The Google search box no longer asks for a keyword. It accepts a brain dump.
Yesterday at I/O 2026, Google announced that AI Mode has surpassed one billion monthly users, with queries more than doubling every quarter since launch. The same announcement introduced what Google called “the biggest upgrade in over 25 years” to the Search box itself, plus a new class of information agents that operate in the background, 24/7, scanning the web on the buyer’s behalf and pushing synthesized updates as conditions change.
For brands that have already done the work to be cited inside AI answers, the next twelve months are decisive. Information agents do not return ten blue links. They return one synthesized recommendation, drawn from a small pool of sources the agent has decided are credible. Sources in the pool are recommended around the clock to buyers who never have to search again. Sources outside the pool become invisible the moment the agent takes over.
What Google actually changed.
The new Search box accepts text, images, files, videos, and open Chrome tabs as inputs. It expands dynamically as the buyer types. It suggests follow-up questions before the buyer finishes formulating theirs. The box is rolling out across every country and language where AI Mode is already available.
Underneath the new box, AI Mode now runs on a version of Gemini 3.5 Flash, Google’s newest frontier model tuned for agents and coding. Per Google’s I/O 2026 announcement, AI Mode has surpassed one billion monthly users with queries more than doubling every quarter since its public rollout. The same announcement reports that nearly one in six AI Mode queries are now non-text, with users searching by image, voice, or attached file.
The most consequential addition is the new category: Search agents. Beginning this summer, AI Mode subscribers will be able to spin up information agents that monitor the web continuously, scanning blogs, news sites, social posts, and Google’s freshest real-time data feeds for changes related to a topic the buyer cares about. The agent sends a synthesized update with sources and an action button. The buyer does not search again. They wake up to a digest assembled while they slept.
The five-step agent loop.
Information agents follow a consistent pattern across categories. With the announcement live, the loop is now publicly described.
1. The brain dump. The buyer describes the problem in full prose. Not “apartments two bedroom Brooklyn.” Closer to: “Two-bedroom in Brooklyn, dishwasher, under $4,500, pet-friendly, walking distance to a 4 or 5 train, available before September 1.” The agent ingests the entire description as a single intent.
2. The fan-out. The agent breaks the brain dump into sub-topics and issues parallel queries against each. Google has been describing this as its “query fan-out technique” since AI Mode launched in 2025. A single dump produces dozens of sub-queries simultaneously.
3. Urgency triage. The agent assigns each sub-topic a freshness requirement. A hot pet-friendly listing needs alerting within hours. A neighborhood demographic shift is monthly. A long-term commute-time pattern is annual. Urgency drives polling frequency.
4. Continuous scan and triggers. The agent sets monitors against every source it has decided is credible. Blogs. News sites. Social posts. Google’s real-time data feeds for finance, shopping, sports, and listings. The scan runs without further input from the buyer.
5. Synthesized update with action. When something material changes, the agent assembles a short brief with source links and an action affordance. The buyer reads. The buyer acts. The buyer does not search.
This is not a paradigm shift in the marketing sense. The architecture is the architecture of every modern AI assistant. The shift is who runs the search. The buyer used to. The agent does now.
The source pool is the new SERP.
The output of a traditional Google search is a results page. Ten blue links the buyer evaluates. The clicks and conversions belong to whichever destination the buyer picks.
The output of an information agent is a synthesized recommendation. The buyer does not see ten options. They see one paragraph with two or three sources cited underneath, and a button. Whether the buyer clicks through to verify the citation is up to them. Most will not. Most will trust the synthesis and act.
This shifts the entire optimization target. The job is no longer to be the ten blue links. The job is to be inside the source pool the agent assembled when it ran the brief. A brand in the pool gets cited. A brand in the pool reliably across the sub-topics the brain dump covered gets cited across the entire engagement. A brand outside the pool will never surface, regardless of how that brand ranks on a traditional SERP for the underlying keyword.
We have been tracking this transition in detail since the first Perplexity citations started moving for clients earlier this year. The pattern is consistent across verticals. Brands cited in answer-engine output were not the highest-traffic destinations. They were the answer-first source pages with schema that agreed with prose and authorship signals real models could verify.
The information-agent loop concentrates that effect by an order of magnitude. The agent does not run once per query. It runs continuously against the same brief. A brand cited in the agent’s source pool gets recommended on every refresh, every alert, every update, to a buyer who is actively monitoring the category and ready to act.
The 90-day window before the summer rollout.
Information agents land for AI Pro and Ultra subscribers this summer. That gives every brand in every category roughly a 90-day window to claim a position in the agent’s source pool before the rollout reaches general availability. The foundational work (answer-first source pages, schema that agrees with prose, real authorship signals, fast site) is the work our AI-optimized websites engagement was built for, and we have written about it at length. The four moves below are specific to information agents and have to happen before the rollout, not after.
1. Write your buyer’s brain dump.
Most brands have never written down what their buyer’s first information-agent query would look like. Information agents do not accept keywords. They accept a four- to eight-sentence problem statement. Until that statement exists on paper, the brand is guessing about which sub-queries the agent will fan out into.
The exercise is short. Pick the highest-revenue topic. Write a paragraph in the buyer’s voice describing their full problem: situation, constraint, deadline, deal-breaker, the thing they are quietly anxious about. Then list the dozen sub-questions an information agent would issue against that paragraph. Each sub-question is a future source page on the brand’s domain. The gap between the dozen sub-questions and the pages the brand actually has is the work.
We have run this exercise across law firms, medical and dental practices, and home-service operations in the last six weeks. Most brands have six to eight gaps per topic. A few have all twelve. None have zero.
2. Map the source-credibility ledger for your vertical.
Information agents do not weigh every source equally. The agent’s source-credibility scoring is anchored on publications it has independently judged authoritative for a vertical. In legal services, that is publications like the ABA Journal, state-bar journals, and circuit-specific trade press. In HVAC, it is Contracting Business and ACHR News. In dental, it is the ADA News and Dental Economics. In plumbing, it is Plumbing & Mechanical and PHCPpros.
The brand that has appeared in two or three of its vertical’s anchor publications in the trailing twelve months becomes the agent’s reference point for the category. The brand that has appeared in none is treated as an unverified source and ranked under brands with editorial mentions. The pre-rollout move is to identify the two anchor publications for your vertical and earn a mention in each before the summer rollout starts. That is a 90-day target, not a 30-day target. Editorial cycles are slow, which is why the window opens now and not in August.
3. Open a freshness ledger.
The information agent runs continuously and weights recent updates more heavily than aged content. A brand whose service-page last dateModified is fourteen months ago gets sampled less often than a brand whose dateModified was three weeks ago. The agent reads this as a proxy for whether the brand is an active participant in its category or an abandoned site running on cached authority.
The fix is to publish a documented monthly cadence of time-stamped updates: one substantive update to each top-revenue source page per month, with the dateModified reflected in schema and visible on the page itself. We call this the freshness ledger. It is not new content. It is documented refresh of existing content, with the change-log visible to the agent. Brands that open this ledger now are operating on a different freshness curve than brands that do not by August.
4. Audit against three named competitors.
Run the brain dump from move 1 against your two largest competitors and the one competitor in your vertical that is moving fastest right now. For each of the dozen sub-questions, log which of the three competitors has a real source page on their domain. The audit produces a per-sub-question scoreboard.
The scoreboard tells the brand exactly where the agent will pick a competitor over them. The sub-questions where the brand has the only source page are positions the agent will hand to the brand by default in August. The sub-questions where two or three competitors have source pages and the brand does not are positions the brand will fight for in arrears. The audit is the planning artifact. The 90-day work plan falls out of it.
The window is open.
Most brands have not built for this. The content depth and editorial authority that lands a brand inside the agent’s source pool is still rare. The brands building it now are claiming citation positions while the rest of the category is still asking whether AI Overviews matters.
Information agents roll out first to AI Pro and Ultra subscribers this summer. The full rollout follows. Once general availability lands, every buyer monitoring a category is, in effect, running a 24/7 background scan against the brands that did the work. The brands in the pool get recommended continuously to active buyers. The brands outside it watch their organic traffic shape-shift around them with no clear cause.
Pricing is set against the actual work scope. The qualifying math is the payback period: a brand with topical cluster gaps and weak schema can usually rebuild into citation-ready form in eight to fourteen weeks, with first citations landing in Perplexity inside a month of source pages going live, and the rest of the platform stack catching up across the following quarter. We quote against the audit, not against a published list.
The work stacks the more agents the buyer is running. A brand cited across one sub-topic in one brief is cited once. A brand cited across an entire category’s worth of sub-topics, across every information agent monitoring that category, becomes the default reference for the category, recommended continuously, by every agent, to every buyer, around the clock. That is the position the brands moving first are taking.
Frequently Asked Questions.
Are information agents replacing traditional Google Search?
No. Google was explicit that the agent layer sits on top of the existing search engine and runs against the same fundamentals. The information agent is a new surface for buyers who want continuous monitoring instead of one-shot queries. Both surfaces are live and both will continue to be optimized against.
How long until information agents materially affect inbound for our category?
The summer rollout is for Pro and Ultra subscribers, which is roughly Google’s most engaged research tier. Effects on inbound show up first in high-research, high-consideration verticals: law firms, medical and dental practices, enterprise services. Service-area trades feel it next, as the agent loop expands to local and shopping queries through fall 2026.
Does this make our existing SEO work obsolete?
No. The schema, authorship, content depth, and speed work that win AI citations are the same work that holds Google rankings. The two surfaces share signal infrastructure. Brands that have been investing in source-page architecture for the AI answer engines are already 60 to 80 percent of the way into source-pool position for information agents.
What is the single most important move to make this quarter?
Audit the brand’s existing service pages against the buyer’s brain dump. Pick the top revenue-driving topic. Write the dozen sub-questions an information agent would issue against a brain dump for that topic. Then check whether each sub-question has a real source page on the brand’s domain, answer-first, schema-aligned, byline-attributed. The gaps are the work. Most brands have six to eight gaps per topic. Closing them is what gets the brand into the pool.
Can a brand land inside the pool without going through a rebuild?
Sometimes. Sites already on static-first architectures (Astro, Next.js static export, Hugo) can usually be retrofitted with schema and content work alone. Sites on heavy CMSes with custom theming usually rebuild faster than they retrofit. The decision lives in the week-one audit. Our agency partnership program covers the rebuild work under the partner’s brand for agencies serving clients in this position.
The takeaway.
The buyer no longer searches. An agent searches on their behalf, continuously, and reports back when something material changes. The agent’s source pool is the new SERP. The brands building topical breadth, schema consistency, authorship, editorial credibility, and speed right now are claiming positions in that pool while the category is still figuring out what just changed.
Run the brand’s top revenue topic through the brain-dump test. If most sub-questions point at competitor source pages, the agent will recommend competitors continuously to every buyer monitoring the category. The work to change that result is unromantic: source pages, real schema, real authors, fast sites. The position holds for as long as the work holds, and stacks across every sub-topic the agent monitors.
