Ad Infinitum<br>Matthias Ott
Power
Reset
Kontrastor 82M – Dynamic Cascading Style Enhancer
Skip to main content
At Google I/O this week, the company announced the biggest change to Search in 25 years. The ten blue links? Gone. Instead, you get “generative UI” – custom interactive widgets, built on the fly by Gemini. You get “information agents” that monitor the web for you around the clock. You get mini-apps you can build right inside the search box, using natural language. And with Gemini Spark, you get a personal AI agent that runs 24/7, connected to your Gmail, your Drive, your calendar, your photos, and soon your local files and third-party services.
It’s ambitious and also impressive at times, no question. It’s also a staggering gamble. Google is dismantling the very product that built its empire – and rebuilding it into something entirely different. Watching the demos, I kept asking myself two questions. First: what if people don’t actually want to use search this way? And second:
Where did the ads go?
Google made $295 billion in advertising revenue last year, out of $403 billion in total. And yet, across all these announcements – the new search box, the generative UI, Spark, the agents – there wasn’t a single mention of how any of this will be monetized. Not one word about ads.
That seemed odd. I mentioned this on Mastodon, and Sijmen replied with a link that sent me down a rabbit hole.
It turns out there’s a small but growing body of research – some of it from Google’s own researchers – exploring exactly how ads could work inside LLM-generated output. And the ideas are as fascinating as they are unsettling.
One approach, proposed by Google Research, is what you might call a “token auction.” In this model, advertisers don’t buy ad slots on a page. Instead, they bid, token by token, on the actual text the model generates. Each advertiser brings their own LLM, and an auction mechanism decides whose model gets to influence the next word. The output is a weighted blend of competing interests, shaped by who’s willing to pay more.
Another approach – also from Google researchers – fits the new “Search” much more precisely. It’s called “prominence allocation.” Here, when a user submits a query with commercial intent, the system runs an auction that doesn’t just decide which ads appear, but how prominently the LLM writes about each one. The auction outputs a prominence score for each advertiser, essentially telling the model: give this product 35 words, that one 20, and this one zero. The ad isn’t next to the answer. The ad is the answer. Or rather, it shapes how much space and enthusiasm each product gets within the answer.
Now think about what Google just announced. Generative UI means there are no more discrete ad slots. The search result is a dynamically generated experience – a synthesized narrative, an interactive widget, a curated set of recommendations. In that world, you can’t place a banner ad next to the output. The only thing you can auction is prominence within it.
But it’s not just the ads that disappeared from the keynote. The links did, too.
For 25 years, Google Search was built on a contract. The web provided the content – billions of pages, freely linked, freely crawled. In return, Google sent people back. The link was the unit of exchange. It’s what made the Web thrive as an information system: you publish, Google indexes, users click through, and value flows back to the source. Win-win.
That contract is now broken. Generative UI doesn’t link to your article. It absorbs your article, synthesizes it into a widget, and presents it as Google’s own answer. Information agents don’t send users to websites. They deliver “synthesized updates” with maybe a link or two buried at the bottom. The web was the scaffolding Google needed to build its index, to train its models, to accumulate the world’s information, and put ads next to it to get filthy rich. Now that the content is inside the system, the scaffolding is no longer needed. Google is creating its own context.
Google thinks it no longer needs the Web to deliver answers. And it no longer needs ad slots to deliver ads. What it needs is you. Your emails, your files, your calendar, your purchase history, your travel plans – all flowing into Spark, all building the richest possible picture of who you are and what you’re likely to click on. That’s exactly the kind of personal context those auction models need to work. The prediction module in the prominence allocation framework doesn’t run on keywords. It runs on knowing you.
None of this was mentioned at I/O. The entire keynote was framed as a story about usefulness and possibility and trust. Will you opt in to sharing your data? Will you let Spark read your inbox? The pitch is productivity. But the infrastructure being built – a deeply personalized agent, connected to commercial services, generating natural-language recommendations with no visible boundary between organic and paid...