How Google rebuilt search with AI: lessons from Robby Stein

Executive overview

Google search wasn't dying — but it was failing users who wanted more than keyword lookups. The fix wasn't a chatbot replacement; it was an expansion layer: AI Overviews, AI Mode, and multimodal search that lets people ask anything, in natural language, with the full depth of Google's data behind it.

Robby Stein, VP of Product for Google Search, built AI Mode in under a year by applying the same principles he used at Instagram — start small, find the "moments of brilliance," add conviction, then scale. The through-line across every product he's built: relentless improvement driven by dissatisfaction with what currently exists.

AI is expansionary, not a replacement

• Core search queries (phone numbers, prices, directions) are unchanged by AI
• Growth comes from new question categories AI makes possible — not cannibalising existing ones
• Google Lens visual search is growing 70% year over year at already massive scale
• Users were already typing "AI" at the end of queries, signalling unmet demand

The three layers of Google's AI search

• AI Overviews — fast AI answers at the top of standard search results
• AI Mode — end-to-end frontier search experience with multi-turn conversation, accessible at google.com/ai
• Multimodal / Lens — camera-first search; photos route into AI Mode for follow-up questions
• All three are converging: AI Mode now powers the depth layer behind Overviews and Lens

What makes AI Mode different from chatbots

• Designed specifically for informational needs, not creativity or productivity tasks
• Performs "query fan-out": model issues dozens of background searches rather than relying on parametric memory
• Taps 50 billion products in Google Shopping Graph (updated 2 billion times/hour), 250 million Maps places, and live financial data
• Less focused on being a conversational companion; focused on returning authoritative, linkable sources

How AI Mode was built in under a year

• Started with 5–10 people, a blank canvas, and the question: what if you could ask Google anything?
• First milestone was purely qualitative — a personal "moment of brilliance" using the prototype
• ~500 external trusted testers before Labs; friends messaged screenshots of what was broken
• Launched in Labs to get real query data at scale; tuned iteratively before US-wide rollout
• Speed came from urgency: "the next year of product will shape how people use AI for many years"

SEO and AEO in an AI search world

• AI constructs responses through query fan-out, so underlying search quality signals still apply
• Helpful, original, well-sourced content that satisfies user intent remains the right target
• Focus content on what people use AI for: advice, how-to, complex multi-part questions
• Core Google quality guidelines ("human rater guidelines") are still the most reliable reference

Embodying relentless improvement

• The best product people never habituate to what's broken — they ask "why does this exist?"
• Tony Fadell's fruit sticker story: the best product thinkers notice irritations others accept
• AI Mode was motivated by watching users type "AI" at the end of queries as a workaround
• Dissatisfaction with your own work is the engine — you are your own harshest critic

How to grow a mature product

• Use jobs to be done (Clayton Christensen, Competing Against Luck): why is someone "hiring" your product?
• Map growth vs. mature vs. declining features; invest where marginal return is still high
• New formats (Stories, AI Mode) complement rather than replace core product — they fill a different need
• Track J-curves (retention at day 7, 30, 90) to know if a new thing is working or dying
• Add new things carefully: they must feel coherent but clearly distinct from existing primitives

The Instagram Stories lessons

• Snapchat invented a great format; Instagram made it fit Instagram's users and expectations
• Key differentiators: allowing camera-roll uploads, adding a pause gesture, different creative tools
• Trying to contort a cemented core product is "usually a bad recipe" — build something new instead
• Formats become primitives (feeds, stories); not building them robs your users of a better product

Close Friends: a three-year failure turned success

• Original launch failed: "close friend" was mistranslated as "best friend" in many markets → users added one person → zero engagement loop
• Also failed because it mixed ephemeral content into the permanent feed, which felt wrong
• Fix: renamed to "Close Friends," added an algorithmic list builder suggesting ~20–30 people, moved the green ring to the outside of the story circle
• Key insight: the emotional job was connection — getting a DM back — not just private sharing
• Lists of 20–30 people meant enough views that users reliably got replies, closing the emotional loop

Three chapters for building great products

1. Deeply understand people — study causation, not just behaviour; Clayton Christensen's "big hire" interview technique; identify both utility jobs and emotional jobs
2. Analytical rigour — root-cause analysis when metrics drop; find the regional, device, or demographic split before treating
3. Design for clarity, not cleverness — Don Norman's Design of Everyday Things; lean into global conventions; don't reinvent icons when a camera is a camera

On team size and when to scale

• "Cult of lean" leads teams to give up too early or let products die on the vine
• Two milestones: (1) internal conviction from early qualitative results; (2) a version good enough to ship
• Close Friends took too long partly because the team stayed too small
• Scale up after conviction is established, not before; but don't stay scrappy forever

What's next for AI search

• Search Live: voice-first conversational AI mode, now out of Labs in the Google app
• Visual AI Mode: image boards for inspirational and shopping tasks; multi-turn with images (announced at Google I/O)
• Shift from text-only AI toward multimodal — visual inspiration, shopping, and discovery use cases