We are at a new inflection point. The biggest change to the way search engines have delivered information to us since the 1990s is happening right now. No more keyword searching. No more sorting through links to click. Instead, we’re entering an era of conversational search. Which means instead of keywords, you use real questions, expressed in natural language. And instead of links, you’ll increasingly be met with answers, written by generative AI and based on live information from all across the internet, delivered the same way.
Literal trillions of dollars are at stake in the outcome as these players jockey to become the next go-to source for information retrieval—the next Google.
People are worried about what these new LLM-powered results will mean for our fundamental shared reality. It could spell the end of the canonical answer.
I think it’s an exciting moment where we have obviously indexed the world. We built deep understanding on top of it with Knowledge Graph. We’ve been using LLMs and generative AI to improve our understanding of all that. But now we are able to generate and compose with that.
If you want to know someone’s darkest secrets, look at their search history. Sometimes the things people ask Google about are extremely dark. Sometimes they are illegal. Google doesn’t just have to be able to deploy its AI Overviews when an answer can be helpful; it has to be extremely careful not to deploy them when an answer may be harmful.
But perhaps the greatest hazard—or biggest unknown—is for anyone downstream of a Google search. Take publishers, who for decades now have relied on search queries to send people their way. What reason will people have to click through to the original source, if all the information they seek is right there in the search result?
OpenAI insists it’s not really trying to compete on search—although frankly this seems to me like a bit of expectation setting. Rather, it says, web search is mostly a means to get more current information than the data in its training models, which tend to have specific cutoff dates that are often months, or even a year or more, in the past.
The model, whether it’s OpenAI’s GPT-4o or Google’s Gemini or Anthropic’s Claude, can be very, very good at explaining things. But the rationale behind its explanations, its reasons for selecting a particular source, and even the language it may use in an answer are all pretty mysterious. Sure, a model can explain very many things, but not when that comes to its own answers.
The search results we see from generative AI are best understood as a waypoint rather than a destination. What’s most important may not be search in itself; rather, it’s that search has given AI model developers a path to incorporating real-time information into their inputs and outputs. And that opens up all sorts of possibilities.
This is the agentic future we’ve been hearing about for some time now, and the more AI models make use of real-time data from the internet, the closer it gets. Let’s say you have a trip coming up in a few weeks. An agent that can get data from the internet in real time can book your flights and hotel rooms, make dinner reservations, and more, based on what it knows about you and your upcoming travel—all without your having to guide it. Another agent could, say, monitor the sewage output of your home for certain diseases, and order tests and treatments in response. You won’t have to search for that weird noise your car is making, because the agent in your vehicle will already have done it and made an appointment to get the issue fixed.
But you can imagine things going a bit further (and they will). Let’s say I want to see a video of how to fix something on my bike. The video doesn’t exist, but the information does. AI-assisted generative search could theoretically find that information somewhere online—in a user manual buried in a company’s website, for example—and create a video to show me exactly how to do what I want, just as it could explain that to me with words today.