For Hema Budaraju, Senior Director Product Management at Google, adding AI to the scale of search, with billions of queries in a day, means unlocking new potential, including opportunities for voice and visual search. In fact, Google is seeing the use of AI search features explode in India with Google Lens usage doubling in just 18 months. Google is also seeing that 15% of the searches are brand new and never seen before. “The more we expand the types of questions that search can answer, the more people expect that we can actually do better in serving them. So part of our responsibility as Google is to keep pace with these user expectations and technology advancements. Both of these are constantly changing,” she tells indianexpress.com on the sidelines of Google For India. Budaraju adds that with Gen AI and Google Gemini models, we are on a journey of reimagining search all over again. Along with the new questions “that people never thought search could answer”, she says, it is also transforming the way information is organised. “We can help you make sense of what's there, connecting people with a combination of insights, like when you think about complex questions, things with nuance, questions with perspective, especially a wide range of perspectives and human voices, like people with shared experiences and information from authoritative sources. When you put all that together, this is where Gen AI really shines, to really bring like a jumping off point for you on search and creating the right touch points so that we can send valuable traffic to publishers and supporting a healthy open web.” She added: “We launched a language toggle buttons to help users easily switch between English and Hindi. People in India are power users of voice search. So when bringing AI overviews to India, we designed a text-to-speech capability to customise the experience for Indian users.” Google’s testing has found that people prefer using search with AI overviews and find search results more helpful with them. In India, users listen to AI overviews more often than in any other country. “We find that people are more likely to spend time on the sites that they visit (from AI overviews), which is an indication of higher quality traffic, to the content creator or the publisher,” she explained. Now, Google will be expanding AI overviews to Tamil, Telugu, Bengali and Marathi. Budaraju said that soon users will be able to use Google Lens to search using video and ask questions. “It is as simple as that: take a video, ask a question. Gemini model will make sense of that video and the question that you asked, and will give you a helpful AI overview with links to learn more,” she said, adding that the feature will first be available in Google Search Labs. Edited excerpts from the interview: At a basic level, what percentage of search do you think will get the AI overviews? Will it be for every possible query, or will you not use AI on certain types of queries? Hema Budaraju: We believe the role of search is really helping take the work out of searching and give you the best experience possible. And sometimes the search results page is perfect for that and you don't need anything else. AI overviews show up when we believe there is additional value that we can bring to on top of the search results page. Where does this additional value come from? It comes from these complex queries, perspectives, nuanced questions, things that you might not just find the answer to through the search results, and you might want a little starting point and a little bit of a frame around it. Yes, we are intentional about when we show AI overviews. Then we show AI overviews when we believe that the quality of that response is the best, and quality is not static. We also do not show you AI overviews for certain categories. For example, when something has like freshness at the centre of it, we are not necessarily sure that AI overviews has all the context of everything that's happening within the last few hours or minutes. How do you filter out the best results for AI overviews, because that is what the user is expecting? Hema Budaraju: Information quality is core to search and we are constantly refining these systems to deliver reliable, helpful and high-quality information. AI overviews are rooted in the same quality systems that power our search. AI overviews apply LLM that has been purposefully trained to carry out tasks specific to search, like identifying relevant, high-quality results from our index to support the info in the AI overview and a customised Gemini model, combined with the search for ranking systems, is really key to this. Our tests show that our accuracy rate for AI overviews is on par with one of our most popular features in search, which is Featured Snippets. Now quality doesn't mean that we have done it all, and there is no room to grow. Should a report come to us when we do get it wrong, we actually have the wherewithal to immediately address it and take it down. We do inordinate amount of adversarial testing, we try to break our system in advance so that we can prevent a whole bunch of other downstream stuff that can happen. So if you put that package together, you'll actually see that AI overviews are like we inherit the principles of search and we are part of the search. Is there a conflict happening within Google when you present AI overviews because with it the chances of people clicking on something and ending on a page which also has Google ads goes down significantly? How do you balance this business part? Hema Budaraju: Fundamentally, users, people who use our product, are front and centre. When we actually talk to people, people prefer search pages with AI overviews for two reasons: the first is we take the work out of searching, and two, they really like the diversity of sources that we have. Users are telling us that these are the two things that make the most sense, then we try and optimise to those two things that people are looking to us towards for. Business will follow, right? There has always been a dearth of good content in smaller Indian languages. With AI is not necessary anymore to have that knowledge in a particular language. How is that helping you with search? Hema Budaraju: There are two parts to your question. One is, whether is there more corpus in English than in other languages; and therefore, like the narrower the language slice, the smaller the corpus. Therefore our ability to bring you the best of the web becomes limited. I actually think this is a fundamental value of LLMs, where we can say here is like the whole of the web at your fingertips with search principles custom trained on a Gemini model that can bring the best of information. I think it's actually a good time to be on a search to think about these problems. And as we start thinking more and more about languages, a lot of these considerations are factors. You mentioned the searches are now different. And as you go into multimodal, I think it will again be different from your learnings so far. How do you evolve for that? Hema Budaraju: You would be surprised as to how natural this has become. It's voice and video, and you are receiving an input and output that is actually a great overview that can also be read or heard. There's a lot going on. One of the reasons why I mentioned this is coming to Labs is do we understand how is that usage? Did we actually get the right form factor? Did we get the right intent? Our labs will continue to be our innovation test bed right now because a lot of the questions that you are asking are things that we will learn too. And the thing that I would emphasise is that we have a huge responsibility, and it's a privilege, to bring these kinds of technologies to really unlock any kind of query in any way possible, text, voice, gesture, radio. We would like to do that responsibly.
