Vizepräsident für Google-Produkte: Es geht nicht darum, Funktionen zu stapeln, sondern AI zu lehren, Menschen zu verstehen.

In March 2025, Google Search launched a new button: AI Mode.

For the first time, it transformed the search experience from entering keywords and clicking links to engaging in continuous conversations. It doesn't just answer questions; it understands why you're asking.

By October, AI Mode was available in over 200 countries and regions and supported more than 35 languages. Data shows that the length of questions users ask in AI Mode is three times that of traditional searches.

On October 11, 2025, Robby Stein, the vice president of Google's product, participated in an interview. What he repeatedly emphasized was not the model's computing power or the technical approach but a more fundamental question:

We're not just adding more features; we're teaching AI to understand humans.

Behind this statement lies a shift in product thinking: In the future, AI won't just be more powerful; it will understand you better.

Section 1

AI Mode Isn't a Chatbot; It's an "Information Comprehension System"

After ChatGPT became popular, many people claimed that Google was finished.

The reason was simple: Who would want to click through a dozen links to browse web pages? People prefer one - sentence answers or direct conversations with AI.

But Robby Stein, the vice president of Google's product, responded:

"We're not creating a chatbot to accompany you; we're designing a system that can understand what you're looking for."

✅ What is AI Mode? It's Different from the Search You're Familiar With

Robby introduced that AI Mode first appeared as a small button at the search entry. When you click it, you won't see the traditional search box and blue links but an interface for continuous conversations.

You can ask a long - winded question and follow up with a second or third one. It remembers the context and can recommend related resources such as maps, official websites, and product links.

How is it different from "AI Overview"?

Robby said that AI Overview only adds an AI summary before traditional search results, providing you with a quick answer. In contrast, AI Mode is a complete interactive process and an entry designed to solve tasks.

It's not about modifying a module but building a new path. It's designed for information, not for casual chatting.

✅ Google Focuses on Comprehension, Not Generation

With ChatGPT, we're used to saying that AI generates an answer. But in AI Mode, Google is more concerned about whether the answer is truly useful to you.

For example, if you're planning a family trip, AI Mode will not only recommend destinations but also:

Show the walking distances between these places;

Provide official website links to confirm opening hours;

Embed maps in the conversation.

Robby gave an example from his personal experience: When traveling with my daughter, I asked AI Mode a question once, and it found all relevant park information, opening hours, walking distances, and official website verification links. Suddenly, I realized that this was not just an enhanced version of a search engine but a new system.

He compared this experience to a perfect shot in golf.

At that moment, you know it's truly helping you.

Section 2

How Was AI Mode "Taught"?

Robby Stein said that AI Mode wasn't a project initiated by Google's top management with the whole company rushing in. Initially, it was just a small project by a few people, even without a name.

"At first, the team had only 5 to 10 people, including engineers, designers, and a technical leader. We weren't trying to overhaul the search; we just wanted to see what would happen if users could ask any questions."

The first thing they did wasn't writing code but creating a blank page: a prototype interface where you could enter questions and get AI responses, with no decorations, just a cursor.

Step 1: Find the Moment When It's Truly Useful

Robby said that the goal at that time was simple: to find out when it was helpful.

During early testing, some users manually added "AI" at the end of the search box, hoping to trigger a more intelligent response. Others took photos of their homework and asked how to solve the second question.

These behaviors made the team realize that users were already teaching the search engine how to evolve.

Robby said that these early signals were more important than any metrics.

Step 2: Involve Real Users and Listen to Real Feedback

Next, they gave this prototype to 500 test users, mostly friends, family, and some internal employees.

The requirement was simple: Just treat this as your new search. If there's a problem, just take a screenshot and send it to me.

"A friend of mine used it very seriously and was also very critical. He sent me screenshots every day, saying 'This answer is wrong', 'This sentence is completely off', 'I can't understand this information'. I'm really grateful for these feedback; they're all valuable."

They didn't extend the project cycle or wait for the product to be perfect before launching. Instead, like a startup team, they improved and tested it while using it.

These real queries became the initial training materials for Google's AI Mode.

Robby described this process in one sentence: "You can't wait for everything to be perfect before launching. You have to find the right direction first and then polish it."

Step 3: Put It in Search Labs and Observe Real Usage Data

After the first - round user feedback, they finally released a public beta version and put it in Google's Search Labs. This meant that anyone could enable AI Mode and conduct real searches.

The goal at this stage changed:

• What questions do users ask most frequently?
• Which questions are answered poorly?
• Will users follow up with more questions?
• When will they exit the page?

All the data was used to figure out what needed to be changed, what to delete, and what was worth keeping.

This stage was crucial: Only when users are actually using it can you find out what's wrong. Real optimization isn't the team lecturing the model; it's letting users' real behaviors teach it how to improve.

Step 4: Roll It Out Globally in Phases

From the initial 5 - person prototype to the launch in Search Labs and finally the official release of the entry button, Google didn't launch it globally all at once. Instead, it tested in phases and expanded country by country.

Why? Robby's answer was straightforward:

"AI Mode isn't just a module; it's a way for people to access information. You can't take the risk of changing the entry for the whole world at once."

It took Google about a year to go from a blank page for 5 people to a global entry covering over 200 countries. The method was simple:

First, create a usable version.

Involve real users to ask questions.

Use every piece of feedback as a direction for improvement.

It's not the team defining how AI should answer; it's users' real needs teaching it how to answer.

Section 3

Making a Product Isn't Difficult; Understanding What Users Really Want Is

Robby Stein doesn't just work on AI. He was the product leader at Instagram and launched features like Stories (ephemeral posts) and Close Friends (a feature for sharing with selected people).

But in this interview, what he was most willing to talk about wasn't the successful one but the initially failed one:

"The Close Friends feature took two or three years to develop. When it was first launched, no one used it at all."

✅ What is Close Friends? Why Didn't People Understand It at First?

This feature allows you to create a list of close friends. You can post a Story that only these people can see, so you don't have to let the whole world see your emotions, your selfie without makeup, or your complaints.

It sounds good, right? But at first, no one used it.

Why?

Robby said: We made it a place where you could post anything: You could post in the Feed, in Stories, and there was also exclusive content for close friends on your profile. It was extremely confusing.

Worse still, the feature was named "Favorites". Many people thought they could only add one or two very close people. As a result, people added two friends, posted a Story, and the friends might not respond. It was extremely awkward.

"We hoped to create a sense of connection, but people posted content and didn't receive responses. The whole experience was broken."

✅ The Real Problem: It's Not the Lack of Features but the High Pressure

Robby recalled that later they did something: They asked users again: Why don't you post Stories?

The answers they got were almost the same:

• "My ex - partner will see it."
• "My boss is following me."
• "There's a friend who always comments on my looks."

In a nutshell: It's not that people didn't want to post; they were afraid of being seen by the wrong people.

This was the core problem: What users wanted wasn't a feature for sharing with special people but a stress - free small circle.

✅ Three Product Principles He Summarized

First: Figure Out Why Users "Hire" Your Product

Robby used a classic analogy to explain:

"People don't want a drill; they want a hole in the wall."

In other words, users don't use a product; they "hire" it to help them accomplish something. Just like when you turn on AI Mode, what users really want isn't to "experience AI" but to find an assistant to help them plan trips, write emails, and look up information.

Second: Don't Just Look at Data; Find the "Key Statement"

Data can tell you that "fewer people are clicking in" or "fewer people are posting content", but the real key is the "why".

Robby said that they asked a user: When did you post content in Close Friends? The user replied: I was feeling bad that day and wanted to post something but didn't want everyone to see it.

This statement became the starting point for product design: It's not about showing yourself but seeking a response.

Third: Don't Be Too Clever; Make It Easy to Understand at a Glance

The initial version was like this:

There was no mark when posting Close Friends content;

Viewers didn't know it was "content just for me";

Posters didn't know if anyone had received it.

Later, the team made a small change and changed the color of the Story circle to green. People immediately knew it was content just for close friends.

This "visible" design made people more willing to post and use it.

This experience deeply influenced Robby's approach to developing AI Mode later.

He no longer focused on the feature list but on one question: What does this product help users accomplish?

The quality of an AI product isn't determined by how fancy its features are but by whether it truly understands what users want to do.

Section 4

It's Okay to Borrow, but Make It Your Own Version

In 2016, when Instagram launched the Stories feature, almost everyone said it was "copying" Snapchat.

Robby Stein, who was in charge of this project, responded:

"This format was indeed invented by Snapchat. But what we care about is: Can it make Instagram better? You don't have to invent everything good. The key is to make it your own version."

✅ Why Isn't Stories Just a Copy of Snapchat?

Robby said that they didn't copy Snapchat's approach but "remade it in the Instagram way".

Here are some specific things he mentioned:

Snapchat doesn't allow uploading existing photos; you have to take them on - the - spot. → Instagram changed it and allowed uploading high - definition photos taken on your phone because many people wanted to preserve memories.

Snapchat's Stories can't be paused while playing. → Instagram added a "long - press to pause" function so users could view them slowly without missing details.

Instagram also added complex filters, neon brushes, and more creative tools. → This made the way of expression more diverse and more suitable for Instagram users who were used to "editing photos".

These weren't technological innovations but choices based on user scenarios.

If users have already formed an idea of what a product should be like, trying to force it into another form will mostly end in failure.

✅ The Same Logic Applies to AI Mode

Now, let's talk about Google's AI Mode. Many people's first reaction is: Isn't this just another ChatGPT?

Robby's response was the same as when he worked on Stories: We're not making a copy of ChatGPT. What we're doing is integrating AI into search and serving users in a search - based way.

What does that mean?

ChatGPT is very powerful, but it's an open - ended conversation platform where you can write poems, create stories, or role - play. In contrast, AI Mode is designed around the task of "accessing information".

For example, if you search for "science exhibitions suitable for 6 - year - olds", AI Mode will:

1. First, understand the meaning of the combination of "6 - year - olds" and "science".
2. Then recommend age - appropriate content, along with maps, ticket prices, and official websites.
3. Finally, you can follow up with questions like "Are there any this weekend?" or "Do I need to make a reservation?"

This isn't AI chatting; it's AI following your questions.

Robby used "usage scenarios" to draw the line:

• If you want to write a letter, polish a text, or find inspiration, ChatGPT is a great choice.
• If you need to find an address, check an exhibition, or compare information, AI Mode is more suitable.

"You wouldn't use ChatGPT as a map, nor would you use it to confirm a restaurant's opening hours. We focus on things that ChatGPT doesn't do well but that users really need."

From Instagram to Google, from Stories to AI Mode, Robby's concept has always been the same: It's okay to borrow inspiration from others, but you need to figure out how to make it suitable for your users.

If other users care about "fun", and your users care about "usefulness", then design around "usefulness".

Section 5

A Good Product Isn't the Smartest One; It's the One That Understands You Best

At the end of the interview, Robby Stein returned to a very basic question: What makes a good AI?

It's not about the size of the model, the strength of the parameters, or the ability to write novels or code. It boils down to one thing:

Can it understand what you're asking?

✅ The Core Value of AI: Understanding What You're Asking

In the past, when using a search engine, we had to figure out the keywords ourselves. When using ChatGPT, we also had to organize our language and then generate content.

But Robby said that AI is becoming "more in line with human thinking". You can just ask whatever you want directly, without worrying about the format or the rigor of your expression.

Even kids can feel this change:

"My kids come home from school and say, 'Can I talk to Google? I want to ask it what zebras like to eat.' Then they open Google, directly state their questions, and start a conversation."

✅ This Isn't Using a Search; It's Practicing Asking Questions

Robby called this phenomenon AI becoming an engine of curiosity.