Google declares the query dead and forgets to mention the ad unit

Google held I/O yesterday and announced, with the calm of a company that has not fully read its own filing, that the search box is over. In its place: "information agents" that run continuously, monitor the web on the user's behalf, and push results without waiting to be asked. The model underneath is Gemini 3.5 Flash, which Google says outperforms Gemini 3.1 Pro on coding benchmarks. A new commercial tier called Google AI Ultra was introduced alongside. Demis Hassabis, on stage, said humanity is at "the foothills of the singularity." ^{1 2}

There is a lot here. Most of it is the same story told from different angles, and the story is that Google has announced the dismantling of its own monetisation surface and has not yet told anyone what replaces it.

What was actually announced. Gemini 3.5 Flash is now the default model powering AI Mode in Search, which has been rolling out in the US since earlier this year. Google's product language calls it "our most powerful model for coding and agentic tasks", with Gemini 3.5 Pro slated for June. ¹ The architectural claim is sharper than the model claim: Google is reframing Search from a pull surface (user types query, index responds, ads sit next to results) to a push surface (agents observe, agents surface, user receives). The blog post calls this an "intelligent search box" and describes agents that work "24/7" on the user's behalf. ^{1 3}

The phrase "the age of the search box is over" is attributed to a Google executive in TNW's coverage. ³ I would treat that as Google's framing of itself rather than a stray quote; it is the thing the company wants to be heard saying.

That is roughly Alphabet's 2025 revenue, the overwhelming majority of which is advertising attached to the query moment. The query moment is the thing Google has just announced it is replacing. This is the part of the announcement that deserves the most attention and received the least.

The ad unit problem, stated plainly. Google Search ads are sold against intent expressed as a query. The user types something, Google infers commercial intent, an auction runs, sponsored results appear above organic ones. Remove the query and you remove the auction. Information agents that push results to users without a triggering query do not have an obvious slot for a sponsored link. You can imagine answers: ads embedded inside agent outputs, sponsored "monitoring topics", AI Ultra subscriptions substituting for the lost ad inventory. Google disclosed none of these. AI Ultra was named as a tier and its pricing was not published. ¹

The charitable reading, and it is genuinely available, is the one the counterpoints in the research file gesture at: Google has navigated architectural pivots before. Featured Snippets, Knowledge Graph, AI Overviews — each was supposed to cannibalise clicks and each coincided with continued ad-revenue growth. The market may be underweighting Google's ability to monetise within the new surface. Possibly. But each prior pivot preserved the query as the unit of attention. This one does not. The pivot is structurally different and the silence on monetisation is structurally louder.

The Gemini 3.5 Flash claim, held at arm's length. The competitive signal embedded in the model announcement is that a cheaper, faster inference tier (Flash) now beats the prior reasoning tier (3.1 Pro) on coding evals. If that holds against third-party benchmarks, it is a real compression of the cost-per-capability curve, and it lands directly on Anthropic's Sonnet pricing rationale and OpenAI's mid-tier positioning. Inference economics is the right frame: when the cheap tier outperforms the expensive tier from one cycle ago, every competitor's pricing schedule has to be re-examined.

The claim is also, as of yesterday, a marketing assertion. Google's own benchmarks are the source. Gemini 3.1 underperformed competitive expectations on third-party evals at launch, and the same caveat applies here until independent evaluators run their own tests. ² I would wait for LMSYS, SWE-Bench, and the usual rerunners before treating "Flash beats 3.1 Pro" as a structural fact rather than a launch slide.

The agent layer as funnel. If the push model works as described, and the "as described" is doing real work, then agents become the chokepoint between publishers and readers. SEO has spent two years adjusting to AI Overviews siphoning clicks. The next shock is not fewer clicks. It is no query to rank for. If users never type, then ranking against typed queries is not the game. The game is being one of the sources an agent decides to surface, which is a different optimisation problem with different (and less transparent) levers.

This is the SaaS-apocalypse frame applied one layer up the stack. B2B content tooling and SEO platforms are built on the assumption that there is a query to rank for, a SERP to appear on, and a click to attribute. Information agents break all three. The vendors most exposed are the ones whose product assumes the pull-query world; the survivors will be the ones that learn to optimise for agent-mediated discovery, which Google has not yet documented and which will, when documented, be Google's to define.

The Hassabis line. "Foothills of the singularity" is not the register of someone disclosing benchmark data; it is the register of someone setting a horizon. ⁴ On a developer stage, with a model release that has not yet been third-party validated, I would read the line as positioning rather than evidence. The gap between the rhetorical escalation and the disclosed benchmarks is itself a data point. Frontier labs have started to talk about the singularity at roughly the rate at which they need to justify capex.

What this is a case of. It is a case of a dominant incumbent announcing the obsolescence of its own monetisation surface, naming a successor product tier (AI Ultra), declining to disclose how that tier is priced, and asking the market to assume the maths will work. Meta did a version of this with Reels in 2022 and ate roughly two quarters of margin compression before the ad unit caught up. Google's surface is bigger and the replacement is more architectural. The compression window, if there is one, will be longer.

It is also a case of competitive catch-up dressed as leadership. OpenAI's SearchGPT and Perplexity have been running versions of persistent and agentic search for over a year. ³ Google's announcement is the largest such launch by distribution, not by novelty. The "competitive signal to Anthropic and OpenAI" reads differently once you note that two of the three were here first.

What to watch. Three things, in order of how much they will tell you.

First, AI Ultra pricing when it is published. Per-seat would be the legacy play and would suggest Google is hedging. Consumption- or outcome-based would suggest Google believes the agent surface is the real revenue centre. The shape of the tier is the shape of the bet.

Second, third-party benchmarks on Gemini 3.5 Flash. If Flash genuinely beats 3.1 Pro on coding outside Google's own evals, the inference-economics story is real and competitors will need to respond on pricing within a quarter.

Third, whether Google discloses, in the next two earnings calls, any monetisation mechanic for the push surface. If it does not, the silence on the ad unit is the story, and the market will eventually price it.

Footnotes

¶