Robby Stein
VP of Product, Google Search (formerly head of product at Instagram)
Where to hear and read about Robby Stein1 episode or test
Robby Stein is VP of Product for Google Search, running AI Overviews, AI Mode, Lens and ranking, after leading Stories, Reels and Close Friends as head of product at Instagram. The thing most worth learning from him is to find the real job people hire a product for, diagnose failures down to the root cause, and keep improving until the product tips.
Don't think of users as using your product. Think of users as hiring you to do something for them.
Inside Google's AI turnaround: AI Mode, AI Overviews, and vision for AI-powered search | Robby Stein
Robby Stein: the story and the lessons
1Robby Stein's path from Stamped to Google Search
Stein moved fast at 25 to land Justin Bieber on Stamped. He and his co-founder emailed Scooter Braun, Bieber's manager, saying they'd be in LA tomorrow, though they weren't. Braun replied with a breakfast spot and they flew out that day.
Bieber then used Stamped to post his favorite things, which drew many people to try the app. Stamped let people stamp favorite things and get recommendations from trusted people. His takeaway: "intense urgency usually wins over thinking about it for a long time".
He started two companies and joined the Artifact founding team. Before his big-company roles, Stein started two companies of his own and was on the founding team of Artifact with Mike Krieger and Kevin Systrom.
At Instagram he led Stories, Reels and Close Friends. As head of product at Instagram, he saw the app grow to half a billion daily active users during his time there.
Stories shipped between 2016 and 2021, an era that also produced Reels, direct messaging updates and feed ranking.
He now runs the whole Google Search experience. Stein is VP of Product for Search, covering AI Overviews, AI Mode, Google Lens and the ranking algorithm.
This is his second stint at Google. He first joined in 2007 and says he cannot speak for the years between.
His wife describes him in one word: dissatisfied. She meant he wants the world to be better. Stein says the products that worked had a spirit of two more moves to make them better, until they tipped into being useful.
He cites Tony Fadell's TED talk: adults habituate to annoyances, but the best product people keep asking why. His example is the fruit sticker that punctures the flesh when you peel it, then misses the bin when you flick it.
2How Stein built AI Mode at Google
Stein's team built AI Mode because users were already asking for it. The team saw that AI Overviews often didn't show up for harder questions, and that people were typing 'AI' at the end of their queries to get an AI answer.
Being 'disgruntled on behalf of the user' pushed them to build a whole new experience instead of patching the old one.
AI Mode started as a 5 to 10 person prototype. The first team was a few technical leaders and a couple of designers, starting from a blank screen with a blinking cursor. The model could search, reason and keep context across turns.
The work began about a year before the interview and ended in a US launch at I/O. Stein's rule: "you got to build something and then you have to feel it yourself".
One brilliant answer from a rough prototype gave the team conviction. A question about an outing with his daughter returned park details, site links and walkable Google Maps information.
Stein compares it to hitting one perfect golf shot. Leaders across the company backed the project, but the decision still came from a small team building something and feeling it themselves.
He staged the launch with testers who tell the truth. AI Mode went to about 500 external trusted testers, including friends and family who sent screenshots of what broke. Then Labs opened it to anyone, and real query data was used to tune it.
Only then did it launch to everyone in the US at I/O. One friend loved it and hated it, and sent constant 'this broke' screenshots. Stein: "you got to have people test it that tell you the truth and tell you when it sucks".
AI Mode works because it sits on top of Google Search. It runs query fanout, firing dozens of background searches and pairing each with web content. It uses Google's spam and quality signals plus live data: 50 billion shopping products and 250 million places.
Stein wants AI Overviews, Lens and AI Mode to blend into one experience. Follow-ups on an Overview open AI Mode, and so does a Lens photo.
3Why Stein says AI expands Google Search instead of replacing it
Old search needs stay while AI adds new questions. People still look for phone numbers, prices, directions and tax payment pages. AI adds questions on top, so core search is not shrinking.
Google Lens visual searches grow 70% year over year and are already at billions of searches. Examples: photograph shoes and ask where to buy them, or photograph a bookshelf and ask which books to get next.
He credits focus and urgency, not one reorg or one leader. Stein says people often credit momentum to a one-time change or a single person. He points instead to leadership, close work with DeepMind, and researchers who had thought about this for years.
Each month the products and models improve until usage tips over. Gemini reaching number one in the app store, above ChatGPT, was one sign. He credits 'many many more people' than himself.
AI Mode targets information needs, not general chat. Typical uses are planning a trip, buying something or researching a question. Answers come with links so people can verify them and reach authoritative sources.
It is less focused on creativity or productivity, such as uploading a spreadsheet to get graphs. Most people come to learn something.
He tells creators to write for the need behind the question. Google's human rater guidelines still decide who shows up: does the page satisfy intent, cite sources, and say something original rather than what has been said 500 times.
His test is what content someone is using AI for, and how yours can be the best for it. People now ask AI more about advice, how-to and complex needs than simple lookups.
Stein sees the next year as setting AI habits for years. He feels an obligation to give users an AI-powered Google that draws on everything Google knows.
He calls it a healthy moment to build quickly, driven by a gap between what users need and what Search could answer.
4How Stein brought Stories into Instagram
Stein tied Stories to why people use Instagram. Instagram exists to share your life and connect with people. Stories did that with less pressure: no likes, ephemeral, full-screen and mobile.
He gives Snapchat credit for inventing the format. His view: "Not every great thing is going to be invented by you".
Stories got their own place instead of changing the feed. Early versions tried to make the main feed ephemeral, and that failed because the feed looks and works a specific way in people's minds.
A teammate said Stories felt like the story holes at the top that completed the product. Users think spatially, so a tile that vanishes or can't be liked would confuse them.
He made the borrowed format feel like Instagram. Instagram added neon drawing, sophisticated filters, camera roll upload and a pause, which Snapchat's version lacked.
Stein warns people fail most by assuming what works in one system works in theirs. In his words: "you never just want to snap in something that's working you have to make it work for your users".
New formats should complement the product. Stories did not replace Instagram, and AI does not replace Google. Both expanded the product, with a coherent look but a different primitive.
Stories share the aesthetic, camera roll and DMs. Google's AI is a full-page experience with follow-up conversation.
5How Stein turned Close Friends from flop to hit
Stein found the root cause in a mistranslation. People added only one person to the list because it was mistranslated as 'best friend' in many markets. With one person, a reply was close to impossible.
With 20 to 30 people on the list, about two usually reply, which closes the loop. The team redesigned around larger lists, and the name was also originally 'favorites', which nudged people to add only two or three.
He treated Close Friends as an emotional job. The real job was feeling connected, so success meant getting a DM back. Christensen separates utility jobs from emotional jobs, and Stein says teams often discount the emotional ones.
The first version spread across feed posts, Stories and a profile, with green markers appearing inconsistently. The rebuild worked only in Stories.
Asking why people didn't post gave the team conviction. Users said an ex, a teacher or a judgy friend could see their Stories. That shared audience problem kept the team going.
Finstas, where people hacked Instagram into smaller private groups, confirmed the demand. The project took two to three years, one of the longest at Instagram. Tests ran in a basket of countries, with Australia among the first.
He chose clarity over cleverness with the green ring. Close Friends only became clear when the green ring went on the outside of the story, instead of a secret story that revealed itself when opened.
He cites Don Norman's glass doors with symmetric handles that don't say push or pull. The same logic is why the plain name AI Mode works.
He thinks Close Friends stayed too small too long. Close Friends was slow partly because the team stayed so small, with slow loop cycles. Stein says "people hold on to small teams too long".
His rule: add people once you have internal conviction and outside validation. Friends use a product to be kind, but by day 30, 60 or 90 they stop unless it is useful.
6How Stein decides what to measure and where to invest
He reads retention curves for young products. Check the share still using it at day 7, 30 and 90. If the curve never flattens, in his words the product is toast.
After retention come two more gates: do people talk about it, and how big can it get. No product starts big. Even fast-growing ones began internally with 100 to 200 people.
He dissects a 5% drop by region, device and use case. In a big product, a 5% drop in a core metric calls for root cause analysis by region, device, demographic and use case.
He compares metrics to a pilot's instruments: they show whether the plane is flying, but not what to do next. A problem gives you the vision.
Stein tracks S-curves and moves when returns approach zero. When 50 more people on a project still would not move the needle, he looks for a new growth driver. That plateau usually coincides with changing expectations or market saturation.
Once a new engine lands, he staffs it up, since each change might be a 4%, 10% or 20% win. Without a quantified goal, teams congratulate themselves on improvements nobody uses.
He asks why people hired the product, not how to improve it. Drawing on Christensen's Competing Against Luck, Stein interviews users about the moment before first use: where they were, who they talked to, what prompted it.
He calls it lowercase jobs to be done, not the rigorous framework. The incremental Instagram move would have been making the square photo do more.
He stays humble about growth in mature products. Stein compares product work to golf: you're always one stroke away from shanking. Inside a big product some parts grow, some are mature and some decline.
On Instagram, sharing shifted from large public feed posts to lightweight formats like Stories and private DMs.
7Stein on building with AI and staying curious
He gives AI models context in plain language, not fine-tuning. Models can now follow an order like a new engineer: describe your data, the API, the schema, the URL, when to use it and what to get right.
The model then decides whether to spend more reasoning budget or run code. He says "I don't think you need to do a lot of this heavy duty fine-tuning", and expects this to democratize who can build.
He wants AI to feel like talking to a person. AI Mode is built for back-and-forth conversation, and Search Live adds a voice version that just left Labs. His kids use it to ask about animals and history topics from school.
He sees natural language replacing prompt hacks, saying it is 'so obvious how humanlike the interface is becoming'.
He sees the next step for AI search as visual. A visual version of AI Mode lets people refine image results by talking to it, such as asking for a mid-century modern office with dark themes, then 'lighter, creamy, more California coastal'.
It finds images on the web, like AI image search, and is not Nano Banana, which is an image editor. Lens and image search growth shows the demand.
He keeps a playful culture alongside his clarity rule. The Nano Banana team was a scrappy, fun group that chose a fun name. Stein says it feels like his first stretch at Google.
The name goes against his advice on clarity, but he says no advice is right universally.
Curiosity is his motto, and user feedback is what he asks for. His motto is to be curious and not assume you have all the answers. He reads old free papers, for example on statistics, uses AI to find them, and blends AI, books and primary sources.
He almost named a company Curious. He is on X at rmstein and asks for problems with Google products and AI.
Inside Google's AI turnaround: AI Mode, AI Overviews, and vision for AI-powered search | Robby Stein
Google Search's VP of Product explains how AI Mode went from a 5 to 10 person prototype to a launched product in about a year, and why AI is expanding search rather than replacing it. He also shares the product principles behind Instagram Stories and Close Friends, including how Close Friends went from flop to hit over two to three years.
Stein moved fast at 25 to land Justin Bieber on Stamped. He and his co-founder emailed Scooter Braun, Bieber's manager, saying they'd be in LA tomorrow, though they weren't. Braun replied with a breakfast spot and they flew out that day.
Bieber then used Stamped to post his favorite things, which drew many people to try the app. Stamped let people stamp favorite things and get recommendations from trusted people. His takeaway: "intense urgency usually wins over thinking about it for a long time".
He started two companies and joined the Artifact founding team. Before his big-company roles, Stein started two companies of his own and was on the founding team of Artifact with Mike Krieger and Kevin Systrom.
At Instagram he led Stories, Reels and Close Friends. As head of product at Instagram, he saw the app grow to half a billion daily active users during his time there.
Stories shipped between 2016 and 2021, an era that also produced Reels, direct messaging updates and feed ranking.
He now runs the whole Google Search experience. Stein is VP of Product for Search, covering AI Overviews, AI Mode, Google Lens and the ranking algorithm.
This is his second stint at Google. He first joined in 2007 and says he cannot speak for the years between.
His wife describes him in one word: dissatisfied. She meant he wants the world to be better. Stein says the products that worked had a spirit of two more moves to make them better, until they tipped into being useful.
He cites Tony Fadell's TED talk: adults habituate to annoyances, but the best product people keep asking why. His example is the fruit sticker that punctures the flesh when you peel it, then misses the bin when you flick it.
Stein's team built AI Mode because users were already asking for it. The team saw that AI Overviews often didn't show up for harder questions, and that people were typing 'AI' at the end of their queries to get an AI answer.
Being 'disgruntled on behalf of the user' pushed them to build a whole new experience instead of patching the old one.
AI Mode started as a 5 to 10 person prototype. The first team was a few technical leaders and a couple of designers, starting from a blank screen with a blinking cursor. The model could search, reason and keep context across turns.
The work began about a year before the interview and ended in a US launch at I/O. Stein's rule: "you got to build something and then you have to feel it yourself".
One brilliant answer from a rough prototype gave the team conviction. A question about an outing with his daughter returned park details, site links and walkable Google Maps information.
Stein compares it to hitting one perfect golf shot. Leaders across the company backed the project, but the decision still came from a small team building something and feeling it themselves.
He staged the launch with testers who tell the truth. AI Mode went to about 500 external trusted testers, including friends and family who sent screenshots of what broke. Then Labs opened it to anyone, and real query data was used to tune it.
Only then did it launch to everyone in the US at I/O. One friend loved it and hated it, and sent constant 'this broke' screenshots. Stein: "you got to have people test it that tell you the truth and tell you when it sucks".
AI Mode works because it sits on top of Google Search. It runs query fanout, firing dozens of background searches and pairing each with web content. It uses Google's spam and quality signals plus live data: 50 billion shopping products and 250 million places.
Stein wants AI Overviews, Lens and AI Mode to blend into one experience. Follow-ups on an Overview open AI Mode, and so does a Lens photo.
Old search needs stay while AI adds new questions. People still look for phone numbers, prices, directions and tax payment pages. AI adds questions on top, so core search is not shrinking.
Google Lens visual searches grow 70% year over year and are already at billions of searches. Examples: photograph shoes and ask where to buy them, or photograph a bookshelf and ask which books to get next.
He credits focus and urgency, not one reorg or one leader. Stein says people often credit momentum to a one-time change or a single person. He points instead to leadership, close work with DeepMind, and researchers who had thought about this for years.
Each month the products and models improve until usage tips over. Gemini reaching number one in the app store, above ChatGPT, was one sign. He credits 'many many more people' than himself.
AI Mode targets information needs, not general chat. Typical uses are planning a trip, buying something or researching a question. Answers come with links so people can verify them and reach authoritative sources.
It is less focused on creativity or productivity, such as uploading a spreadsheet to get graphs. Most people come to learn something.
He tells creators to write for the need behind the question. Google's human rater guidelines still decide who shows up: does the page satisfy intent, cite sources, and say something original rather than what has been said 500 times.
His test is what content someone is using AI for, and how yours can be the best for it. People now ask AI more about advice, how-to and complex needs than simple lookups.
Stein sees the next year as setting AI habits for years. He feels an obligation to give users an AI-powered Google that draws on everything Google knows.
He calls it a healthy moment to build quickly, driven by a gap between what users need and what Search could answer.
Stein tied Stories to why people use Instagram. Instagram exists to share your life and connect with people. Stories did that with less pressure: no likes, ephemeral, full-screen and mobile.
He gives Snapchat credit for inventing the format. His view: "Not every great thing is going to be invented by you".
Stories got their own place instead of changing the feed. Early versions tried to make the main feed ephemeral, and that failed because the feed looks and works a specific way in people's minds.
A teammate said Stories felt like the story holes at the top that completed the product. Users think spatially, so a tile that vanishes or can't be liked would confuse them.
He made the borrowed format feel like Instagram. Instagram added neon drawing, sophisticated filters, camera roll upload and a pause, which Snapchat's version lacked.
Stein warns people fail most by assuming what works in one system works in theirs. In his words: "you never just want to snap in something that's working you have to make it work for your users".
New formats should complement the product. Stories did not replace Instagram, and AI does not replace Google. Both expanded the product, with a coherent look but a different primitive.
Stories share the aesthetic, camera roll and DMs. Google's AI is a full-page experience with follow-up conversation.
Stein found the root cause in a mistranslation. People added only one person to the list because it was mistranslated as 'best friend' in many markets. With one person, a reply was close to impossible.
With 20 to 30 people on the list, about two usually reply, which closes the loop. The team redesigned around larger lists, and the name was also originally 'favorites', which nudged people to add only two or three.
He treated Close Friends as an emotional job. The real job was feeling connected, so success meant getting a DM back. Christensen separates utility jobs from emotional jobs, and Stein says teams often discount the emotional ones.
The first version spread across feed posts, Stories and a profile, with green markers appearing inconsistently. The rebuild worked only in Stories.
Asking why people didn't post gave the team conviction. Users said an ex, a teacher or a judgy friend could see their Stories. That shared audience problem kept the team going.
Finstas, where people hacked Instagram into smaller private groups, confirmed the demand. The project took two to three years, one of the longest at Instagram. Tests ran in a basket of countries, with Australia among the first.
He chose clarity over cleverness with the green ring. Close Friends only became clear when the green ring went on the outside of the story, instead of a secret story that revealed itself when opened.
He cites Don Norman's glass doors with symmetric handles that don't say push or pull. The same logic is why the plain name AI Mode works.
He thinks Close Friends stayed too small too long. Close Friends was slow partly because the team stayed so small, with slow loop cycles. Stein says "people hold on to small teams too long".
His rule: add people once you have internal conviction and outside validation. Friends use a product to be kind, but by day 30, 60 or 90 they stop unless it is useful.
He reads retention curves for young products. Check the share still using it at day 7, 30 and 90. If the curve never flattens, in his words the product is toast.
After retention come two more gates: do people talk about it, and how big can it get. No product starts big. Even fast-growing ones began internally with 100 to 200 people.
He dissects a 5% drop by region, device and use case. In a big product, a 5% drop in a core metric calls for root cause analysis by region, device, demographic and use case.
He compares metrics to a pilot's instruments: they show whether the plane is flying, but not what to do next. A problem gives you the vision.
Stein tracks S-curves and moves when returns approach zero. When 50 more people on a project still would not move the needle, he looks for a new growth driver. That plateau usually coincides with changing expectations or market saturation.
Once a new engine lands, he staffs it up, since each change might be a 4%, 10% or 20% win. Without a quantified goal, teams congratulate themselves on improvements nobody uses.
He asks why people hired the product, not how to improve it. Drawing on Christensen's Competing Against Luck, Stein interviews users about the moment before first use: where they were, who they talked to, what prompted it.
He calls it lowercase jobs to be done, not the rigorous framework. The incremental Instagram move would have been making the square photo do more.
He stays humble about growth in mature products. Stein compares product work to golf: you're always one stroke away from shanking. Inside a big product some parts grow, some are mature and some decline.
On Instagram, sharing shifted from large public feed posts to lightweight formats like Stories and private DMs.
He gives AI models context in plain language, not fine-tuning. Models can now follow an order like a new engineer: describe your data, the API, the schema, the URL, when to use it and what to get right.
The model then decides whether to spend more reasoning budget or run code. He says "I don't think you need to do a lot of this heavy duty fine-tuning", and expects this to democratize who can build.
He wants AI to feel like talking to a person. AI Mode is built for back-and-forth conversation, and Search Live adds a voice version that just left Labs. His kids use it to ask about animals and history topics from school.
He sees natural language replacing prompt hacks, saying it is 'so obvious how humanlike the interface is becoming'.
He sees the next step for AI search as visual. A visual version of AI Mode lets people refine image results by talking to it, such as asking for a mid-century modern office with dark themes, then 'lighter, creamy, more California coastal'.
It finds images on the web, like AI image search, and is not Nano Banana, which is an image editor. Lens and image search growth shows the demand.
He keeps a playful culture alongside his clarity rule. The Nano Banana team was a scrappy, fun group that chose a fun name. Stein says it feels like his first stretch at Google.
The name goes against his advice on clarity, but he says no advice is right universally.
Curiosity is his motto, and user feedback is what he asks for. His motto is to be curious and not assume you have all the answers. He reads old free papers, for example on statistics, uses AI to find them, and blends AI, books and primary sources.
He almost named a company Curious. He is on X at rmstein and asks for problems with Google products and AI.
Google
