Part 1 of 4: what changed at Google I/O 2026, and what didn't.
TL;DR
- At I/O 2026, Google announced "Information Agents": persistent, 24/7 AI agents that monitor the live web on a user's behalf, inside Search.
- This is Live Search Agent traffic - a category we've been measuring since 2024 - with a Google brand on it and is being shipped this summer.
- The dual-dependency thesis (~50% training data share, ~40% live retrieval) stopped being theoretical last Monday. ~40% live retrieval is no longer a forecast; it's a product launch.
- Every existing ad infrastructure layer - header bidding, OpenRTB, Prebid - sits at page load. Information Agents don't trigger any of it.
- For publishers, the question is no longer whether to monetise agentic traffic, it's a matter of when now.
On 19 May, Liz Reid stood on the Google I/O stage and announced what she called the biggest upgrade to the search box in over 25 years.
Most of the coverage focused on the redesigned box itself to be more in line with what's become standard for AI users aka the generative UI.
A move from blue links to AI generated answers.
But the piece of the announcement that should have made every publisher executive sit up was buried in the middle of the keynote,
and Google gave it a name we'd been working without: Information Agents.
What Google actually announced
Information Agents are persistent AI processes that run inside Google Search,
in the background,
24 hours a day.
A user describes what they want to track - flat listings under £1,500 in Hackney, a sneaker drop from a specific athlete, a competitor's earnings cycle, a flight price - and the agent scans the web continuously, returning synthesised updates with the ability to take action.
"continuously" is a scary word.
Google's own framing, from Liz Reid's blog post the same day:
"Your agent will intelligently look across everything on the web, like blogs, news sites and social posts, plus our freshest data, such as real-time info on finance, shopping and sports, to monitor for changes related to your specific question."
It's pitched as the successor to Google Alerts.
It is not.
Alerts pinged email.
Information Agents reason across sources, surface comparisons between perspectives, and hand off to other agents in the Google stack - Universal Cart for the basket, AP2 for autonomous payment, agentic booking for the reservation - to complete tasks without a human ever loading a publisher's page.
Information Agents ship this summer for Google AI Pro and Ultra subscribers,
US first, with more markets to follow.
The underlying stack is Gemini 3.5 Flash, with the agentic coding layer running on Google Antigravity - the same agent-first development platform that used to power Project Mariner, before it was closed down*.
The thesis, made product
For the last two years, we've been telling publishers a version of the following: AI is not one traffic source. It is two.
The first is training data ingestion - the slow-moving, periodic crawl that feeds the foundation models.
That's roughly 50% of the AI dependency on the open web today.
It's the layer everyone understands.
It's the layer the "should we let OpenAI train on us" conversations have been about.
And publishers have been rightfully trying to block training bots.
The second is live retrieval - AI agents that fetch publisher pages at query time, in response to a user prompt the model can't reliably answer from its training data alone.
That's roughly 40%.
This is the layer that bypasses every existing monetisation infrastructure.
It loads pages without rendering ads.
It synthesises content without sending users back to the source.
And now it generates ad impressions for the AI assistant like ChatGPT, not the publisher.
The Information Agent is that second category, named, productised, and shipping with a launch date.
What's changed in the last week isn't the thesis.
The thesis was already in the data.
What's changed is that the thesis now has a Google logo on it.
A category name.
A user count that will be public within a quarter.
A press cycle that's about to put it in every publisher boardroom.
Why the existing ad stack doesn't see it
This is the part most analyses of last week's keynote are missing. Information Agents don't trigger any of a publisher monetisation stack.
The display ad stack - header bidding, OpenRTB, Prebid, every supply-side and demand-side platform built over the last decade - all of it runs at page load, in the browser, against a human user with a session, a viewport, and a cookie.
Information Agents do none of this.
They fetch via Google-Agent, a user-triggered fetcher that Google quietly documented in March 2026*.
They identify themselves cryptographically using the Web Bot Auth protocol, signing every request with a key tied to the agent.bot.goog identity.
They render server-side, in Google Cloud.
The user never hits the publisher's page.
Result: the publisher serves the content.
The user gets the answer.
Google captures the relationship - and the ad impression, if it surfaces one inside the AI response.
The existing ad stack records nothing.
For publishers running CPM-based monetisation, this is the architectural problem.
Not "AI will take search clicks."
Not "zero-click is up."
Those are downstream symptoms, and we'll cover them later in this series.
The upstream cause is that there is now a layer of traffic - documented, identifiable, growing - that the ad infrastructure was never built to see.
What this means for publishers, briefly
We'll publish the practical playbook in part 2 of this series, but the short version is: the question stopped being theoretical last Monday.
Three things are now true that weren't two weeks ago.
First, Live Search Agent traffic has a Google-branded name - Information Agents - and a the press are loving it. The "agentic traffic" conversation just moved from technical fringe to executive briefing.
Second, the user agent (Google-Agent) and the IP range file (user-triggered-agents.json) are both public.
Any publisher can identify this traffic in server logs today.
Not next quarter.
Today.
Third, the cryptographic identity layer (Web Bot Auth) is moving fast, and the major WAF providers - Cloudflare, Akamai, AWS - are already implementing it.
The "block or allow" binary at the firewall is about to become a much more nuanced "verify, classify, and negotiate per-agent" decision.
That decision is contractual, not technical.
And the publishers ready to have that contractual conversation will compound an advantage.
The window
The publishers who instrument early, treat agentic traffic as a distinct category in their analytics, and prepare for contractual rather than blanket allow-or-block decisions will be in a different commercial position by Q4 than the ones who wait.
The ones who wait will be making the same decisions a year from now, with less data and a worse seat at the table.
We built blankspace because we saw this coming.
Last Monday at Google I/O, Google made it official.
Part 2 in this series: what publishers should actually do now, and the four week instrumentation plan we're recommending to our partners.
Sources: *https://www.theverge.com/tech/925559/google-project-mariner-shut-down *https://nohacks.co/blog/google-agent-user-agent