your top topics into their fan-out sub-questions and audit whether you answer each one.
• Tighten entity signals: consistent naming, structured data, clear "who and what" pages.
• Fact-check and date your content so it is safe for a synthesizer to quote.
• Start measuring AI Mode citations and AI Overview presence, not only classic rankings.
• Watch the agentic features (Mariner, checkout, Information Agents) and prepare structured, actionable data for when action, not just answers, is the default. For the longer horizon, see the future of search marketing.