Google's AI Overviews just started building interfaces

On 19 August 2026, Google confirmed that generative UI, the feature that lets AI Mode assemble a custom chart, table or interactive tool instead of a paragraph, is expanding into AI Overviews too. Search Engine Journal reported the confirmation the same day. Google framed the release around back to school study tools, but the mechanism behind it reaches well past exam season.

Generative UI first shipped inside AI Mode in November 2025, alongside Gemini 3. At Google I/O in May 2026, Google said the capability would reach everyone in Search that summer, free of charge. Until 19 August, that promise had stayed inside AI Mode. Overviews, the box that sits above ordinary results on most searches, had only ever given you a few lines of text and a source link.

What actually changed on 19 August

Generative UI is already live globally in English inside AI Mode. From 19 August it started rolling out inside AI Overviews as well, with Google saying full availability is expected over the coming weeks. Search Engine Journal's example is a query for "pH scale": instead of a paragraph explaining the concept, Overviews can now generate an interactive scale a searcher can move along.

Google has not said which query categories will get generated interfaces most often, or how quickly the rollout reaches beyond the English, US led launch. That is worth sitting with rather than filling in with a guess. What is confirmed is narrower than the headlines: Overviews can now build a component on the fly, on some queries, in one market, expanding from there.

What generative UI is actually assembling

Google's own description of the feature is a model that returns a custom visual layout, tables, charts, simulations, controls, built for that specific query rather than picked from a template. That is a different task from writing a sentence. To draw a chart, the model needs numbers it can plot with confidence. To sequence a set of steps, it needs steps it can tell apart. To build a comparison, it needs two sides it can label correctly.

A model can extract that from unstructured prose if it has to. It does it far more reliably, and far more cheaply in compute, when the source page already hands the numbers over in a table, the steps over in an ordered list, and the entities over with clean, unambiguous names.

Our reading: structure just became UI material

We have argued for a year that structure and schema are levers for AI citation, alongside crawlability, correct delivery, freshness and original data. Generative UI does not change that list. In our view, it raises what each lever is actually for.

Being quotable used to be enough: a clear sentence with a number in it could get lifted into an AI answer as text. Being assemblable is a higher bar. A paragraph that buries three figures in prose gives a generative interface nothing clean to chart. The same three figures in an actual table give it exactly the material it needs, and a plausible reason to build something interactive around your page rather than a competitor's.

This is our reading of where the incentive points, not a mechanism Google has confirmed. Nobody outside Google can say today which pages get chosen as component source material and which get quoted as a caption underneath one. What we can say is that the pages least likely to be chosen are the ones a model cannot parse cleanly in the first place, and that is a problem site owners can actually go and check.

There is a plainer commercial reading too, and it is fair to state it alongside the structural one. An interactive component built directly inside the results page is a reason to stay on that page rather than click through to yours. Google earns the same credit either way, a quoted line or a built interface, so the incentive to keep a searcher on Overviews longer does not disappear here. That does not make the feature a bad one. It is a genuinely useful answer format for a query like the pH scale example above. It does mean that being the source underneath an interface, in a visible link or citation, matters more than ever if you want the visit as well as the mention.

What we would check on a site before assuming this helps you

• Whether your comparison data sits in a real <table> element or is buried inside a paragraph. A model can chart the first. It has to guess at the second.
• Whether a sequential process is marked up as an ordered list, or written as a run of sentences with "first" and "then" doing the structural work instead of HTML.
• Whether the content is there at all in the HTML a crawler receives, rather than only in what a browser renders after JavaScript runs. We wrote up 46 pages on our own site that were invisible to every AI crawler for a much simpler reason than this one, and it is worth ruling out first.
• Whether the page carries the schema that names its own entities and figures explicitly, rather than leaving a model to infer them from formatting alone.

Generative UI in Overviews is days old and still narrow. The direction is not hard to read. If a search result is becoming a built component instead of a quoted line, the raw material it gets built from is the same structure and schema work we already tell clients to do, just with a sharper reason to do it properly. Our GEO audit checklist is the place to start checking your own pages, and the full method sits in the AI Search Visibility Field Manual.