What it takes to be a top PM today | Robby Stein (Google Search)
Robby Stein, drawing on his work on Instagram Stories, Reels, and AI-era Google Search, argues that the core craft of product management is making great decisions. He lays out a three-part process: understand people's underlying needs through jobs to be done, diagnose why a product is failing through root cause analysis and iteration, and refine the craft details that make a product work well and feel delightful.
Robby Stein opens by calling this one of the most interesting moments in history to build products, while admitting the PM role is changing. His career includes startups, leading teams through major shifts at Instagram (Stories, Reels, feed ranking, direct messaging), and now AI-era Google Search, where the goal is to let people ask anything and get the most helpful answer across video, audio, text, and the web. He is explicit that this talk is not about how AI transforms products but about the craft of product: what it takes to build something people love when AI agents let almost anyone build almost anything in a fraction of the time that once took years and whole teams.
The PM’s real value is judgment
As building gets cheaper, Stein argues that the traditional PM value of getting things done, organizing people, and creating energy and momentum matters less than judgment and taste. He says the craft of PM comes down to one thing you should be the best at: decision-making. He calls it humbling, because many decisions are wrong, and he has treated the lessons from his career as a book he has been writing in his head, organized into three chapters: understanding people deeply, diagnosing root causes, and refining the craft.
Chapter one: understanding people deeply
Stein starts with the jobs-to-be-done framework from Clayton Christensen’s Competing Against Luck. The core idea is that people do not really use products; they hire them to make progress in their lives, and you have to know what that progress is. The best way to extract it is through a story: reconstruct the exact moment of a past decision, including where the person was, who they were with, what they were reading or wearing, and whether they were in a store or online.
He illustrates this with buying a bed. He and his wife narrowed it to three or four options, and she picked one for a reason that had nothing to do with cooling, eco-friendliness, or price. The transcript renders the key concept as transference, which in context means motion isolation: she needed a bed that would not wake her when he moved. A PM who never asked the question would never have discovered that the decision hinged on it.
He then gives two Google examples. Early in the AI experiences, the team noticed people came for inspiration, such as choosing a couch. Text-only chatbots could not handle questions like how to describe a fuzzy couch or how a throw pillow draws attention, so the team invested early in state-of-the-art multimodal understanding, retrieval, and knowledge systems that support multi-turn visual conversation. The second example is AI Mode for food and neighborhood questions. The first version was missing what people expected from Google: maps, star ratings, closing times, price levels, and trusted descriptions of places. Once those were brought together into a rich, entity-based experience, it became one of the most popular features based on thumbs-up ratings and other signals.
He closes the chapter by suggesting AI can scale this work. Transcripts from user interviews could be mined to source jobs to be done automatically, and an agent could be built to interview users using the same methodology, asking why they use a product and what underlying problems they were solving.
Chapter two: diagnosing root causes
The second chapter starts from the reality that the first version of nearly everything he has built was not great. Stein compares the process to an epoch of model training: maintain a clear, analytical view of the problems in your product space, rank them by priority, understand them deeply, fix them, and loop again. Each iteration moves you closer to product-market fit.
The Instagram Stories example shows the technique. Stories took off early, but not everyone adopted it or felt comfortable sharing. Instead of asking why people did the thing, the team asked why they were not doing it. The answers clustered around audience: an ex, a teacher, an aunt, a sibling would see the post. A large quantitative survey of thousands of people then weighted these burdens, and the conclusion was blunt: if you cannot overcome the audience problem, no creative tool will work. Two years of iteration led to Close Friends, but the first versions were confusing, with Favorites, a hidden private profile view, and a separate feed label that nobody understood. The only piece that worked was inside Stories, so the team cut everything else and relaunched it as a Stories experience.
The Reels example is the inverse. A launch in Brazil assumed that goofy dances were ephemeral content that should disappear, and it was a huge failure. The deeper insight was that people who invested time in content wanted it to last, to build an audience, a business, or a viral following. The update made Reels permanent and visible to a whole school, and that became a major success.
Stein says AI makes this loop far easier. Opted-in users can be asked directly where a product missed the mark, and a model can combine qualitative comments with quantitative signals. He uses a backpack shopping example: the assistant correctly surfaced sizes and features but should have asked about the child’s height and weight. Turning the interaction into a collaborative conversation helps users reach their objective even when they do not know what to ask, and this approach drove some of the biggest gains in engagement and perceived helpfulness.
Chapter three: craft, pain, and delight
Craft is the sense people get about whether the creators cared. Stein identifies two drivers: whether the product works perfectly without user pain, and how it makes the user feel.
On pain, he describes an internal tool built with Antigravity that uses an agent to act like a user. It searches, asks a follow-up, takes screenshots, and evaluates the results against a comprehensive rubric. It checks things like whether math renders properly in LaTeX, whether stray characters appear, and whether a visual question about bioluminescence returns an image tray instead of plain text. The team has built agents that identify breakages and increasingly fix them, which replaces much of the manual QA work. The PM or team member who obsessively uses every part of the product is still valuable, but this can now be scaled much more efficiently.
On delight, he points to the redesigned Google search box, the first real change to it in about 20 years. It was reimagined as an AI input that grows with you and can accept files, videos, and photos. The team wanted it to still feel like Google while signaling that AI is inside, so they used a color gradient to represent AI power. Small details carry the feeling: the blinking cursor cycles through Google’s colors, and tapping the box makes it react with motion and haptic feedback. Stein argues that these intentional details, through motion, color, and touch, communicate humanity and give a product a soul, and that many teams overlook this part of the creative process.
The playbook
Stein ends by synthesizing the three chapters into a playbook. Understand people deeply and find their core needs and problems. Be rigorous and analytical about what must be fixed for the product to win. Build something that is simple, works well, and makes people feel good using it. He admits much of this is not new, but AI changes what is possible. He encourages PMs to spend less time chasing clever features or new technology and more time thinking about the people on the other side of the screen.