Anyone watching Game 1 of the NBA finals between the Dallas Mavericks and Boston Celtics Thursday evening might have been slightly perplexed by an advertisement for something called Perplexity. But look closer and it could offer a glimpse into the future of how we find information online.

Perplexity is a San Francisco company valued at more than $1 billion that has picked a fight with Google, worth more than $2 trillion, to challenge its supremacy in internet search.

That would have seemed like a hopeless task before the release of powerful artificial intelligence programs able to produce and summarize text like a human — quickly and mostly accurately. The best known of those programs, OpenAI’s ChatGPT, was released in November 2022. Perplexity was founded the following month.

The startup is betting on technology called generative search: using AI to answer a question by crawling the best pages it can find online and then summarizing them for a curious user. If that approach turns out to be superior to traditional search, and a startup like Perplexity beats Google to perfecting it, it could claim the giant’s crown jewel and biggest revenue source.

“This has been the holy grail for search,” said Krishna Gade, a former Facebook engineer who also worked at Microsoft on the Bing search engine and is now the CEO of startup Fiddler AI, which makes tools to improve the safety and trustworthiness of AI.

“We think about this as ‘people have questions and we’re trying to answer them,’ ” instead of pointing them to places where the answer might be, said Dmitry Shevelenko, Perplexity’s chief business officer. Instead of a search engine, “we call ourselves an answer engine,” he said.

With his company now seeing hundreds of millions of queries a month, he said, Shevelenko has reason to be confident. But Google is not resting on its links. In its search results, Google has long offered magnified snippets pulled from relevant documents to give users quick answers to their questions. And it began rolling out some generative search results a year ago.

So what is generative search, exactly? How does it work? And most importantly, can it be trusted?

Instead of sniffing out keywords to return the best links — as Google’s search engine has for decades — generative search can understand concepts more broadly to get better results and then summarize them into a fluent answer for a user, said Gade, the Fiddler CEO.

For example, when a person asks “What are the latest cars on the market?” the program knows that car means Tesla or BMW or Rivian, and will collect a range of pages based on that understanding, Gade said.

Using that broader knowledge of how information is connected, like in our own brains, rather than just matching keywords, the program ranks the documents it finds, summarizing the ones it thinks are the most relevant, he said.

Another example would be asking Google for information about Perplexity’s valuation. The search engine thinks for a moment and then its bot generates a few paragraphs of text explaining what the startup is worth and about its latest funding round, with the main documents where it found the information summarized just below.

In this case the information came from Bloomberg and tech news website Fast Company — two reliable sources — with others listed beneath.

One disruptive effect of generative search is its potential financial impact on suppliers of information. Google’s traditional list of links send users on a mission to hunt down what they’re looking for. But if “crawler” bots pull information from those pages and then summarize it for users, that does nothing to drive traffic to the original sources — and undermines the concomitant advertising revenue those visits might generate.

Shevelenko said he couldn’t share exactly what percentage of users clicked through from Perplexity to source documents, but said it was in the double digits and varied across web and mobile searches.

And, of course, sometimes generative search fails to find answers — or produces comically wrong results. Since Google started showing more such results to users, summaries have recommended adding glue to pizza recipes or confidently responded to questions about how many rocks a person should eat when asked, among other bizarre answers.

A blog post by Google VP and head of search Liz Reid argued that in those cases the technology was returning results for questions that didn’t include much high quality data. People had not really been asking about eating rocks until the screenshots of that search went viral, Reid said, “so when someone put that question into Search, an AI Overview appeared that faithfully linked to one of the only websites that tackled the question.”

Other results tracked to what Reid called “sarcastic or troll-y content from discussion forums,” and returned the best results it could. The company is disabling the generative search feature on some questions, and trying to make it better at recognizing non-serious questions, but the feature is still not 100% reliable.

Given that imperfection, “more and more (AI) tools should say ‘I’m not sure,’ ” said Matei Zaharia, a UC Berkeley professor and computer scientist, and the co-founder and chief technical officer at San Francisco tech firm Databricks.

That’s because while generative search programs excel at summarizing their findings, “they’re not good at reasoning,” Zaharia said, such as whether a result is a joke or what it means when there aren’t a lot of high quality pages for it to summarize.

Shevelenko of Perplexity said his company’s model is trained to say it doesn’t know when it can’t find enough high quality information, or when a question violates its terms of use. “If there isn’t a good answer we acknowledge that,” he said.

The bigger problem with summarizing the best information out there is that it’s impossible to ensure everything in a given article being summarized is 100% accurate, he noted.

“Many questions don’t have perfect answers to them,” Shevelenko said. That’s why Perplexity’s bot has features such as automatically asking the same question multiple ways to source from a variety of pages.

When and whether this new kind of search will totally and reliably overtake keyword searches that have been with us in some form since the 1990s is difficult to say. But if companies like Perplexity can get it right, “keyword search will become the backend,” said Gade, the Fiddler AI CEO. In that world, he said, Google’s blue links could be gone forever.

Reach Chase DiFeliciantonio: chase.difeliciantonio@sfchronicle.com; Twitter: @ChaseDiFelice