Google’s generative UI is being introduced to AI Overviews this week, enabling interactive tools to be generated directly within search results. I covered the announcement on Wednesday.
This expansion is especially important for websites that depend on search traffic from embedded tools. In a previously published Google paper, the team evaluated generated interfaces against real websites, and results suggest that the rollout may extend beyond educational aids.
What Happened #
Generative UI is now accessible worldwide in English for AI Mode and is starting to be introduced in AI Overviews. Google shared this update on Wednesday as part of its back-to-school features, though the announcement doesn’t specify when the full generative UI rollout in AI Overviews will be complete.
This feature creates a personalized visual layout whenever Google’s model finds it helpful, which might include an interactive tool or a simulation tailored to your query.
For instance, if you search for “pH scale,” you might see an interactive visualization in an AI Overview. If you ask a follow-up question, it shifts to AI Mode, providing a more specific response, such as plotting citrus fruits on the same scale. Google has geared this launch towards education and studying, even adding practice quizzes to both features as part of this rollout. But the generative UI system is versatile and already supports other tasks, such as a mortgage calculator in AI Mode that helps compare different loan options.
From Answers To Interfaces #
Google Research introduced generative UI in November 2025, along with Gemini 3, and launched it within the Gemini app and AI Mode, where it was available to AI Pro and Ultra subscribers in the U.S. At I/O 2026 in May, Google announced that generative UI would become free for everyone in Search that summer. The recent rollout is the next step, expanding from paid subscribers to a global English version in AI Mode, and now also available in AI Overviews.
Sundar Pichai shared at I/O that AI Overviews now reach over 2.5 billion monthly active users, and that AI Mode has surpassed 1 billion.
One key difference between the two is that people need to actively select AI Mode, whereas AI Overviews automatically appear on the default results page. People also click traditional search results less often when an AI summary appears. A Pew Research Center analysis found that only 8% of visits with an AI summary led to a click, compared with 15% without. I took a closer look at this in my coverage of AI Overviews and how it impacts publishers. Summaries compete with the pages people read, while generated interfaces compete with the pages people use.
What Google’s Testing Found #
The paper supporting the November launch, “Generative UI: LLMs are Effective UI Generators,” reached arXiv in February. It tested how people respond when a generated interface was compared directly with a real website.
The team showed participants five results for the same prompt: a custom website created by a human, the top Google Search result for the query, a generated interface, and two plain LLM outputs in markdown and raw text. Participants saw pre-cached results, so generation speed wasn’t a factor.
Participants favored the generated interface over the top search result in 90% of head-to-head comparisons on the main prompt set. The custom-built sites performed well, with 50% of wins against generative UI, which won 35.3%. The rest were rated neutral.
The mix of prompts limits how far the 90% figure can go. The main set of prompts came from LMArena, a chatbot evaluation platform, and these prompts resembled requests for chatbot interactions more than traditional search queries. On Google’s second set, focused on information seeking, the generative UI still came out on top 73.5% of the time, compared to just 19% for the top search result. The paper suggests that the narrower gap is due to differences in the types of prompts.
The custom sites in the test weren’t from well-established websites. Google hired freelance developers, paying $100 to $130 per site, with about three to five hours of work each. They asked these developers to prioritize user experience over SEO. So, these weren’t mature, highly refined tool pages.
The gap with the custom sites is a snapshot, not a ceiling. Google describes a robust generative UI as an emergent capability of its latest models. Each new generation of the model scored better than the last in their tests. For example, on the LMArena set, the reported output-error rate dropped from 29% with Gemini 2.0 Flash to zero with Gemini 3.
Google is grading its own feature here, and it names the current limitations. Generation times can be a minute or two, and errors might sometimes slip into the output. To support further research, Google is also releasing PAGEN, the dataset of contracted expert sites, so researchers outside the company can perform similar comparisons.
Which Tools Are Easiest To Reproduce #
Generative UI produces a single web page with scripts, built from the query plus search and image tools. That scope shapes which site tools it can plausibly match.
The easiest tools for generative UI to reproduce involve taking generic inputs and running public math. Mortgage calculators, unit converters, and learning simulations all fit this description, and Google has already demoed several of them.
Tools that depend on non-public data, logged-in workflows, or integrations beyond the page’s reach may be harder to reproduce. The paper calls generated interfaces “ephemeral,” created for one prompt rather than stored.
Google describes it as standing in for the team that would have built the tool:
“Generative UI enables us to spin up an instant (AI-based) product management, UX design, and engineering teams, to build an interactive experience, over the course of a minute, for a specific prompt. While not as competent as human experts, Generative UI enables custom experiences for any prompt.”
That formula-versus-data boundary isn’t fixed, though. At I/O, Google previewed mini apps and dashboards for subscribers alongside generative UI, so the distance between a one-off generated page and a standing tool is already shrinking.
The more cautious interpretation is that formula-based tools are increasingly overlapping with generative UI, but proprietary data and workflows might remain distinct and relevant for longer.
Why This Matters For Search Professionals #
Education websites face the most immediate overlap from generative UI, as Google initially launched study tools such as interactive explainers, simulations, and practice tests. The practice quizzes added to Search directly compete with test-prep content.
Publishers with tool sections face broader overlap with generative UI; for example, a finance site’s payment calculator or a health site’s BMI tool can now be matched by a generated interface inside the result that would have sent the click.
Over the next quarter, analyzing Search Console data for impressions and clicks on tool-page queries will reveal whether this rollout has an impact, even if the cause cannot be pinpointed.
SaaS companies offering free tools are also at risk. For example, a free converter or grader used to initiate customer engagement could be replaced by an interface generated directly in search results, potentially capturing the first interaction.
Agencies advising on content strategies should now consider whether building a tool makes sense before pitching it. A calculator designed to attract links and traffic might be competing with a version that Google can produce instantly at the query stage. A stronger pitch is developing a tool based on the client’s proprietary data, which may have less direct overlap with Google’s offerings.
What Nobody Can Measure Yet #
Search Console’s new Generative AI performance report, which has been gradually rolling out to some properties since June, displays impressions from AI Overviews and AI Mode, categorized by page, country, device, and date. However, it does not differentiate between a generated interface and a standard AI Overview, nor does it track clicks. Therefore, if generated tools lead to fewer clicks on utility pages, this decline will be indistinguishable from other effects of AI Overviews.
As of this week’s announcement, Google has not clarified which query types trigger the generative UI, nor has it shared how frequently the feature appears. Without this trigger data, site owners can only estimate their exposure based on the types of tools.
Additionally, the preference data reflects user choices in a controlled lab setting rather than real-world behavior. Raters compared completed pages side-by-side, with the generation wait omitted. According to the research version, this process often took 1 to 2 minutes, but streaming reduced this wait time by about half. The paper does not provide latency data for the current Search features, nor does it test whether user preferences remain consistent after that wait.
Looking Ahead #
The generative UI rollout in AI Overviews is currently in progress, with no specified end date in the announcement. It also does not specify which languages are included, while the AI Mode launch is global in English.
Google’s paper concludes with a statement about their future goals:
“It is early days for Generative UI, and important limitations exist. Yet, we are excited about a future where users don’t have to pick from a finite library of applications or visual pages, but instead, they have access to an infinite catalog, where the right ephemeral interface is generated on the spot tailored for their need.”
You don’t need trigger data to get ready. List the tools on your site that a generated page could rebuild from public information, and note which ones depend solely on data you hold. When generated interfaces start appearing for tracked queries, that list indicates where to look first.
More Resources
Google Revamps Image Search, Adds Generation To AI OverviewsGoogle’s AI Generated Landing Page Patent Is Limited To Shopping & AdsGoogle’s New Search Box Hands Queries To AI Agents, I/O Reveals
Featured Image: Golden Dayz/Shutterstock