Search Isn't Dying, It's Expanding: Robby Stein on Google's AI Mode and Building Products Relentlessly

Host Lenny Rachitsky opens by noting that Gemini had just hit number one in the App Store, above ChatGPT, something few people predicted. Robby Stein, VP of Product for Google Search, says he can't point to one reorg or single leader behind the change. What he feels now, in his second stint at Google after starting in 2007, is a strong sense of focus and urgency to ship great products quickly.

Stein credits close collaboration between product teams and researchers at Google DeepMind, plus a group of technical thinkers who have worked on these problems for years. In his view people overattribute momentum to one-time changes. He describes instead a compounding effect: ruthlessly improving the product and models every month until it tips and people simply like using it more.

Asked whether Nano Banana drove the surge, he says people are excited about it but are also discovering how powerful the wider set of Google tools has become. He is regularly surprised that features sitting right inside core search are news to people on X.

Against the "Google is dead" narrative, Stein says people come to search for a ridiculously wide range of needs, from phone numbers and prices to directions and tax payment pages, and that breadth is underappreciated. Those foundational needs haven't changed. What AI adds is a wave of new questions and curiosity that can now be satisfied.

He points to Google Lens, which is growing about 70% year over year in visual searches at a scale of billions. People photograph shoes to find where to buy them, a homework problem they're stuck on, or a bookshelf to get book recommendations. He frames the picture as three connected pieces: AI Overviews for quick answers, multimodal search with Lens, and AI Mode, which ties everything into an end-to-end conversational experience.

AI Mode taps into Google's rich data, including a shopping graph of 50 billion products updated 2 billion times an hour and 250 million places in Maps. Stein says the goal is that users shouldn't have to think about where to ask. Hard multi-sentence questions already trigger AI previews that lead into AI Mode, and the entry point at google.com/ai exists mainly to gather feedback on a new product.

The big behavioral shift, he says, is natural language. People used to type "keywordese"; now they can ask long, complex questions, like finding an outdoor-dining date-night spot while avoiding four restaurants already tried and accounting for a friend's allergy.

On the rise of AEO and GEO, Stein explains what happens under the hood. When Google's AI builds a response it uses "query fan-out," treating search as a tool and issuing possibly dozens of background queries, each paired with content. So the question for publishers is whether their page is genuinely helpful for the query being run.

He points creators to Google's long-standing human rater guidelines: satisfy user intent, cite sources, and offer original information rather than something repeated 500 times. Those signals remain valid because an AI is still doing research and looking for good information. His one new piece of advice is to think about what people use AI for, which increasingly means advice, how-to and more complex needs, and to make content best for those.

Stein says this design is unique to Google's AI: the model is built for informational tasks, with access to search signals such as spam detection and authority, and it checks its own work. It links to the sources so users can verify and dig in.

Asked what he has learned about building AI products, Stein says the most recent lesson is how human-like the interface has become. Only months ago you needed prompt incantations, hacks like "act like a coach," or post-training that updated model weights to get sophisticated behavior.

Now, he says, you can write instructions much like a brief to a new engineer: here is our data, here are the APIs and schema, here is when to use them, and be careful on this type of question. Models have been trained to allocate more reasoning budget or use code execution and tools when needed. Stein sees this as democratizing access, since the most sophisticated outcomes increasingly don't require heavy fine-tuning.

Stein's core message for builders is "embodying relentless improvement": relentless effort directed at productive ends, plus never being content and always making things better. He traces it to an icebreaker at an Instagram all-hands where he texted his wife for one word to describe himself and she replied "dissatisfied." She then clarified it wasn't unhappiness but a deep drive to make the world better.

For the products that succeeded, he says, there was always a spirit of making two more moves until the product tipped into being net useful. He also cites Tony Fadell's TED talk on thinking younger, with the example of a peeled fruit sticker, which argues that adults habituate to annoyances and the best product people refuse to.

AI Mode itself grew from that mindset. The team saw AI Overviews failing on harder questions and users appending the word "AI" to queries to force it. They decided to be disgruntled on the user's behalf and build a whole new experience, even though that was hard.

On the tension between craft and KPIs, Stein says the two aren't opposed. Start from a problem or vision, then use instrumentation: retention J-curves at day 7, 30 and 90, whether people talk about and share the product, and how big it can get. On mature products, use metrics for root-cause analysis, checking whether a 5% drop sits in a region, device, demographic or use case, then fix it.

Reflecting on launching Instagram Stories, Stein says the lesson is to understand why people use your product and recognize when an alternative format is an existential question. Not every great idea will be invented by you. Instagram is about sharing life and connecting with people, and an ephemeral, full-screen, no-likes format lowered the pressure, so the team credited Snapchat for inventing it and focused on making it Instagram's own.

Earlier attempts to make the main feed ephemeral failed, he says, because contorting a cemented core product is usually a bad recipe. The Stories team instead added neon drawing and filters, camera-roll uploads that Snapchat lacked, and the ability to pause a story with your finger. Responding to the backlash, he argues formats like feeds become core primitives, and refusing to adopt one just robs users of a better product.

On stepping into mature products like Instagram or Google, Stein advises humility, calling product like golf: you're always one stroke from shanking. Find which parts are growing, and use the jobs-to-be-done lens, citing Clayton Christensen's *Competing Against Luck*, to move beyond incrementally tweaking the current tool. New formats should complement rather than replace the old, feel coherent but distinct, and respect user expectations, because people think spatially about interfaces.

To balance optimizing versus big bets, he tracks growth curves by feature and watches for diminishing marginal returns. When effort stops moving the needle, usually because something fundamental changed, it's time to find a new growth driver and then optimize it with steady 4%, 10% and 20% wins.

Stein says AI Mode grew out of AI Overviews, where Google saw people asking natural-language questions and wanting follow-ups, something hard to build into the classic results page. A very small team, perhaps five to ten people including technical leaders and designers, started with a blank page and a blinking cursor, wiring into the same AI but adding search, reasoning and multi-turn context.

The early version, built around last summer, wasn't good but had moments of brilliance. Stein recalls asking about planning an outing with his daughter and getting park information, links and walkable Google Maps details. That experience gave the team conviction, he says, and the project was entrepreneurial in that someone had to build it and feel it.

The rollout progressed from about 500 trusted testers, including friends and family who sent screenshots of what broke, to an opt-in Labs release that supplied real query data, to a US launch and expansion across countries and languages. Stein says urgency came from users not getting answers Google could now provide, and a belief that the next year of products will set habits for years.

On positioning against ChatGPT, Claude or Perplexity, Stein says AI Mode is designed for information: planning trips, shopping and research, with links to authoritative sources. It is less focused on creativity or productivity tasks like turning spreadsheets into graphs, though it will rewrite text if asked.

If Stein wrote a book on great products, he says it would have three chapters. The first is to deeply understand people, using Christensen's idea that users "hire" a product to do a job. He quotes Theodore Levitt's line that people want a quarter-inch hole, not a drill, and describes interviewing users about the moment they first hired a product.

The second is analytical rigor about problems. Close Friends initially failed, he says, because the data showed people added only one friend: the feature had been mistranslated as "best friend" in many markets, so the odds of getting a reply were near zero. The third is clarity over cleverness. Using Don Norman's door example, he argues that standard patterns beat reinvention, which is why a camera icon should just be a camera. A coda: be humble and open to being wrong.

Rachitsky notes the name AI Mode is itself an example of clarity, and Stein agrees the team discussed this internally. The host jokes that Nano Banana is the opposite; Stein concedes no advice is universally right.

Stein tells the story of Close Friends, the private-list feature behind the green circle, which took two or three years to work. The first version spanned feed posts, stories and profiles, which was confusing: a vulnerable, blurry moment would appear right after a polished feed photo, and the green marker behaved inconsistently across surfaces.

Applying his framework, the team recognized the real job was emotional: share something vulnerable and get a DM back that makes you feel connected. With only two people on a list the product was broken, but with 20 to 30 people a couple would reply and close the loop. They renamed it Close Friends, built a list builder with recommended people, limited it to Stories, and put the green ring on the outside of the story tray for clarity.

The investment was justified by research: when asked why they didn't post to Stories, people said an ex, a teacher or a judgy friend was watching, an audience problem. Stein says finstas and similar hacks showed users were trying to create smaller private groups. Features were tested in a basket of countries, often Australia or Canada, because it is easier for teams to get feedback there.

Stein pushes back on the "cult of lean" popular on Twitter. Small teams make sense for building internal conviction, but he says a product based on a technological breakthrough often needs real investment. He points to the years and hundreds of people behind foundational models, and says inside big companies he sees projects stay too scrappy, never get momentum and die on the vine.

Close Friends again serves as an example: the team stayed so small that the loop cycle was slow, which would have killed a startup. He suggests thinking from first principles about the group you need to build a great version rather than blindly staying at two people until market fit.

On when to scale up, Stein says it is when you hit conviction: you believe in it, backed by external validation. He cautions that friends won't use a product daily at day 30, 60 or 90 unless it is useful. Shipping a broken version wouldn't win, so invest enough to put out the best version you can.

In the AI Corner, Stein says AI is moving beyond its text origins into visual and inspirational needs such as shopping and design. He previews a visual version of AI Mode, announced at I/O and rolling out, where you could ask for a mid-century modern office with dark themes, get an image board, and refine it by asking for a lighter, more coastal feel. He distinguishes it from Nano Banana, which edits images, saying this is more like AI image search with natural-language turns.

His closing theme is curiosity: wanting to know why people do things, why someone disagrees, why something isn't working. He calls AI the ultimate curiosity engine but recommends blending it with original sources such as old papers and books. For his kids, the best AI-native tool has been Search Live, a voice mode in the Google app that they use to ask about animals or school history.

In the lightning round he recommends *Competing Against Luck*, *The Design of Everyday Things*, and the novel *Aurora*, and tells the story of his earlier startup Stamped. At 25, he and his co-founder emailed Justin Bieber's manager and flew to LA the next day for a meeting, which brought a celebrity user to the app. His lesson: be scrappy and immediate, because urgency usually beats overthinking.