This Startup, Focused on AI Search, Now Has to Stare Down Google and OpenAI

When I was in junior high, I needed to know who the secretary of state was for a homework assignment. Today, you can just Google it. Back then, you had to be more industrious.

My stepfather, always keen to teach a lesson, told me to call the library and ask for the reference librarian, who would tell me the secretary of state was — and he mouthed the answer. I read his lips and exclaimed “Warren Christopher!” but he made me call the library anyway, thinking I was picking up a valuable life skill.

Soon enough, the role of the reference librarian might be played by generative AI, and more specifically gen AI-powered search. It’s a field that has the full attention of the heavyweights in artificial intelligence.

Both search giant Google and ChatGPT maker OpenAI have their own visions for the future of search, called AI Overviews and SearchGPT, respectively. They seek to provide more direct responses via AI-generated summaries. This ideally results in faster answers than scrolling. As of June, AI Overviews automatically appeared at the top of approximately 8% of all Google search results. Meanwhile, SearchGPT is reportedly being tested by some 10,000 users.

AI Atlas art badge tag

There’s also a  startup called Perplexity.ai, which quickly made waves in the tech sector with its novel blend of AI and search. In his review, my colleague Imad Khan said Perplexity’s use of the open web and ability to pull from social media platforms like Reddit and X make it a competitive alternative to Google Gemini and ChatGPT 3.5. However, he said AI startup Anthropic’s Claude model “feels better tuned to give more nuanced answers with greater informational synthesis.” 

At the end of the day, these companies are all after the same thing: A stake in the search engine market projected to reach $430 billion by 2032.

‘A large language model as a summarizer’

San Francisco-based Perplexity refers to itself as an AI answer engine. Instead of having you click through links as in traditional search, Perplexity finds what it calls trustworthy sources, identifies relevant facts from those sources and combines them into answers with citations to the sites where the information came from, which makes fact-checking easier.

It’s a lot like OpenAI’s SearchGPT, but Chief Business Officer Dmitry Shevelenko prefers to think SearchGPT was inspired by Perplexity, which rolled out to the public in December 2022. (OpenAI just unveiled SearchGPT in July, a couple months after Google’s troubled introduction of AI Overviews.)

He distinguishes the two by noting that Perplexity is focused on answering questions while SearchGPT is focused on chat. But that may be six of one and half dozen of the other.

Three of Perplexity’s four co-founders have doctoral degrees. According to Shevelenko, this experience in academia “where the currency of the realm is citations”led to Perplexity in its current form.

“This idea of using a large language model not as a source of information but as a synthesizer and summarizer of information and then using high-quality, trustworthy sources on the real-time internet as the base of knowledge, that was kind of like the aha moment for them,” Shevelenko said.

(A large language model is an AI model that uses machine learning to understand text. It is trained to predict future words based on previous words. [Machine learning is a branch of AI that uses data and algorithms to imitate the way humans learn.])

To do so, Perplexity created its own web index and ranking system. Shevelenko said it uses 50 data signals to determine the trustworthiness and reputation of domains. (By way of comparison, Google has at least 200 ranking signals.)

“Things that will always be reputable [are] news organizations,” he added. “The facts that are embedded inside of their articles will always have priority, and that gets reflected in many data signals we look at.”

The startup has had a complicated relationship with those news organizations. Earlier this year, Wired and Forbes said Perplexity was using their content without permission. Perplexity CEO Aravind Srinivas denied the allegations. One month later, the startup unveiled an initiative called the Perplexity Publishers’ Program, which includes a revenue-sharing component for media partners like Time, Der Spiegel and Fortune.

Perplexity offers two plans for consumers. A free version lets you do unlimited “Quick” searches and five “Pro” searches per day. The $20-per-month Pro plan lets you do more than 300 Pro searches a day.

According to a Perplexity blog post, the Pro version, which came out in July, can answer more complex queries, perform advanced math and programming tasks, and generate more thoroughly researched answers.

Perplexity also offers enterprise subscriptions and an API — which developers use to build new applications based on Perplexity technology — and it plans to launch advertising in the third quarter. The startup is developing an ad unit called Sponsored Question, which will allow brands to sponsor follow-up questions in search results.

“We anticipate within the next, say, 18 months, that will become our largest source of revenue,” Shevelenko said of advertising.

He declined to share user numbers but said Perplexity answered 250 million questions in July. That’s still a drop in the bucket compared with you-know-who.

Google doesn’t share search volume data, but it’s estimated the search giant gets 8.5 billion searches daily, which translates to 255 billion searches monthly.

Where did the phrase holy cow come from?

As I was writing this, I talked with an SEO friend about ranking signals and he said Google’s official number is 200, but a leak of internal documents earlier this year suggested there could be as many as 8,000.

“Holy cow,” I said. And then I wondered where that phrase came from. So I went to Perplexity.

Citing moving company Square Cow Movers, answer site HowStuffWorks and news site Indian Express, Perplexity told me the phrase dates back to at least 1913 when baseball announcers used it to avoid swearing. But, Perplexity continued, it could also be a variation of “Holy Christ!” and/or influenced by Irish immigrants who said, “Holy cathu.” The word cathu refers to sorrow in Gaelic and sounds like cow in English.

Google’s AI Overview, meanwhile, cited Wikipedia and the same Indian Express article, along with a YouTube video from a creator called English with Jackie. It highlighted the cathu explanation first.

In a regular Google search, a Wikipedia entry is the top search result for the phrase in Google, but that’s followed by Square Cow Movers, social forum site Reddit, Q&A site Quora, HowStuffWorks and Indian Express.

I was honestly surprised to see so much overlap in sourcing.

When I asked, “When will the Braves play tonight?” both Perplexity and Google told me they would be playing the Los Angeles Angels at 9:38 ET. Perplexity cited broadcaster website CBSSports.com, while baseball site MLB.com has the top ranking in Google, followed by the New York Times and CBSSports.com.

A search for recipes with tomatoes perhaps illustrated Perplexity’s limitations as it highlighted a single recipe for panzanella while Google had much more variety.

Looking ahead, Perplexity wants to add more visual elements to its search results, Shevelenko said.

The company has raised $165 million to date, including $63 million in a fourth round earlier this year.

Shevelenko confirmed Perplexity has raised “over $100 million” and has “no specific plans” for more.