Google tests letting an LLM build the whole results page

The search results page, as a stable object with a knowable anatomy, is the thing SEO has been reverse-engineering for 25 years. Google Research's new paper, flagged by iPullRank, proposes doing away with it entirely. "Generative UI: LLMs are Effective UI Generators" describes a model that writes the HTML, CSS and JavaScript for a results page from scratch, per query, with no fixed template underneath. iPullRank reports that this is a continuation of a shift already visible in Google's results pages over the past two years, away from the ten-blue-links grid and fixed modules toward assembled, query-specific layouts.

The distinction matters more than it sounds. Featured snippets, knowledge panels and AI Overviews are all variations on a theme: a model selects content and drops it into a pre-built container. The container has rules. Marketers can study those rules, the way a navigator studies currents. A generative UI has no container. The page is the output of an inference pass, as bespoke as the prose in a ChatGPT answer. There is no template to reverse-engineer because there is no template.

Why a stable SERP was the last fixed point

Everything the SEO and content industry does rests on the assumption that a ranking position is a recoverable, comparable thing: you can check it today, check it next week, attribute a change in traffic to a specific cause. Generative UI removes that fixed point. If the page is rebuilt from first principles for every query, "ranking" stops being a meaningful concept in the way it has been for two decades. What replaces it is something closer to the logic already governing ChatGPT and Perplexity answers: inclusion, not position. Either a brand's content informed what the model generated, or it did not. There is no eleventh position to climb into.

For an industrial conglomerate or a multilateral institution with a communications team built around ranking reports and share-of-voice dashboards, that is a structural problem, not a tactical one. These organisations have spent years optimising for the idea that a page exists somewhere between position one and position fifty. A generative results page does not have fifty positions. It has one assembled output, and the brand is either a visible ingredient in that output or it is invisible. The World Bank's CGAP arm, Holcim's sustainability reporting, IEEE's standards documentation: all currently measurable, however imperfectly, by rank-tracking tools. None of those tools has an answer for a page that does not exist until the query is typed.

The paper is research, not a roadmap, but the direction is not in question

Google Research publishing a paper is not the same as Google shipping a product, and iPullRank is careful to frame this as a signpost rather than an announcement. Google has shipped nothing resembling full generative UI in production search. But the direction of travel across the past two years (AI Overviews, AI Mode, the steady erosion of fixed SERP furniture) points the same way, and Google has a known pattern of previewing architectural intent in research well before productisation. Treating this paper as an academic curiosity would be the same mistake as treating the first AI Overviews pilot as a sideshow in 2023.

What should change now, before the architecture exists, is the unit of measurement brands use to judge their own visibility. Rank tracking assumes a stable page. The more durable metric is whether a brand's own content, data, and named expertise show up as inputs when a model is asked to synthesise an answer in that brand's domain, regardless of what container the answer eventually gets poured into. That is already the right way to think about ChatGPT and AI Overviews. Generative UI just makes it obligatory, because soon there may be no container left to track at all.

A results page assembled per query, by a model, with no fixed template, also quietly resolves the debate over whether AI answers "cannibalise" clicks from traditional rankings. There will be no traditional ranking left to cannibalise. The comparison brands should be running now is not this week's position versus last week's; it is whether their content is the kind a model would choose to build with.