From keywords to questions

For 25 years, people have relied on Google Search as their most trusted source for information. Over 2 billion use it daily. But the existing paradigm still asked them to think in keywords, not questions — complex, open-ended conversation wasn’t possible within its technical and interaction constraints.

By 2024, AI had changed that. For the first time, you could ask questions the way you actually think: complex, open-ended, conversational. AI could finally adapt to people instead of the other way around.

But changing Search’s core experience is risky. Billions of people rely on it every day for the authority, the reliability, the sense that it simply gets things right. It still drives the majority of the company’s revenue. Getting it right wasn’t optional. The answer was a new surface: AI-native in design, Search-grounded in values. AI Mode.

My team’s job was to humanize the experience and give the product voice: Who is Search when it can speak? What does it know about you, and how does it use that? What can it do on your behalf and how does it earn your trust to do so?

In AI Mode, people talk directly to the model. The UI is still there, but the experience hinges on something less tangible: how it speaks, what it knows, and what it does on your behalf.

Three things made this hard.

1. Search’s brand principles favor speed, efficiency, and authority. These sat in direct tension with what people expect from conversational AI: warmth, collaboration, and personality.
2. Search’s content designers wrote scoped, bounded UI copy. Now the job was to define a personality.
3. Training a model to adopt a new persona, grounded in the web, with Search’s factuality requirements, hadn’t been done before.

The core problem was definitional: what “good” meant for a conversational Search experience wasn’t established.

Early testing showed Search’s AI capabilities were strong, but the experience fell short on collaboration and tone, especially for younger people. Even when the model was accurate, it was tone deaf. If someone said thank you, the model responded with Assistance can be provided. When someone asked how are you, it replied: “how are you” is a common greeting phrase. Correct but not conversational.

The most consequential decision was first-person voice. My team championed it at the most senior levels, against real concern about what it might do to the brand and the trust people had built with it over 25 years. Google Search had always been third-person, neutral, authoritative. The use of “I” fundamentally changed the relationship between the product and the user. We revised the brand guidelines to make it possible: a bounded expansion that let the model engage collaboratively without ceding authority.

Persona improvements drove gains in AI Mode usage and thread depth. More durably, the workstream established a discipline that hadn’t existed before at Google: defining and training a voice into a search product.

Google had a longstanding perception problem: people overestimated how much personalization was happening in Search, and that gap was also a missed opportunity. Personalization had become a baseline expectation across modern apps.

In 2024, I co-led a presentation of our long-term personalization strategy directly to Google’s CEO. The core insight was that Google’s products already held tremendous user value, but in isolation: Gmail, Photos, and YouTube each knew something meaningful about you. Connecting them would unlock a kind of helpfulness no single product could offer on its own.

Personal Intelligence became the umbrella for how that personalization works — giving people the choice to connect their Google apps, history, and preferences on their own terms. Asking the model for sneaker recommendations becomes more powerful when it can draw on the purchase history from your Gmail receipts and understand which brands and styles you prefer.

Personal Intelligence launched publicly in early 2026.

Thirty days before Google I/O, senior leadership asked the team: can we ship an agentic feature? We had been incubating early tech, and I immediately put a principal designer on the problem, backed by three junior designers and a researcher, and partnered with a Senior Director PM to pressure-test what could actually ship in a month.

One research finding shaped our focus: people value offloading cognitive and emotional effort over time savings. They were also wary of AI overstepping. People want to pick their own restaurant; they don’t want to go through the cumbersome task of booking the reservation.

Speed set the other constraint. We had to be faster than doing the task manually and faster than any competing tool.

Agentic Search launched at Google I/O 2025 with tasks completing in under two minutes on average, never exceeding five. Competing tools took up to 45 minutes for equivalent requests.