The Shortlist You’re Not On

Last updated: 2026-05-25

Type: Original article / reactive commentary — Google I/O 2026

At Google I/O last week, Google announced that AI Mode has crossed one billion monthly users. Queries are more than doubling every quarter. The search bar is being redesigned for the first time in 25 years.

Those are the headline numbers. The more important announcement was buried beneath them.

Google is launching information agents: persistent, background AI processes that monitor the web 24/7 and surface synthesised findings without being asked. No query required. The agent watches topics on a user’s behalf and delivers what it finds — relevant brand updates, price changes, competitor activity, vendor options — as push notifications, at whatever moment the information matters.

And alongside it: Universal Cart, a cross-merchant purchasing layer built into Search, Gemini, YouTube, and Gmail. Agent Payments Protocol (AP2), a framework allowing AI agents to complete purchases on users’ behalf within conditions the user specifies upfront.

Conditions like: preferred brands.

This is not zero-click search. Zero-click still starts with a query. The user types something. A result appears. A brand gets cited or it doesn’t. The interception point exists.

What Google announced last week eliminates the interception point.

Not Zero-Click. Zero-Query.

The progression is worth stating plainly.

Traditional search: user forms a query, gets ten blue links, clicks one or two. Brand visibility = ranking.

Zero-click search: user forms a query, gets an AI-generated answer at the top of the page, doesn’t click. Brand visibility = citation in the AI answer.

Zero-search discovery: user delegates a monitoring task to a persistent agent. The agent researches, shortlists, and reports back. The user sees the conclusion. They never formed the query. Brand visibility = being in the agent’s knowledge before the task was assigned.

Each stage doesn’t merely reduce traffic. It moves the moment of brand inclusion further upstream — further away from any content your marketing team can optimise in real time, and closer to what the AI already believed about your brand before it was asked.

Google’s information agents are the first mass-market implementation of zero-search discovery. The concept was documented as theoretical in December 2025. It shipped in May 2026. One billion monthly AI Mode users now have access to it.

The Shortlist Was Already Closed

Here is the detail that changes the strategic picture.

When Google launched Agent Payments Protocol, they gave users the ability to specify conditions for autonomous agent purchases: preferred brands, product types, spending limits. The agent acts within those constraints. It doesn’t ask at checkout what brand you want. It was told upfront.

That preference list was formed before any agent was launched. It reflects what the user already believed about brands in a category — the brands they think of as credible, relevant, and worth considering. That belief was shaped by what they have read, heard, and asked AI assistants over weeks and months.

Which is exactly what AI training knowledge and entity confidence determines.

The shortlist a user gives their agent is, in practice, a projection of the AI’s own brand knowledge back through the user. A user who regularly asks ChatGPT about project management tools, CRM platforms, or accounting software has absorbed an AI-mediated view of their category. The brands that appeared reliably in those answers — cited clearly, described accurately, associated with the right attributes — are the brands that make the agent’s preference list.

The brands that didn’t appear, or appeared inconsistently, or were confused with a competitor, are not on the list. They cannot optimise their way onto it now. The list is closed.

The Agent Doesn’t Browse. It Recalls.

There is an important distinction between what retrieval-optimised content does and what the agent draws on when it shortlists.

Retrieval optimisation — the discipline of structured data, BLUF formatting, semantic HTML, citation-ready content — affects what happens when an AI engine executes a live query. The agent fetches pages, scores them, extracts information, and constructs an answer. Content quality determines whether your brand makes that answer.

But when an information agent is monitoring a topic on a user’s behalf, it is not starting from scratch on every cycle. It is drawing on its established understanding of the category — who the credible players are, what they offer, which ones are reliably accurate — before it retrieves anything new. The shortlist it builds reflects both what it retrieves and what it already knew.

This is the distinction the wiki captured as “training mode” versus “retrieval mode.” For popular entities, training-time knowledge dominates. The model’s answer about who the leading vendors in a category are reflects patterns absorbed across billions of documents, not the page it retrieved this morning. Retrieved content updates and corrects at the margins; it does not override.

For mid-sized brands — the ones not prominent enough to have saturated training data — retrieval matters more. Fresh, well-structured earned coverage in credible sources can shift what the agent finds and reports. But the baseline the agent works from is still its existing category understanding.

Which means: retrieval optimisation keeps you visible to agents actively searching. Entity confidence determines whether you were on the shortlist before the search began.

What This Means in Practice

Google’s information agents launch first for paid subscribers (Google AI Pro and Ultra) in the US this summer. Universal Cart rolls out across Search and Gemini, with YouTube and Gmail to follow. The distribution is initially constrained.

But AI Mode already has one billion monthly users. Queries are doubling every quarter. The agents are the next layer, not a separate product. The direction of travel is unambiguous: AI-mediated purchase journeys are moving from “the user searches and decides” to “the user delegates and approves.”

In that model, the purchase decision has already been shaped before the agent receives its instructions. The category shortlist, the trusted vendors, the brands worth considering — these are background knowledge that the user carries into the agent conversation. They were formed over months of AI-assisted research, not in the moment of purchase.

For brands, the practical implication is uncomfortable. The content sprint, the PR push, the structured data audit — these are retrieval optimisation. They affect what the agent finds when it searches. They do not reach back in time to correct the category knowledge the user already has. They don’t change what the user tells their agent to prefer.

The gap between “what we optimised for” and “what the agent knew before it started” is the gap where most brands are losing.

The Entity Confidence Response

The question is not whether information agents change things. They do. The question is what brands can do about it, and on what timeline.

Three things are in scope:

Build category knowledge in AI systems now, not later. The training-mode channel — community presence, third-party editorial coverage, Wikipedia-adjacent authority signals — shapes what AI systems know about a category before any retrieval happens. This is the long cycle. It cannot be compressed. Brands that start it now are building the knowledge base that will be in the next generation of training data; brands that wait are not.

Make your entity unambiguous. If an AI agent is shortlisting vendors in your category, the first thing it needs to do is correctly identify who you are. Identity drift — the failure mode where a brand appears under multiple name variants without canonical linking — ensures you are either absent from the shortlist or confused with a competitor. Before any agent can recommend you, it must know that the company the user mentioned, the domain it found, and the brand in the editorial coverage are the same entity. That is a structured data and canonical naming problem. It is solvable today.

Measure what agents can see, not what analytics can count. Zero-search discovery is invisible to standard analytics. No query, no impression. No click, no session. No session, no attribution. The brands that will navigate this era are the ones that begin measuring AI citation frequency and entity representation now — not as a supplement to traffic data, but as the leading indicator for a category of brand influence that traffic cannot capture.

The Closing Argument

Google’s information agents are not a feature. They are the endpoint of a direction that has been visible for three years: AI moving from answering questions to completing tasks; from citing brands when asked to recalling brands when needed.

The brands that win in that model are not the ones that appear when searched. They are the ones that were already part of the AI’s understanding of the category — present in training data, clearly identified as entities, accurately described across multiple independent sources — before any user opened a search box.

You cannot optimise your way onto a shortlist that closed before the query.

The work that determines whether your brand is on that list is happening now, in the slow accumulation of editorial coverage, entity signals, and community mentions that shape what AI systems know. It is not a campaign. It is not a sprint. It is the background condition on which every future agent decision will be based.

Start building it before the agents start working.