Inside Google's AI turnaround: AI Mode, AI Overviews, and vision for AI-powered search | Robby Stein
Summary
Google’s AI turnaround is less a single reorganization than the visible payoff from years of compounding investment, tighter product-research collaboration, and “an incredible sense of focus and urgency.” Lenny opens with Gemini reaching No. 1 in the App Store and excitement around Nano Banana. Stein describes the broader monthly product-and-model improvements as reaching a tipping point: frontier models have become useful enough that consumers can finally feel the accumulated gains.
Stein rejects the “Google is dead” thesis because AI is expanding what people search for rather than replacing Search’s vast base of navigational, transactional, and factual jobs. Google Lens visual searches are growing 70% year over year from an already “billions and billions and billions” scale, as users photograph shoes, homework, or bookshelves and ask questions that were impractical in keyword search.
Google’s competitive AI asset is the combination of frontier models, distribution, and live structured information—not merely a chatbot embedded in Search. AI Mode can tap 50 billion products updated two billion times an hour, 250 million places in Maps, finance data, and the web; AI Overviews and Lens increasingly become previews that lead into the same conversational system. The investor-relevant claim is that Google can turn existing intent into richer queries without requiring users to form a new habit elsewhere. For a forthcoming visual version, Lenny frames the opportunity as a potential threat to Pinterest; Stein distinguishes it from Nano Banana, which is an image editor.
AI Mode’s query fanout turns one prompt into potentially dozens of background searches, pairing model reasoning with real-time information, spam detection, authority signals, and links for verification. Stein’s answer to AEO/GEO is therefore evolutionary: satisfy intent, provide original and well-sourced information, and focus on the advice, how-to, and complex questions AI is causing people to ask more often. “At the end of the day, actually something’s searching.”
AI Mode is positioned as an information product, not a general-purpose therapist, creative companion, or spreadsheet workbench. Google is betting that users will move from “keywordese” to five-sentence natural-language requests—such as finding outdoor date-night options after excluding four restaurants and accommodating an allergy—while retaining access to authoritative sources and follow-up questions.
The speed of the launch is evidence that Google can still operate like a startup when conviction is high. A five-to-10-person team began roughly a year before the conversation, found a few “moments of brilliance,” tested with about 500 outsiders who were encouraged to say when it sucked, expanded through Labs, and then launched broadly in the US. Stein believes “the next year or so of product” might establish consumer habits for many years.
Stein’s operating system combines relentless dissatisfaction with hard instrumentation: vision identifies the better world, while retention curves and root-cause analysis show whether the product is actually getting there. Teams should watch day-seven, day-30, and day-90 retention, find where an S-curve is flattening, and move investment toward new engines where individual changes can still generate 10%, 20%, or 4% wins. His counter to the “cult of lean” is that difficult breakthroughs often die because teams remain understaffed after internal conviction arrives.
Instagram Stories and Close Friends illustrate how to adopt proven formats without merely cloning them—and how long compounding improvement can take. Stories became Instagram-native through camera-roll uploads, pausing, creative tools, and a coherent placement; Close Friends took two or three years to recover from confusing design and mistranslation, then worked when lists reached roughly 20–30 people and could produce two DM replies in Stein’s example. The broader rule is “clarity instead of cleverness,” paired with enough humility to admit the first version failed.
Deep dive
1. Google’s AI momentum came from compounding, not one dramatic intervention
Lenny opens with the surprise that Gemini had reached No. 1 in the App Store, above ChatGPT, after years of questions about why Google’s research strength had not translated into a winning consumer product. Stein confirms that Nano Banana was generating excitement, while noting that users were also newly discovering powerful AI already embedded across Google’s products.
Stein cannot diagnose every organizational change across his years away from Google, but his current experience is unambiguous: “an incredible sense of focus and urgency to deliver great products quickly.” Product teams, Google DeepMind researchers, technical thinkers, and leadership are working closely enough to turn longstanding research investment into consumer experiences.
His explanation resists hero narratives and one-time turning points. Momentum comes from “every month, like ruthlessly improving the product or the models,” until the compounded gains cross a threshold where people like the product, use it more, and suddenly notice what has been building underneath.
2. AI is expanding Search rather than erasing its core jobs
Lenny preserves the bear case plainly: ChatGPT and Perplexity appeared to make result pages and link-clicking obsolete, producing the refrain that “Google is dead.” Stein’s answer is categorical about the observed product: “The core Google Search isn’t really changing, in my opinion. We’re not seeing that.”
The reason is Search’s underappreciated breadth. People still need a phone number, a price, directions, or the payment page for their taxes; AI has not removed those foundational needs, even if conversational answers now address a new class of questions.
Stein instead calls AI “expansionary”: more curiosity can be fulfilled, so people ask more questions and express needs that keyword search handled poorly. Growth comes from adding addressable queries rather than assuming every AI interaction substitutes for an old search.
Google Lens is his clearest evidence. Visual searches are growing 70% year over year at an already “billions and billions and billions” scale, with users photographing shoes to find sellers, homework to request help on question two, or a bookshelf to ask what they should read next.
3. AI Overviews, Lens, and AI Mode are converging into one Search system
Stein divides Google’s AI-search product into three components: AI Overviews provides a quick answer atop ordinary results; Lens handles visual and multimodal searches; AI Mode combines those capabilities in an end-to-end conversational experience built on frontier models and designed specifically for Search.
The underlying information base is central to the pitch. Google’s Shopping Graph contains 50 billion products and receives two billion merchant updates an hour, while Maps contains 250 million places; AI Mode can also draw on finance information, web context, and links that let users investigate further.
Integration is already reducing the need to choose a surface. A difficult natural-language query can produce an AI Overview and then continue in AI Mode; a Lens photograph can provide an initial interpretation and move into the same follow-up conversation.
Stein’s eventual design goal is that people “shouldn’t have to think about where you’re asking a question.” Google launched the explicit google.com/ai entry point because the product was new and needed focused feedback, but the intended experience is coherent across text, camera, quick answers, and deeper dialogue.
4. Google is betting that natural language unlocks its installed intent
Lenny invokes Alex Rampell’s framing that startups win by securing distribution before incumbents innovate fast enough, then suggests Google has reached the “now here comes Google” moment. Stein’s response is that users were already asking Google for these outcomes; the missing piece was AI capable of answering the chemistry image or hard calculation they submitted.
The behavioral hurdle is teaching people that Google no longer requires “keywordese.” Users can enter a five-sentence request about date-night restaurants, exclude four places they have visited, specify outdoor dining, and include a friend’s allergy—the kind of query they historically would not imagine typing into Search.
Ask Jeeves, Lenny observes, was “surprisingly prescient”: it had the idea of asking a question as if speaking to a human before the technology was ready. Stein agrees that the idea was simply ahead of its time.
5. AI Mode is specialized for information, not every chatbot use case
Stein defines AI Mode as “a way to ask Search anything you want,” optimized for trip planning, shopping, research, and other needs grounded in information. Its intended advantages are context, current data, links, and the ability to verify an answer against authoritative sources.
It can still rewrite text, but Google is not principally targeting creative companionship, productivity workflows, or tasks such as uploading a spreadsheet and producing charts. Lenny’s shorthand captures the distinction: “AI Mode is not your therapist.”
That narrower positioning also shapes interface and model design. People may occasionally greet it conversationally, but observed usage centers on learning or completing an informational task, so Google is optimizing the experience around effortless retrieval and deeper exploration rather than a general chatbot persona.
6. Query fanout keeps web discovery inside the AI response loop
When constructing an answer, AI Mode performs “query fanout”: the model can append dozens of related searches to the user’s prompt, issue them against Google’s systems, retrieve current data when necessary, and assemble the response from the resulting material. “At the end of the day, something’s searching. It’s not a person, but there are searches happening.”
Stein presents that capability as distinctive because the model can combine parametric memory and reasoning with Google Search’s machinery for identifying spam, evaluating authority, checking work, and deciding when to link directly to a particularly useful source.
His AEO/GEO advice begins with traditional quality signals, not a new bag of tricks: satisfy the user’s intent, cite sources, contribute original information, and avoid repeating what has already been repeated “500 times.” Google’s human-rater guidelines remain relevant because the AI is still researching and selecting information.
The genuine change is demand. Creators should study what people now ask AI—especially advice, how-to questions, and complicated needs—and build the best material for those expanding categories, rather than treating AI visibility as detached from usefulness.
7. Natural-language steering is lowering the cost of building AI products
Stein’s freshest lesson is how quickly AI interfaces have become human-like. Only months earlier, users needed prompting “incantations,” role instructions, or heavy post-training to make a model behave reliably; increasingly, they can describe the desired behavior almost as they would brief an engineer.
A startup might now give a model its internal data, API documentation, schema, URL, usage conditions, and warnings about questions requiring extra care. The model can infer when to spend more reasoning budget, call a tool, or execute code, without every behavior being encoded through weight updates.
The implication is broader access to sophisticated product development: “increasingly, I don’t think you need to do a lot of this heavy-duty fine-tuning.” Lenny connects that to designing AI against the standard of a human interaction, not merely against the previous generation of software interfaces.
8. Visual conversation is the next expansion beyond text chat
Stein argues that AI “was born and grew up in this text modality,” so even a visual request such as redecorating a bookshelf traditionally produced prose. Multimodal systems can instead help directly with inspiration, shopping, and other needs where seeing possibilities matters more than reading a description.
A visual version of AI Mode, announced at I/O and in the process of rolling out, can produce an inspirational image board for a “mid-century modern” office with dark themes. The user could then request something lighter, creamier, more Californian, or more coastal, with the system understanding both the images and the conversational refinement.
Lenny frames this kind of visual inspiration as a potential threat to Pinterest and guesses it is Nano Banana inside AI Mode. Stein corrects the mechanism: Nano Banana is an image editor, while this experience searches for images on the web and lets users converse with visual results—though editing a photograph of one’s living room might eventually be complementary.
9. Relentless improvement begins with productive dissatisfaction
Stein defines his central product trait as the physical embodiment of two ideas: “complete effort” continuously directed toward positive productivity, and the refusal to stop making things better. Relentlessness without improvement is insufficient; improvement without sustained pressure rarely reaches the tipping point.
The phrase began when his wife answered an icebreaker asking for one word to describe him: “Dissatisfied.” Her explanation transformed the apparent insult—he was not simply unhappy, but unwilling to accept what the world supplied when he believed it could be better.
Tony Fadell’s fruit-sticker story is Stein’s model of this sensibility: the sticker punctures the peach, the fruit “bleeds,” the discarded label misses the bin, and the customer bends down to retrieve it. Most adults habituate to such friction; strong product thinkers keep asking, “Why am I tolerating this?”
AI Mode emerged from the same irritation. Google saw users appending the word “AI” to queries in hopes of forcing an AI response, while harder questions often received no AI Overview at all. The team’s reaction was, “This is ridiculous,” followed by the larger question: why could Google not do this for everything?
10. Metrics are instruments, not substitutes for product judgment
Stein rejects a choice between making something better and driving KPIs. The process starts with a problem—or its inverse, a vision of a better state—then uses instrumentation to determine whether the built product actually changes behavior.
For a young product, he might look at the retention “J curve”: what percentage remains on day seven, day 30, and day 90, and whether the curve flattens or keeps draining toward zero. Surviving that gate precedes growth, word of mouth, and the question of whether the opportunity can become large.
For a mature product whose core metric falls 5% in a week, metrics locate the disease: region, device, demographic, or use case. Root-cause analysis identifies where treatment is needed, but the dashboard cannot prescribe the fix; “You have to think for yourself how to make it better.”
Resource allocation follows S-curves and diminishing marginal returns. If another 50 people will barely move an established feature, the team needs a new growth driver; once that driver produces changes worth 10%, 20%, or 4%, it deserves more investment. Otherwise, teams risk “congratulating ourselves” without evidence anyone cares.
11. Stories succeeded by treating a competitor’s invention as a format
Instagram’s existential question was not whether it had invented ephemeral vertical sharing, but whether that format better served its core job of sharing life and connecting people. Stories lowered posting pressure, removed likes, disappeared, and fit mobile screens; Stein gives Snapchat explicit credit for inventing a genuinely strong format.
Early attempts to make Instagram’s established feed ephemeral failed because they contorted a product users already understood. The team instead built a distinct surface that fit Instagram’s system while adding its own creative tools, sophisticated filters, neon drawing, and high-resolution camera-roll uploads.
Small acts of dissatisfaction mattered. Snapchat did not then allow uploaded photos, and its Stories could not be paused; Instagram let users preserve a camera-roll memory and hold a finger down when content moved too quickly. Those details made the feature feel native rather than mechanically copied.
Lenny retains the founders’ objection that Instagram “stole” Snapchat’s idea. Stein’s rebuttal is user-centered: feeds and Stories became product primitives, just as many products adopted feeds after Facebook; refusing a useful format can mean “robbing your user base of the opportunity to have a better product.”
12. Mature products grow by adding coherent but distinct new primitives
Stein approaches large products with humility: product is like golf because “you’re always one stroke away from shanking.” He first asks why people hire the product, which parts are growing, mature, or declining, and how user needs have shifted—for Instagram, from public feed broadcasts toward Stories, DMs, and private sharing.
Jobs-to-be-done analysis prevents teams from assuming the current interface must supply the next answer. Instagram did not need a square photo that did more; Google did not need only another subtle results-page tweak. Both needed a first-principles response to the user’s underlying job.
New formats should be complementary, coherent, and visibly different. A disappearing item hidden inside an ordinary feed slot would violate spatial expectations; Stories therefore occupied its own recognizable row, while AI Mode uses a full-page conversational experience that remains connected to core Search.
Stein warns against importing another company’s successful feature unchanged. Its users, context, and expectations may be completely different: the task is to learn what the external format proves, then rebuild it for the essence and conventions of one’s own product.
13. AI Mode’s year-long build shows when small teams should scale
AI Overviews supplied the starting signal: people were already asking natural-language questions, wanted more direct access, and needed follow-ups that did not fit comfortably inside the traditional results page. A small team created a blank-screen prototype with search, reasoning, multi-turn memory, and a more powerful version of the AI behind Overviews.
The initial group was roughly five to 10 people, formed around the previous summer and fall. An early version was poor overall but produced “moments of brilliance,” including a response for an outing with Stein’s daughter that combined park details, useful links, Maps information, and walkability.
Before Labs, about 500 external trusted testers—including friends and family—were encouraged to report every breakage and nonsensical answer. Once their feedback improved, Google opened Labs for larger-scale query data, launched to users in the US, and began expanding across countries and languages.
Stein’s resourcing lesson cuts against the “cult of lean.” Small teams are useful until internal conviction, but technically difficult products often “die on the vine” because staffing stays minimal too long; even Close Friends’ slow iteration reflected underinvestment. After validation, leaders should fund the group required to build a genuinely strong external product.
14. Clarity, causation, and humility rescued Close Friends
Stein’s hypothetical product book has three main chapters: deeply understand people, apply analytical rigor to the problem, and “design for clarity instead of cleverness.” Humility is the coda—question yourself, listen to users, and remain open to being wrong.
His jobs-to-be-done test is not merely how someone uses a product but why they first “hire” it. He favors interrogating the causal moment—where the person was, what they were doing, and what triggered the decision—because the “big hire” reveals more than a feature wishlist.
Close Friends initially mixed private Feed and Story posts, added a special profile, and used inconsistent green indicators. Worse, Stein thinks it may originally have been called “Favorites”; in some markets, the term was mistranslated as “best friend,” leading users to add one person. That person rarely saw and answered the post, so the emotional job of connection failed.
Data showed the loop working with 20–30 people: perhaps two would reply by DM, creating the desired feeling of connection. Over two or three years, Instagram limited the feature to Stories, renamed it Close Friends, recommended list members, and exposed the green ring outside the Story. Finsta behavior also informed the team’s understanding of the need for smaller-group sharing.
15. Curiosity closes the loop between AI assistance and original sources
Stein’s parting principle is “be curious”: keep asking why a product fails, why another person disagrees, and why the world works as it does. AI is an “ultimate curiosity engine,” but he pairs it with old papers, freely available PDFs, books, and original sources rather than relying only on summaries.
Search Live, which had moved out of Labs that week, makes AI Mode available as a full-screen voice conversation in the Google app. Stein’s young children ask to “talk to Google” about animals, history, or school topics, an experience he believes is making them naturally AI-native.
His Stamped story supplies the action-oriented version of curiosity. At 25, facing a cold-start problem for a recommendations app, Stein and his co-founder emailed Scooter Braun, claimed they would be in Los Angeles the next day, received a breakfast invitation, and immediately flew from New York.
Braun offered to help and possibly advise; Stein and his co-founder then met Justin Bieber, who used the product to recommend favorite songs and other things. Bieber’s participation helped attract users. Stein’s lesson is not celebrity strategy so much as tempo: “Do it now, be scrappy, be immediate. Intense urgency usually wins over thinking about it for a long time.”