The Future of Google Search: Talking AI Mode with Robby Stein

00:26:22
https://www.youtube.com/watch?v=dVZ_WmKIkoc

Ringkasan

TLDRIn this episode of the Imaginative Podcast, Robbie Stein, VP of Product at Google Search, discusses the transformative role of AI in search technology. He introduces AI overviews, which provide quick, relevant information, and AI mode, allowing users to engage in natural language queries. Robbie emphasizes the importance of accuracy and quality in AI responses, leveraging Google's extensive search history to enhance results. He addresses trust concerns, the potential for businesses to adapt their content strategies, and the future of search as AI continues to evolve, aiming for a seamless user experience.

Takeaways

  • 🤖 AI is transforming search technology.
  • 📊 AI overviews provide quick snapshots of information.
  • 🗣️ AI mode allows natural language queries.
  • 🔍 Businesses should focus on high-quality content.
  • 💡 Google leverages extensive search history for accuracy.
  • 📈 AI mode is currently available in the US.
  • 🔗 Query fan out generates related questions.
  • 🛠️ Google is experimenting with ads in AI experiences.
  • 🔒 Trust in AI is crucial for user acceptance.
  • 🌐 The future of search will integrate AI seamlessly.

Garis waktu

  • 00:00:00 - 00:05:00

    The podcast introduces Robbie Stein, VP of Product at Google Search, discussing the impact of AI on search. Businesses are encouraged to adapt their content strategies to rank effectively in AI-driven search results as Google innovates with AI overviews and AI mode.

  • 00:05:00 - 00:10:00

    Robbie explains AI overviews, which provide quick snapshots of information with links for deeper exploration. This feature aims to enhance user experience by allowing natural language queries and multimodal interactions, such as using images or voice to ask questions.

  • 00:10:00 - 00:15:00

    The conversation highlights Google's cautious approach to AI integration in search, emphasizing the importance of understanding user queries and providing relevant information. AI mode is introduced as a new feature that allows users to engage in a more conversational manner with the search engine.

  • 00:15:00 - 00:20:00

    AI mode utilizes advanced models to generate responses based on user queries, incorporating real-time data and a query fan-out technique to explore related questions. This aims to enhance the search experience by providing diverse and relevant information.

  • 00:20:00 - 00:26:22

    Robbie reassures businesses that traditional content quality guidelines still apply in the AI context. Google is exploring ways to monetize AI features while ensuring high-quality sources are prioritized, addressing concerns about trust and accuracy in AI-generated content.

Tampilkan lebih banyak

Peta Pikiran

Video Tanya Jawab

  • What are AI overviews?

    AI overviews provide quick snapshots of information with links for deeper exploration, designed to answer specific questions.

  • What is AI mode?

    AI mode allows users to ask questions in natural language and receive comprehensive responses, including links to relevant information.

  • How can businesses adapt to AI in search?

    Businesses should focus on creating high-quality, accurate content that answers user questions, following Google's content guidelines.

  • Will ads be included in AI overviews?

    Google is experimenting with how ads could work in AI experiences, focusing on providing helpful information.

  • How does Google ensure the quality of AI responses?

    Google uses its extensive search signals and systems to prioritize high-quality sources and reduce low-quality information.

  • What is query fan out?

    Query fan out generates related questions based on the user's initial query to provide a broader range of information.

  • How can users access AI mode?

    Users in the US can access AI mode by searching for 'AI mode' and enabling it on the Google Labs page.

  • What is the future of AI in search?

    The future involves expanding the capabilities of AI to provide more personalized and efficient search experiences.

  • How does Google handle AI hallucinations?

    Google trains its models to minimize mistakes and improve accuracy, especially in critical areas like health.

  • What is the role of Gemini in AI mode?

    Gemini is the foundational model used in AI mode, designed to enhance information retrieval and reasoning.

Lihat lebih banyak ringkasan video

Dapatkan akses instan ke ringkasan video YouTube gratis yang didukung oleh AI!
Teks
en
Gulir Otomatis:
  • 00:00:00
    These AI systems are new overall in the
  • 00:00:01
    world and everyone's learning on how to
  • 00:00:03
    best implement them and make them as
  • 00:00:05
    accurate and useful as possible,
  • 00:00:06
    including Google. If I'm a business and
  • 00:00:08
    I'm listening, is there anything that I
  • 00:00:09
    need to be doing differently in terms of
  • 00:00:11
    getting my content to rank in AI mode or
  • 00:00:14
    in AI overviews? So, we're still in the
  • 00:00:17
    early phases here, but I think that our
  • 00:00:19
    goal is that this is as effortless and
  • 00:00:20
    easy as possible. Hey everyone, welcome
  • 00:00:22
    to another episode of the Imaginative
  • 00:00:24
    Podcast. Today, I'm super excited to be
  • 00:00:26
    joined by Robbie Stein. Robbie is the VP
  • 00:00:29
    of product at Google search. A lot is
  • 00:00:32
    changing in the world of search due to
  • 00:00:34
    AI. We have new entrance in the space,
  • 00:00:36
    but of course, Google has been
  • 00:00:38
    innovating. We have seen them release AI
  • 00:00:40
    overviews and now the new AI mode. So,
  • 00:00:44
    Robbie's joining us to dive deeper into
  • 00:00:46
    Google's plans for search, how they're
  • 00:00:48
    thinking about product innovation in a
  • 00:00:50
    space that is rapidly changing. Robbie,
  • 00:00:52
    it's great to have you on. Thank you for
  • 00:00:53
    joining us. Thanks so much for having
  • 00:00:55
    me. I wanted to start Robin by asking a
  • 00:00:58
    question that I heard on another podcast
  • 00:01:00
    that I thought was really cool. What is
  • 00:01:02
    it that people should know about you
  • 00:01:03
    that they probably don't? That's
  • 00:01:05
    interesting. It might be that I was when
  • 00:01:11
    I was younger and still am a magician
  • 00:01:13
    really and did some birthday parties.
  • 00:01:16
    So, I think you could say that I am a
  • 00:01:18
    hobbyist but technically at some point
  • 00:01:20
    was professional on some birthday
  • 00:01:21
    circuit. And I've loved magic my whole
  • 00:01:23
    life since I was eight and I spent a lot
  • 00:01:25
    of time practicing magic and enjoying
  • 00:01:27
    it. And I think maybe a little bit of a
  • 00:01:29
    bridge to technology because of our
  • 00:01:30
    products I think can truly be magical
  • 00:01:32
    and blow people away. I guess I'm like
  • 00:01:34
    consistent and like to do that even like
  • 00:01:36
    in my free time as a hobby sometimes and
  • 00:01:38
    with two little kids it's starting to be
  • 00:01:40
    something I do again. I'm sure that is
  • 00:01:42
    amazing. I'm a big fan of magic also.
  • 00:01:45
    I'm sure we're going to have lots of
  • 00:01:46
    sidebar conversations. So, thanks for
  • 00:01:48
    joining us, Robbie. And I wanted to make
  • 00:01:51
    the most of this time that we have
  • 00:01:52
    together and dive right in. Before we
  • 00:01:55
    get started, though, I wanted to just
  • 00:01:57
    spend a few minutes to level set
  • 00:01:58
    everyone on some of the new releases
  • 00:02:00
    that Google has put out in terms of
  • 00:02:02
    their search product. So, do you mind
  • 00:02:04
    just taking a step back explaining to
  • 00:02:06
    the listeners what is AI overviews and
  • 00:02:09
    what is the new AI mode? How are they
  • 00:02:11
    different and why should they be excited
  • 00:02:13
    about it? Yeah, sure. Obviously, AI has
  • 00:02:14
    not been new to search, but last year we
  • 00:02:17
    took a big step forward with an AI
  • 00:02:19
    ccentric experience called AI overviews.
  • 00:02:21
    And what it allows you to do is to get a
  • 00:02:23
    quick snapshot of information with links
  • 00:02:25
    to to dig deeper for really things that
  • 00:02:27
    you have on your mind. We call these
  • 00:02:29
    kind of like answer seeking questions.
  • 00:02:30
    When you have something pretty specific,
  • 00:02:32
    you can put that into Google. Might be,
  • 00:02:33
    hey, how do I get a coffee stain out of
  • 00:02:36
    my couch with when I have toddlers or
  • 00:02:38
    something? And what frequently see is a
  • 00:02:41
    short blurb that gets generated by AI at
  • 00:02:44
    the top with helpful information and
  • 00:02:46
    links to dig deeper and learn more. It
  • 00:02:48
    may include links to other articles or
  • 00:02:50
    how-to videos, links to YouTube or
  • 00:02:53
    embedded videos. And really, it's to
  • 00:02:55
    give you that snapshot of information
  • 00:02:56
    and the context. And that's something
  • 00:02:58
    that we've been rolling out across the
  • 00:02:59
    world. It's now available to most of the
  • 00:03:01
    world via the last set of announcements.
  • 00:03:04
    And it's also something we've been
  • 00:03:05
    increasingly connecting to our
  • 00:03:07
    multimodal experiences, which for people
  • 00:03:08
    who aren't familiar is how you can ask
  • 00:03:10
    questions with video or voice or photos.
  • 00:03:14
    So you could take a picture of something
  • 00:03:16
    and ask about it or actually take a
  • 00:03:18
    screenshot on a device and also ask
  • 00:03:20
    questions about that as well. It might
  • 00:03:22
    be you take a picture of a piece of art
  • 00:03:24
    and you ask more about the background of
  • 00:03:26
    it. We can answer those, too. So, we're
  • 00:03:28
    excited because it really is the
  • 00:03:29
    beginning of starting to pretty widely
  • 00:03:31
    expand what Google search can do for you
  • 00:03:34
    and provide as natural as possible of an
  • 00:03:37
    interface for those experiences. All
  • 00:03:39
    right. Obviously, there are lots of new
  • 00:03:41
    entrance into this space. Open AI is
  • 00:03:43
    exploring search features. You have
  • 00:03:45
    tools like Perplexity. Google kind of
  • 00:03:47
    took their time to roll out a generative
  • 00:03:50
    AI solution. Why was that the case? Were
  • 00:03:53
    there technical challenges that you were
  • 00:03:54
    looking to solve? How has Google been
  • 00:03:56
    intentional in it in this roll out of AI
  • 00:03:58
    mode and previously AI overviews? Yeah,
  • 00:04:00
    obviously when you have a product that
  • 00:04:02
    people use at such a large scale, you
  • 00:04:04
    want to be sure you get it right and you
  • 00:04:05
    do you're really thoughtful about how
  • 00:04:07
    you approach this. And I think we
  • 00:04:09
    started by understanding what are the
  • 00:04:10
    kinds of questions people are asking of
  • 00:04:12
    Google and how can we make those allow
  • 00:04:14
    you to ask even more and in the more
  • 00:04:15
    natural language and ways that people
  • 00:04:17
    can now ask and really that created AI
  • 00:04:20
    overviews and the way it works is
  • 00:04:22
    because people come to Google for such a
  • 00:04:24
    wide set of questions. You might be just
  • 00:04:26
    learning about a new band. You just type
  • 00:04:27
    in the artist's name. Nothing else. It's
  • 00:04:30
    not a question. There's no You just want
  • 00:04:31
    to know what's up, right? And so in some
  • 00:04:33
    cases, it's you actually want a totally
  • 00:04:36
    different experience than something that
  • 00:04:37
    might feel more AIcentric with rich
  • 00:04:39
    information. You actually just might
  • 00:04:41
    want to browse or you might look for
  • 00:04:42
    gardening ideas and you actually want
  • 00:04:44
    images, right? So Google accommodates
  • 00:04:46
    such a wide set of questions. We really
  • 00:04:48
    got to get it right and bring AI to the
  • 00:04:51
    front where it makes sense. And so the
  • 00:04:53
    system actually learns this and that's
  • 00:04:54
    what AI overviews really offers which is
  • 00:04:56
    AI where we found it to be most helpful.
  • 00:04:58
    And what we learned in rolling out AI
  • 00:05:00
    overview is that people wanted more of
  • 00:05:01
    it and they wanted to be able to control
  • 00:05:03
    wanted the control and choice to say hey
  • 00:05:05
    I have a question that I I'm pretty sure
  • 00:05:07
    AI can be helpful with and by the way
  • 00:05:09
    want to also ask follow-up questions and
  • 00:05:11
    go back and forth and that's something
  • 00:05:12
    that's pretty hard to do in the
  • 00:05:13
    construct of the current search
  • 00:05:15
    experience. And that led us to build
  • 00:05:17
    this experiment called AI mode, which is
  • 00:05:19
    a place you can go as a mode selection
  • 00:05:22
    or as a shortcut in a few places like in
  • 00:05:24
    the app to just ask whatever is on your
  • 00:05:26
    mind. Again, you can use plain English,
  • 00:05:28
    natural language. And as of recently,
  • 00:05:30
    the last few weeks, you can take a photo
  • 00:05:31
    or use voice as well. And then what will
  • 00:05:34
    happen is the AI mode will go find the
  • 00:05:36
    most helpful information for you along
  • 00:05:38
    with links and ways to dig deeper. bring
  • 00:05:40
    that all to you in this new experience
  • 00:05:42
    that we built out which is currently in
  • 00:05:44
    labs and we're really excited about and
  • 00:05:45
    happy to share more about. So let's talk
  • 00:05:47
    AI mode because that sounds super
  • 00:05:48
    interesting reading through when you
  • 00:05:50
    released the press release around about
  • 00:05:53
    AI query fan out. Can you double click
  • 00:05:56
    on that and explain how that works? How
  • 00:05:57
    it integrates with your ranking
  • 00:05:59
    algorithms and real-time data? Yeah, so
  • 00:06:01
    with AI mode, it's really very powerful
  • 00:06:03
    model that brings together the power of
  • 00:06:06
    Gemini's foundational models with
  • 00:06:08
    thinking and reasoning with the
  • 00:06:10
    real-time information systems of search.
  • 00:06:12
    And so it's really magic. We've really
  • 00:06:14
    designed it to be focused on
  • 00:06:16
    informational needs like what are the
  • 00:06:17
    people the kinds of needs people we see
  • 00:06:19
    coming to search but really being able
  • 00:06:21
    to expand what's possible in those
  • 00:06:22
    spaces. So we didn't design it to be
  • 00:06:25
    like a general purpose AI product that
  • 00:06:27
    can generate images for you or upload
  • 00:06:29
    huge amounts of data to but we really
  • 00:06:32
    designed it to be great at finding and
  • 00:06:34
    understanding information. And so what
  • 00:06:36
    that means is now when you ask a
  • 00:06:38
    question, we've trained the model to use
  • 00:06:40
    Google search. And so what will happen
  • 00:06:41
    is you ask one question and it will do
  • 00:06:44
    something we call query fan out which
  • 00:06:45
    may generate multiple other questions
  • 00:06:47
    that could be related. It will then go
  • 00:06:49
    out and find the most useful information
  • 00:06:52
    both in terms of the web but then also
  • 00:06:55
    across all of the interesting real-time
  • 00:06:57
    systems that Google has built out. So
  • 00:06:59
    that might be the billions of product
  • 00:07:01
    catalog data, products and catalog data
  • 00:07:03
    that are available, real-time
  • 00:07:04
    information from local listings,
  • 00:07:06
    businesses, real world information.
  • 00:07:09
    There's knowledge panel which is a whole
  • 00:07:11
    incredibly rich knowledge base of
  • 00:07:13
    information and all of that information
  • 00:07:16
    can then come back to the model and then
  • 00:07:18
    the model can generate a response that
  • 00:07:20
    gives context and links to dig deeper.
  • 00:07:22
    And that's really this really unique
  • 00:07:25
    experience we've built out. And then and
  • 00:07:27
    now hopefully the idea is that you can
  • 00:07:28
    come to Google with really any question
  • 00:07:30
    on your mind and get a great experience.
  • 00:07:32
    All right, I'm buying that so far. But
  • 00:07:34
    two things that I think every creator
  • 00:07:36
    has on their mind and then just people
  • 00:07:38
    generally for a creator is AI mode going
  • 00:07:42
    to really help in terms of driving
  • 00:07:44
    discovery or is it about keeping users
  • 00:07:47
    in the Google ecosystem? What we're
  • 00:07:49
    finding is that when we add AI products
  • 00:07:51
    and experiences like this, it's
  • 00:07:52
    expansionary. people use these services
  • 00:07:54
    more and it creates growth. So what
  • 00:07:56
    we've seen is for instance with AI
  • 00:07:58
    overviews people are using Google more
  • 00:08:00
    searching for more things and newer
  • 00:08:02
    things. An example of this is with with
  • 00:08:04
    lens which is our photo-based experience
  • 00:08:06
    which is also available in AI mode. You
  • 00:08:08
    can take a picture of a bookshelf and
  • 00:08:10
    say what are other interesting books
  • 00:08:11
    here. These are hard these things that
  • 00:08:13
    were just not possible before. And so
  • 00:08:15
    when we generate an AI response and it
  • 00:08:18
    includes links to interesting great good
  • 00:08:20
    readads, authoritative articles, there's
  • 00:08:23
    opportunities hopefully for websites to
  • 00:08:25
    rank and be discovered. And because of
  • 00:08:27
    the query fanout technique I mentioned,
  • 00:08:30
    it gives a broader and more diverse set
  • 00:08:31
    of websites that could show up there too
  • 00:08:33
    because it's not it's going beyond the
  • 00:08:35
    original query in the thinking and the
  • 00:08:37
    reasoning steps of this experience. And
  • 00:08:39
    what we do find is that people they want
  • 00:08:41
    this context up front. They want to be
  • 00:08:43
    able to get a sense of things, but then
  • 00:08:45
    they do click in. And I think what we
  • 00:08:47
    hear from users is they want to
  • 00:08:49
    ultimately see where information comes
  • 00:08:50
    from themsel. They don't want to take
  • 00:08:52
    anyone's word for it. They want to see a
  • 00:08:54
    diversity of perspectives and they they
  • 00:08:56
    trust certain sources and that's that
  • 00:08:58
    hasn't changed. And we've really made I
  • 00:09:01
    think connecting to the web and diving
  • 00:09:03
    deeper an absolutely critical and
  • 00:09:05
    central design principle both with AIO
  • 00:09:07
    AI overviews and AI mode. That is
  • 00:09:10
    because as you and a few things we've
  • 00:09:11
    done there is actually train the model
  • 00:09:13
    to link to other websites and have them
  • 00:09:16
    show up where most helpful as you read
  • 00:09:18
    AI mode responses like it will actually
  • 00:09:19
    link directly to websites where it finds
  • 00:09:21
    that helpful. It could be booking
  • 00:09:23
    things, reading more, seeing lists, and
  • 00:09:25
    we're working on this as a key design
  • 00:09:26
    principle because we see that as
  • 00:09:28
    something as people really want and is
  • 00:09:29
    really useful. Now, in in terms of the
  • 00:09:32
    link into websites, Google obviously has
  • 00:09:36
    so much experience with their ranking
  • 00:09:37
    algorithm and that's an advantage that
  • 00:09:39
    many of the newer startups don't have
  • 00:09:41
    because we have seen instances where the
  • 00:09:44
    quality of the sources are pretty low.
  • 00:09:46
    Maybe they're from just a chat room on
  • 00:09:49
    Reddit. Maybe it's just from a website
  • 00:09:51
    that somebody decided to write an
  • 00:09:53
    article about. How is Google thinking
  • 00:09:55
    about bringing highquality sources into
  • 00:09:58
    AI conversations and to reduce the
  • 00:10:02
    ability to get lowquality sources. Yeah.
  • 00:10:05
    What's cool about AI overviews and AI
  • 00:10:08
    mode is that it builds on the 25 years
  • 00:10:10
    of search signals and systems and
  • 00:10:12
    knowledge that exist. And so during this
  • 00:10:14
    query fanout technique, the systems are
  • 00:10:16
    using these underlying search signals to
  • 00:10:18
    understand, hey, what pieces of
  • 00:10:20
    information, what websites have been the
  • 00:10:22
    most helpful given the context of these
  • 00:10:25
    kinds of questions before. And all of
  • 00:10:26
    that's available in helping the system
  • 00:10:29
    determine what to show the user. And and
  • 00:10:32
    in the same case, there might be
  • 00:10:33
    websites or sources of information that
  • 00:10:35
    have been maybe manipulative or
  • 00:10:37
    generated specifically to attract
  • 00:10:38
    attention in ways that haven't been
  • 00:10:40
    helpful for people. Our systems have
  • 00:10:42
    detections and ways of finding those
  • 00:10:44
    too. And it's not a brand new thing in
  • 00:10:46
    that regard. It's really an evolution, a
  • 00:10:48
    core part of our search experience,
  • 00:10:50
    building on that foundation. Okay. And
  • 00:10:53
    so if I'm a business and I'm listening,
  • 00:10:55
    is there anything that I need to be
  • 00:10:57
    doing differently in terms of getting my
  • 00:10:59
    content to rank in AI mode or in AI
  • 00:11:02
    overviews? How should businesses be
  • 00:11:04
    thinking about this? Yeah, I think that
  • 00:11:05
    in general the principles are very
  • 00:11:07
    similar. you ultimately what we
  • 00:11:10
    recommend in terms of if you look up
  • 00:11:11
    Google's history of content guidelines
  • 00:11:14
    and rating guidelines and search best
  • 00:11:15
    practices is really tied to something
  • 00:11:18
    that's outside of products. It's really
  • 00:11:19
    like how do humans find in for certain
  • 00:11:22
    information helpful. Is it a does it
  • 00:11:24
    answer the user's question? Does is it
  • 00:11:26
    accurate? Is it factual? There's a whole
  • 00:11:28
    bunch of things that if you're doing
  • 00:11:30
    it's great and that is not has nothing
  • 00:11:32
    to do with any product. It's just as a
  • 00:11:35
    person when you're experiencing
  • 00:11:36
    information, you're going to trust it,
  • 00:11:38
    value it, and find it helpful. And so
  • 00:11:39
    the more that people are producing that
  • 00:11:41
    information, the more that should be
  • 00:11:43
    something that is detectable and is
  • 00:11:46
    found in the systems and prioritized.
  • 00:11:48
    And so I'd say it's a very similar set
  • 00:11:49
    of best practices. That's nice here
  • 00:11:51
    because we're using very similar signals
  • 00:11:53
    to determine quality in this context.
  • 00:11:55
    Right. Awesome. I I'm sure a lot of
  • 00:11:57
    businesses have been trying to think of
  • 00:11:59
    how they can have the equivalent of SEO
  • 00:12:02
    or AI and so that that is good to know.
  • 00:12:06
    So search historically has been Google's
  • 00:12:08
    cash code. It's a massive part of your
  • 00:12:10
    business. And so when you think of it,
  • 00:12:14
    it's not just a simple UX change. This
  • 00:12:16
    is really business transformation if
  • 00:12:19
    this entire experience is to evolve. And
  • 00:12:21
    we haven't been seeing ads on search in
  • 00:12:24
    the AI overview so far. Is Google
  • 00:12:26
    thinking through a way of monetizing AI
  • 00:12:29
    overviews. Should we be looking to see
  • 00:12:31
    ads in our conversations in the future?
  • 00:12:35
    We are doing experiments to understand
  • 00:12:37
    how ads could work in these experiences.
  • 00:12:39
    I think if you abstract away the product
  • 00:12:41
    again just think about information and
  • 00:12:43
    people's needs. People want to find
  • 00:12:45
    information and in many cases the ads
  • 00:12:47
    and sponsored content is as are more
  • 00:12:49
    helpful because getting an offer or a
  • 00:12:51
    specific way to get the thing you're
  • 00:12:53
    looking for is incredibly useful and
  • 00:12:55
    that that connection is valuable just in
  • 00:12:57
    the context of search today and also in
  • 00:12:59
    the context of these new AI experiences.
  • 00:13:00
    It's it's really no different there. And
  • 00:13:02
    so we are experimenting with ways that
  • 00:13:04
    advertising can work in these new
  • 00:13:06
    surfaces, but we've been really focused
  • 00:13:08
    on getting the particularly for AI mode
  • 00:13:10
    in the short term, building a great
  • 00:13:12
    experience that's working for people and
  • 00:13:14
    especially given it's a labs and it's
  • 00:13:15
    fairly early for us over time. Yeah. The
  • 00:13:18
    other big question that businesses and
  • 00:13:19
    users have is really our own trust. Can
  • 00:13:22
    they trust the systems? We have seen
  • 00:13:24
    hallucinations in the past. Why should I
  • 00:13:27
    trust an AI overview or a result that I
  • 00:13:30
    see in AI mode versus just traditional
  • 00:13:32
    search? And how is Google tackling
  • 00:13:34
    things like verifying content and
  • 00:13:36
    reducing hallucinations?
  • 00:13:39
    Yeah. So, these AI systems are new
  • 00:13:41
    overall in the world and everyone's
  • 00:13:43
    learning on how to best implement them
  • 00:13:45
    and make them as accurate and useful as
  • 00:13:46
    possible, including Google. And I think
  • 00:13:48
    that in our case, there's there's so
  • 00:13:51
    many interesting underlying signals that
  • 00:13:54
    can be used to back up information
  • 00:13:57
    because of this long history that I
  • 00:13:59
    think that we're uniquely able to build
  • 00:14:01
    something that is as accurate as
  • 00:14:02
    possible. Now, it's not to say that it
  • 00:14:04
    won't make mistakes. It will, and it
  • 00:14:06
    has, but for example, in something like
  • 00:14:08
    AI mode, what we've tried to do is
  • 00:14:11
    really train the models for accuracy. So
  • 00:14:16
    what'll happen is as it's learning in
  • 00:14:17
    the training process, it's effectively
  • 00:14:20
    punished for making these kinds of
  • 00:14:21
    mistakes. And we say bad model if you're
  • 00:14:24
    effectively like making a mistake or a
  • 00:14:26
    hallucination. And by the way, we say
  • 00:14:27
    it's especially bad if it's in certain
  • 00:14:29
    categories like if it's in health.
  • 00:14:32
    Whereas if maybe it has a piece of
  • 00:14:33
    information that's it's kind of like
  • 00:14:35
    blurry if it's accurate. It's like
  • 00:14:37
    pretty accurate. It's hard you maybe
  • 00:14:38
    debate it. So in in it actually with
  • 00:14:43
    this model it tries very hard to back up
  • 00:14:46
    what it says and so what we are finding
  • 00:14:47
    I think what your users are finding is
  • 00:14:48
    that it does have pretty high quality
  • 00:14:50
    responses and I think that's something
  • 00:14:52
    that Google does uniquely well now again
  • 00:14:55
    having said that these are new systems
  • 00:14:56
    it's going to make mistakes and I think
  • 00:14:58
    part of this test period is you actually
  • 00:15:00
    want to get these products out there and
  • 00:15:02
    I think a learning for me and many of us
  • 00:15:03
    building an in AI is that you need this
  • 00:15:05
    incubation period because you can't
  • 00:15:07
    really know just from internal testing
  • 00:15:09
    despite our best efforts to test as
  • 00:15:11
    rigorously as possible where because
  • 00:15:14
    there's so many degrees of freedom. You
  • 00:15:15
    can ask any question, any kind of
  • 00:15:16
    combination, you can follow up with it
  • 00:15:18
    in an infinite number of ways. You got
  • 00:15:20
    to get it out there and you have to get
  • 00:15:21
    some trusted testers really using it and
  • 00:15:23
    then very quickly diagnose problems and
  • 00:15:26
    fix it. And then it goes through this
  • 00:15:28
    process of maturation where I think what
  • 00:15:29
    you see over time is that these AI
  • 00:15:31
    systems will become increasingly
  • 00:15:33
    accurate and more dependable very
  • 00:15:35
    quickly. And I think that's what we've
  • 00:15:36
    seen as well. I love that. Now, one of
  • 00:15:40
    the things I've seen over the last year
  • 00:15:42
    from Google is this emphasis on learning
  • 00:15:45
    and working with AI models to figure out
  • 00:15:47
    how they can optimize things like
  • 00:15:48
    research and things like learn LM. You
  • 00:15:51
    mentioned this being a specially trained
  • 00:15:53
    model. Are you able to dive a little bit
  • 00:15:55
    deeper into the backend models that are
  • 00:15:57
    in the system? Are we using Gemini 2.5
  • 00:16:00
    as yet? Are we leveraging things like
  • 00:16:02
    deep research? And you mentioned
  • 00:16:04
    reasoning. Um, how are you balancing
  • 00:16:06
    like the chain of thought that might go
  • 00:16:09
    into test time compute that could take a
  • 00:16:12
    long time to generate a response versus
  • 00:16:13
    the immediiacy that people typically
  • 00:16:15
    expect from search?
  • 00:16:18
    Yeah, absolutely. So, this is a custom
  • 00:16:20
    model that's built on the Gemini 2
  • 00:16:22
    series, but it's been specifically
  • 00:16:24
    trained for informationational needs and
  • 00:16:26
    using Google search. And so what that
  • 00:16:28
    means is that as the model accepts a
  • 00:16:30
    query or someone's question, it's
  • 00:16:33
    learning the process of generating a
  • 00:16:35
    plan, figuring out what these fanout
  • 00:16:37
    queries are, finding information with
  • 00:16:39
    certain search signals, and then
  • 00:16:41
    rendering a response based on that. And
  • 00:16:43
    those are all things that have been
  • 00:16:44
    done. And by the way, the other thing is
  • 00:16:45
    it's learned to put attention on the
  • 00:16:48
    right kinds of information. So for
  • 00:16:50
    example, if you're asking for just an
  • 00:16:51
    immediate maybe sports score like we
  • 00:16:53
    have you type into Google right now any
  • 00:16:56
    sports score you get this one box at the
  • 00:16:57
    top pretty much immediately like there's
  • 00:16:59
    a bunch of really rich real-time
  • 00:17:01
    information just available the model
  • 00:17:03
    also has access to that and so has been
  • 00:17:05
    trained to understand and use that
  • 00:17:07
    information which makes it even more
  • 00:17:08
    powerful and accurate for many of these
  • 00:17:10
    questions and so that's really what goes
  • 00:17:12
    into the custom training and then we do
  • 00:17:14
    a bunch on reinforcement learning and
  • 00:17:16
    agentic reinforcement learning where the
  • 00:17:18
    model actually learns this reasoning
  • 00:17:20
    process to produce highquality accurate
  • 00:17:22
    and factual responses and is rewarded
  • 00:17:24
    for that obviously and if it makes
  • 00:17:25
    mistakes it's punished and then you
  • 00:17:28
    hopefully like through that process it
  • 00:17:30
    learns to do a pretty good job most of
  • 00:17:31
    the time. Now in terms of looking ahead,
  • 00:17:34
    what should users expect in terms of
  • 00:17:36
    just AI in search and the search
  • 00:17:38
    experience evolving overall? Are they to
  • 00:17:41
    expect that AI mode or its descendants
  • 00:17:45
    will become the main sort of search
  • 00:17:47
    experience or is Google still thinking
  • 00:17:49
    of maintaining traditional search and
  • 00:17:51
    having AI search live alongside it?
  • 00:17:55
    Yeah, the good news is there's already
  • 00:17:56
    for in AI mode in this lab, there's
  • 00:17:59
    already a really nice connection here
  • 00:18:00
    where if you search for something and it
  • 00:18:03
    generates an AI overview, which was
  • 00:18:04
    where we said we found that AI is most
  • 00:18:07
    helpful for that use case, you can
  • 00:18:08
    actually follow up in AI mode from that
  • 00:18:10
    and then ask follow-up conversations and
  • 00:18:12
    go deeper. For a user who's just not
  • 00:18:14
    thinking much, they put a question in to
  • 00:18:16
    Google. That is a really easy and fast
  • 00:18:19
    way to get more information, go deeper,
  • 00:18:21
    and have a deeper journey on on Google.
  • 00:18:24
    And then for a power user who might know
  • 00:18:25
    that they have a really hard thing,
  • 00:18:27
    there's also ways that you can go right
  • 00:18:29
    to AI mode and just start asking that
  • 00:18:31
    question and going back and forth. And I
  • 00:18:32
    think that choice and that those options
  • 00:18:34
    are important to maintain, especially
  • 00:18:36
    during the fact that these are new
  • 00:18:38
    things and we're learning, but also
  • 00:18:39
    people have different needs and we've
  • 00:18:41
    found that is a really important thing
  • 00:18:42
    to offer that kind of choice to users.
  • 00:18:45
    So we're still in the early phases here,
  • 00:18:47
    but I think that our goal is that this
  • 00:18:48
    is as effortless and easy as possible. I
  • 00:18:50
    don't think we want people to worry too
  • 00:18:52
    much about where you use Google. You
  • 00:18:54
    just put your question into Google. You
  • 00:18:56
    get this AI is helpful AI overview if it
  • 00:18:58
    feels like that's the most helpful.
  • 00:18:59
    Maybe you get this grid of images if it
  • 00:19:01
    feels like that's the most helpful. And
  • 00:19:03
    then you can follow up and go deeper in
  • 00:19:04
    AI mode. And that's a really simple
  • 00:19:05
    thing. And I think we want it to also be
  • 00:19:08
    simple across modalities. So whether
  • 00:19:10
    it's a a photo or a video or audio or
  • 00:19:13
    text, like it should be as easy and
  • 00:19:14
    seamless as talking to a person just
  • 00:19:16
    going back and forth. And I think that's
  • 00:19:18
    really one of the big unlocks with AI is
  • 00:19:20
    that you don't have to have this keyword
  • 00:19:22
    ease framework. You can just ask really
  • 00:19:24
    whatever you want and that's why it's so
  • 00:19:27
    expansionary and we could do so much
  • 00:19:29
    more for you now given that interface. I
  • 00:19:31
    love that and I I love the fact that
  • 00:19:34
    you're still retaining this principled
  • 00:19:36
    approach of solving a user's problem. I
  • 00:19:39
    think so many people looked at AI and
  • 00:19:40
    they thought how can AI be applied to
  • 00:19:43
    search and I think it sounds from how
  • 00:19:46
    you're explaining it that Google kept
  • 00:19:48
    saying okay what is the user trying to
  • 00:19:49
    do and how can we best do that and if AI
  • 00:19:52
    overviews is that solution if AI mode is
  • 00:19:55
    that solution then we'll present that
  • 00:19:56
    and we'll see how the product evolves to
  • 00:19:58
    to keep on solving that problem I think
  • 00:20:00
    that will that that sort of northstar is
  • 00:20:02
    super helpful all right so the last set
  • 00:20:04
    of questions that I want to look at are
  • 00:20:06
    really just forward facing When you
  • 00:20:09
    think of search in the next 3 to 5
  • 00:20:11
    years, given the pace of AI that you're
  • 00:20:14
    seeing today, personally, what are you
  • 00:20:16
    thinking? What are you looking forward
  • 00:20:18
    to the most? Are there any specific use
  • 00:20:20
    cases that the current AI mode doesn't
  • 00:20:23
    really solve well that you think the
  • 00:20:25
    evolution of the product you're hoping
  • 00:20:27
    that it will be able to tackle better?
  • 00:20:30
    Yeah, there's so much. It's dizzying at
  • 00:20:32
    times to consider all the things that it
  • 00:20:34
    could do, but I think for me the reality
  • 00:20:36
    is it's about just expanding what's
  • 00:20:38
    possible so that you can get really the
  • 00:20:42
    value of reasoning and information for
  • 00:20:43
    any of your problems. And I think if you
  • 00:20:45
    think about what's next, there's so much
  • 00:20:48
    information just in the world and in all
  • 00:20:50
    kinds of different areas. And
  • 00:20:51
    increasingly some of that information
  • 00:20:52
    may actually be in other AI systems and
  • 00:20:54
    they're all AI is starting to talk to AI
  • 00:20:56
    and collaborate and do really
  • 00:20:58
    interesting things like that. And so
  • 00:21:00
    obviously at Google search it's an
  • 00:21:01
    information system predominantly. And so
  • 00:21:03
    I think there's a lot I'm very intrigued
  • 00:21:05
    with how it can expand what is what
  • 00:21:07
    information is available to it to be
  • 00:21:09
    most helpful. Whether that's information
  • 00:21:11
    that might not be available as readily
  • 00:21:13
    today or information that you have that
  • 00:21:15
    you want to help augment with. there
  • 00:21:16
    seems to be a lot of opportunities to
  • 00:21:19
    really just make this do more for you. I
  • 00:21:22
    think the other area is these systems. I
  • 00:21:24
    think people come to Google looking to
  • 00:21:26
    do something like they're not always
  • 00:21:27
    just looking for information. It's part
  • 00:21:29
    of a journey. And so we have a lot of
  • 00:21:31
    opportunities to help actually help you
  • 00:21:34
    go further with these areas. We actually
  • 00:21:35
    launched a a service that is called
  • 00:21:37
    Bismatch that actually helps connect to
  • 00:21:39
    businesses but through phone calls where
  • 00:21:42
    an AI can help book and make
  • 00:21:44
    reservations for you. That's an area
  • 00:21:46
    that I feel like is really exciting
  • 00:21:47
    because that really ties together the
  • 00:21:49
    whole journey of what people are really
  • 00:21:51
    trying to do. And that should be so
  • 00:21:52
    easy. It should be effortless. You
  • 00:21:54
    really just should be able to explain
  • 00:21:55
    what you're trying to get done. And like
  • 00:21:56
    I said, it makes a plan. It finds
  • 00:21:58
    information. It makes stuff happen. And
  • 00:22:00
    I think increasingly that's the
  • 00:22:01
    expectation for people is that you have
  • 00:22:03
    these superpowers now that you really
  • 00:22:04
    couldn't do yourself as easily. And
  • 00:22:06
    that's going to and I think it's going
  • 00:22:07
    to be great. I think obviously it's not
  • 00:22:08
    part of the core Google search
  • 00:22:10
    dayto-day, but there's so many I'm
  • 00:22:12
    personally very excited for scientific
  • 00:22:14
    discovery and drug discovery and we're
  • 00:22:16
    starting to see like early progress of
  • 00:22:18
    almost like AI scientists that can do
  • 00:22:20
    breakthroughs, collaborate and reason
  • 00:22:22
    and just the potential. I'm an optimist
  • 00:22:25
    on this. I think that the potential for
  • 00:22:27
    solving world problems and people's
  • 00:22:29
    problems is just absolutely enormous.
  • 00:22:31
    There's a whole group of people that I
  • 00:22:32
    respect and admire working on that side
  • 00:22:34
    of the problem and I'm personally very
  • 00:22:35
    excited to see that evolve in the next 5
  • 00:22:38
    to 10 years. I agree. I'm excited about
  • 00:22:40
    education and learning which is why I'm
  • 00:22:42
    particularly intrigued about what Google
  • 00:22:44
    is doing. I've seen all the products
  • 00:22:46
    notebook LM deep research I think is
  • 00:22:49
    going to be fascinating. I have two kids
  • 00:22:50
    also and the prospects of them being
  • 00:22:53
    able to have personalized AI tutors that
  • 00:22:56
    can help them to digest information
  • 00:22:58
    quickly and process in a way that they
  • 00:23:00
    probably can't do in a traditional
  • 00:23:02
    learning environment. I think that's so
  • 00:23:04
    fascinating. I think that's a great one.
  • 00:23:05
    It's absolutely one of mine too. But I
  • 00:23:07
    think like I'm a I love to learn and I
  • 00:23:10
    always find that it's incredibly hard
  • 00:23:11
    because you either need to read a
  • 00:23:13
    textbook and it's you can't really ask
  • 00:23:15
    many questions of the textbook. Yeah.
  • 00:23:17
    And then you also want to try doing it
  • 00:23:18
    yourself but then when you try doing it
  • 00:23:19
    yourself you can't really get help doing
  • 00:23:21
    it because no one else is with you and
  • 00:23:24
    with an AI that has access to all of us.
  • 00:23:25
    By the way, that was one of the things I
  • 00:23:26
    was mentioning in terms of access to
  • 00:23:28
    different types of information. like AI
  • 00:23:30
    had access to that quality of
  • 00:23:31
    educational content and then it
  • 00:23:33
    obviously has access to seeing and
  • 00:23:35
    talking to you in the same way a person
  • 00:23:37
    does. I think you can do some pretty
  • 00:23:38
    magical things. Let me ask this before
  • 00:23:41
    we wrap up. Is Google worried or
  • 00:23:43
    thinking about product cannibalization
  • 00:23:46
    meaning we see tools like Gemini
  • 00:23:49
    becoming increasingly popular, becoming
  • 00:23:51
    increasingly the UI that people want to
  • 00:23:55
    interact with the web through. Do you
  • 00:23:57
    see a future where Gemini becomes that
  • 00:24:00
    default interface and Google search as a
  • 00:24:02
    standalone product almost disappears? I
  • 00:24:05
    think people I really have started to
  • 00:24:06
    appreciate this much more deeply since
  • 00:24:08
    over the last year or so working in
  • 00:24:10
    trying to bring AI into search in this
  • 00:24:12
    way. There is a just an enormity of
  • 00:24:14
    needs that people have in the world
  • 00:24:16
    every day. People want a specific piece
  • 00:24:18
    of information. They want to know a
  • 00:24:19
    phone number. They want to go pay their
  • 00:24:20
    tax bill. they want to and you really
  • 00:24:23
    need to make sure the information is in
  • 00:24:24
    the format that people need and it's as
  • 00:24:26
    efficient and fast as possible and in
  • 00:24:28
    some cases we find that something like
  • 00:24:30
    an AI overview is like really efficient
  • 00:24:32
    and helpful for people and maybe if you
  • 00:24:33
    combine that with AI mode you have this
  • 00:24:35
    natural back and forth that can occur
  • 00:24:36
    and for other cases it may not be
  • 00:24:38
    actually the best and we see that from
  • 00:24:40
    users and I think you see some of these
  • 00:24:41
    more exploratory questions people have
  • 00:24:43
    as an example where it wants to be more
  • 00:24:45
    visual you want to understand what's out
  • 00:24:47
    there see what people are saying click
  • 00:24:48
    through it to a different hey here's
  • 00:24:50
    like some videos on the topic. Here's
  • 00:24:51
    some cool web articles on the topic. AI
  • 00:24:53
    systems aren't as dialed on those kinds
  • 00:24:55
    of on those kinds of things. And I'm
  • 00:24:56
    mostly of a believer that that the world
  • 00:24:58
    is pretty vast and that there's going to
  • 00:25:00
    be many ways and formats that people
  • 00:25:02
    interact to get
  • 00:25:04
    information. And I do think that this is
  • 00:25:06
    one though where as we're adding the
  • 00:25:08
    this is a growing one clearly and as we
  • 00:25:10
    build these some products like AI
  • 00:25:12
    overviews and AI mode like we are seeing
  • 00:25:14
    people use search more and wanting to
  • 00:25:16
    use it more. I think that we're going to
  • 00:25:18
    he's going to be a key part of what we
  • 00:25:19
    do, but I don't think it's going to be
  • 00:25:20
    all we do. Awesome. Robbie, thank you so
  • 00:25:23
    much for joining us. I thought it was
  • 00:25:25
    super helpful. I learned a lot. I'm sure
  • 00:25:27
    our listeners learned a lot. If users
  • 00:25:29
    want to try AI mode or to dive deeper
  • 00:25:32
    into AI overviews, can you just wrap up
  • 00:25:34
    by telling them how they can access it?
  • 00:25:37
    Yeah, just Google AI mode and then
  • 00:25:39
    you'll go to the Google Labs page and
  • 00:25:41
    you can turn it on and then you'll be
  • 00:25:43
    able to access AI mode on your next time
  • 00:25:46
    using Google. And the AI mode is
  • 00:25:47
    currently limited for just for the it's
  • 00:25:49
    in the US for US specific right now. So
  • 00:25:52
    if you're in the US, you can turn that
  • 00:25:53
    on and obviously we're looking to learn
  • 00:25:55
    quickly of how it's going and as we can
  • 00:25:57
    bring it to more people. But we're
  • 00:25:59
    looking to just make something great and
  • 00:26:00
    make sure that it's working well and
  • 00:26:02
    people are loving it. It's a fascinating
  • 00:26:04
    product. I'm looking forward to tracking
  • 00:26:05
    and seeing how it progresses. Robbie,
  • 00:26:07
    awesome. Thank you. This was an amazing
  • 00:26:09
    conversation. Thank you so much for
  • 00:26:10
    joining us. I'm looking forward to
  • 00:26:11
    having you back on sometime. Appreciate
  • 00:26:13
    that. And thank you so much. It was
  • 00:26:14
    great to be here. All right. Take care.
  • 00:26:15
    See you soon. All right. Bye-bye.
Tags
  • AI
  • Google Search
  • AI Overviews
  • AI Mode
  • Search Technology
  • Content Strategy
  • User Experience
  • Trust in AI
  • Gemini
  • Query Fan Out