For more than 25 years, Google’s search box was a narrow white rectangle that trained billions of people to compress their thoughts into a handful of keywords. That box is gone. At its I/O 2026 developer conference in May 2026, Google unveiled a search interface that dynamically expands as you type, stretching to accommodate full paragraphs, multi-sentence instructions, and the kind of detailed requests that used to require five separate searches. The company’s product blog called it “the biggest upgrade to our Search box in over 25 years,” and the timing is not accidental: AI Mode, Google’s conversational search layer, has already surpassed 1 billion monthly users, a figure disclosed by Liz Reid, Google’s head of Search, at I/O 2026 in May 2025 and reported by the Associated Press.
The old box was a bottleneck. A billion people were already using AI-powered search to ask complex, conversational questions, and they were doing it through an input field designed for “cheap flights LAX.” The redesign removes that friction. Type a few words and the box behaves like it always did. But start describing a week-long trip to Japan with a $3,000 budget, a preference for direct flights, and a need for a hotel near public transit, and the box grows to fit the whole request. Google’s AI systems then decompose that paragraph into subtasks, research each one, and return a consolidated plan.
Search becomes a task engine
The expanded box is the most visible piece of a deeper architectural shift. Google also introduced what it calls Search agents: background systems that gather information across multiple sources to handle multi-step queries without requiring users to manually refine their searches. During the I/O keynote, a travel-planning demo showed an agent that could research destinations, compare airline prices, check hotel availability, and assemble a bookable itinerary from a single detailed prompt. These agentic travel-booking capabilities were demonstrated on stage but have not been independently verified as live or functioning as described outside of Google’s controlled presentation.
AI Overviews and AI Mode are now built to handle questions “no matter how long or complex,” with a seamless transition from a single AI-generated summary into a back-and-forth conversation. The practical effect is that Search is becoming less of a lookup tool and more of a task-completion engine. Instead of typing “flights to Tokyo” and then clicking through filters, tabs, and separate hotel searches, users are nudged to describe the entire job they want done in one go.
The interface assumes people will paste itineraries, draft emails for feedback, or pose multi-part research questions. Google’s AI orchestrates the steps required to respond. That is a fundamentally different interaction model from the keyword-and-links paradigm that defined web search for a generation.
What the independent research shows
Google’s announcements describe capabilities and user counts, but the company has not released internal telemetry or A/B testing data showing how paragraph-length queries change user behavior. There are no public metrics on whether longer prompts lead to deeper sessions, more clicks on external links, or fewer visits to publisher websites.
Independent researchers have started filling some of those gaps, though their view is necessarily from the outside. A large-scale measurement study published as a preprint on arXiv examining AI Overview behavior found differences in how AI-generated answers select and cite sources compared with traditional ranked results. The authors measured atomic-claim support rates and source-selection patterns, suggesting that AI Overviews may pull from a narrower or different set of publishers than the classic blue-link format.
A separate empirical preprint comparing outputs across Google Search, AI Overviews, and Gemini variants found meaningful differences between retrieval-based and generation-based answers. The researchers documented variation in factual support, citation behavior, and the mix of sources surfaced, indicating that generative systems do not simply restate the same information users would have found through traditional ranking.
Neither study had access to Google’s internal ranking algorithms or source-selection logic. They measured external outputs rather than internal decision-making, which means their findings describe patterns without confirming causes. Google has not publicly responded to these assessments or offered its own accuracy benchmarks for AI-generated answers.
The publisher question no one can answer yet
If AI agents complete tasks inside Search without sending users to external websites, traffic to news outlets, travel blogs, and e-commerce platforms could decline. The travel-booking workflows showcased at I/O envision a user staying within Google’s interface from initial inspiration through final reservation. That is a product design choice with real economic implications for every business that depends on Google referral traffic.
Yet no longitudinal publisher revenue data tied to AI Mode usage has been made public, either by Google or by independent researchers. The early preprint research on AI Overviews suggests that generative answers reshape which sources are surfaced and how often they are cited, but it does not quantify downstream effects on traffic or revenue. The direction of the trend seems implied by the product design; the magnitude remains unknown.
There are also open questions about reliability. Google has not shared aggregate error rates for AI Overviews or for the new agents when executing multi-step tasks like booking travel. Without benchmarks, it is difficult for users, regulators, or publishers to assess how often the system hallucinates, omits key constraints, or misinterprets complex instructions. The company has emphasized safety layers and human evaluation in broad terms but has not provided task-level reliability data.
What this means for the people typing into that box
For everyday users, the change is straightforward: the search box now accepts and encourages longer, more detailed queries, and Google’s AI will attempt to handle them as multi-step tasks rather than single lookups. Complex planning and research should feel easier, especially for people comfortable describing their needs in natural language. But the opacity around how answers are sourced and how often agents get things wrong means caution is warranted, particularly when decisions involve money, health, or legal risk.
For publishers and businesses that depend on search traffic, the central uncertainty is how much of the user journey will stay inside Google’s interface. Until Google or independent teams release data linking AI Mode usage to referral patterns, projections about winners and losers remain speculative.
Why the gap between shipped features and public evidence defines this moment
The distance between what Google ships and what outside researchers can measure is the defining tension of this redesign. A bigger search box and more capable agents clearly signal a strategic bet on conversational, task-oriented search at a scale no competitor has matched. Whether that bet pays off for the broader web, not just for Google, depends on evidence that as of June 2026 has not yet been made public.
More from Morning Overview
*This article was researched with the help of AI, with human editors creating the final content.
