Sandberg had to quickly assemble a human wave of screeners to zip through thousands of ads for instant review. She contacted a temp agency, which sent over fifty people. She also pulled people from other areas of Google on an emergency basis. Still, Google’s software for approving ads and putting them into the approval bin wasn’t built to handle such volume. Sandberg told everyone to stop and go home. “Over the weekend, our ads engineering team built us a new approval bin,” she says.
The job was trickier than anyone thought—to do it right, the screeners actually needed deft judgment to quickly determine whether thumbs would be up or down on each ad. Two weeks later, only one of the fifty workers was deemed worth keeping. Sandberg found some better suppliers. One was a specialized temp agency that had done some hiring for Microsoft. Another was craigslist, which was just emerging. Both had access to unemployed victims of the recent tech bust. “We could hire all these graduates of the Ivy League and great state schools to be temps,” says Sandberg.
Later, Google figured out how to scale the process by using better algorithms and data. “Google does a hundred thousand ads a day, and most of it’s automated,” Sandberg later said. “We had to get fast and good because we grew so quickly.”
But the ad policy had more implications than just pleasing AOL—it turned out that Page and Brin had their own ideas about what was proper in an ad. When Alana Karen arrived at Google in late 2001 and volunteered to work on ad policy issues, she found only a rudimentary set of rules, such as a ban on pornography. But she learned that Page and Brin were concerned that Google ads uphold their “make the world better” standard. That presented tricky problems. Products such as tobacco were clearly out. Also, Google started to ban liquor ads. Then came the discovery that some advertisers were offering wine or other soft liquor in gift baskets. That didn’t seem wrong. So the policy was amended to allow beer and wine. Later Google realized that it was appropriate to have different standards in different countries. For instance, in Japan hard-liquor advertising was more culturally acceptable in mass media. Eventually Google figured out a way to balance its corporate conscience with the concept of running ads that didn’t meet a standard of healthy living. In 2003, Alana Karen took charge of a program called Google Grants, which gives free ads to nonprofit, socially beneficial organizations. “It’s like carbon offsets,” she explains.
Google came to see the AOL deal as a tipping point. Before AOL, it had a limited inventory of searches that would be relevant to a given keyword—if you were selling ski equipment in early 2002, your ads would run only in response to people who typed a keyword that involved winter sports. But after AOL the number of pages multiplied, so the inventory of popular terms almost always met the demand.
No one was sure how advertisers would respond when their opportunities were just about infinite; Google now could give them more customers for keywords than they could possibly budget for. Yet money spent for Google ads seemed well spent. “We didn’t know how valuable increased inventory would be—what happens if you double it overnight?” says Wojcicki. “It turns out that advertisers will keep using it up.”
In July 2003, Yahoo bought Overture for $1.63 billion, sending shockwaves through the Googleplex. Overture’s ad technology would be linked to an effective search engine—and perched on the world’s biggest portal. In addition, Overture had an active lawsuit going against Google. Even though Overture had failed to nail down the patent for the core of its ad system, it claimed that Google was infringing on the “obscure, silly patents” (in Bill Gross’s words) it did own.
Google’s biggest fear was that Yahoo would begin innovating with Overture and improve its system to Google’s level. Yahoo had already decided to replace Google’s search engine with its own system. Its then CEO, a former Hollywood executive named Terry Semel, recalls that after that announcement, Page and Brin came to his office and told him the two companies were now at war. Semel was amused. “Are you going to bomb us?” he asked. Semel knew that there were profits to be made even as a runner-up to Google. But Yahoo never figured out how to innovate with Overture.
“We used to benchmark ourselves against Overture,” says David Fischer, a former Google ad executive who worked under Sheryl Sandberg. “But at some point Sergey just said, ‘Why are we paying attention to them?’ That’s the Google way—we don’t confine ourselves to catching other people.”
Years later, Gary Flake, the chief science officer at Overture and leader of Yahoo’s search efforts in the mid-2000s, would amuse audiences with a slide show that documented Overture’s failures to respond to Google’s advances. “How did I lose so badly?” he would ask, calling it a classic case of the innovator’s dilemma, where the pioneer in a field—in this case, search advertising—found itself locked into the model that had initially brought it success. Google innovated circles around Overture, focusing on its core obsessions of speed and scale. Overture required its advertisers to pick specific keywords; Google would match an ad to many keywords, some of them with subtle connections discovered by analysis of the behavior of its millions of users. Overture concentrated on high-value accounts that it sold by hand. Google built a self-service system that allowed it to accommodate hundreds of thousands of advertisers. Overture did implement some of Google’s innovations, such as the second-price auction. But by then AdWords had left Overture and Yahoo in the dust.
(Bill Gross would later shrug off the fact that his ideas involving pay per click and ad auctions had made billionaires at Google but not at Idealab. “I feel we won,” he says. “There was the satisfaction of breaking the code. We originally invested $200,000 in GoTo, and when we sold Overture, we made $200 million. That was a pretty great return. And we learned our lessons about patent protection.”)
AdWords Select rolled out in February 2002. The AOL deal went into effect in May. Suddenly, Google’s financial crisis was over. Now Google had a cash cow that would fund the next decade’s worth of projects, from brilliant to lunatic. In 2007, writing about the “spectacular commercial success” of the second-price auction model, economists at Stanford, Harvard, and the University of California at Berkeley described it as “the dominant transaction mechanism in a large and rapidly growing industry.”
Before AdWords Select and the AOL deal, Eric Schmidt often passed by Sheryl Sandberg’s cubicle and asked her how many advertisers Google had. “Not many,” she would say. Later in the day, he’d ask her the same question. “Eric,” she’d say, “not many more than we had three hours ago.” In 2002, all that changed. AdWords Select was drawing new advertisers to the Internet, and the AOL partnership was pulling in ones that had resisted Google. “So we just started growing,” says Sandberg. “It went unbelievably well. And nobody knew just how well until the IPO.”
3
“When the money keeps rolling in, you don’t ask how.”
AdWords would soon get a sibling that would be a mighty companion in piling up revenues. It would expand Google’s advertising power beyond search pages, establishing the company as a provider of ads to all sorts of online properties—and giving it a foothold to making all the world a platform for Google ads.
Typically, it began as an engineering obsession. Georges Harik, one of Google’s first ten employees, had impressed Larry Page during his initial interview in 1999 when he described his longtime goal of using artificial intelligence to analyze data, winnowing down digital content to themes that human beings would recognize. If you did that, he told Page, you might use the information to target ads to web pages. Or, who knew, maybe something else. “It was one of ten ideas I brought up which were equally nonobvious,” he says. And Larry Page said, “Why don’t you work here?” (Harik actually had him at “artificial intelligence.”) Harik, who had a doctorate in machine learning from the University of Michigan, was also impressed with Page. “He was the smartest guy I talked to in Silicon Valley, and I told my parents I just interviewed with a six-person company that in five or six years was going to be th
e biggest on the Internet,” he says. So Harik left his job at Silicon Graphics and joined Google.
Harik began by helping Urs Hölzle create Google’s infrastructure, but he kept thinking about data analysis and artificial intelligence. One day, while talking to Ben Gomes in the kitchen in the Googleplex at 2400 Bayshore Avenue, he described his concept of how compressing data was equivalent in many ways to understanding it. That concept, he argued, could be a key to algorithmically squeezing meaning from web pages. Gomes told him that another Googler, Noam Shazeer, had similar ideas. (While studying at Duke, Shazeer had worked on a computerized crossword puzzle solver.) From that point, Harik and Shazeer, two of Google’s best engineers, stopped working on projects related to Google’s overstressed operations and began an artificial intelligence project that would have seemed more appropriate in a research lab.
“A large number of people thought it was a really bad thing to spend our talents on,” says Harik. But one of Google’s star engineers, Sanjay Ghemawat, thought the project was really cool. So Harik would posit the following argument to doubters: Sanjay thought it was a good idea, and just about no one in the world was as smart as Sanjay. So why should I accept your view that it’s a bad idea?
For the next year and a half, Harik and Shazeer studied probabilistic models of things such as why people often use clusters of words in the same phrases. “For instance,” he says, “when people write the word ‘gray,’ what words are they willing to write afterwards, like ‘elephant’?” The secret to compressing web pages into themes, they discovered, turned out to be prediction: if you can predict what will happen next, you can compress the page. The payoff is that as you get better at predicting a page, you get better at understanding it. Since Harik and Shazeer had the benefit of many terabytes of data documenting the web and the way Google’s users interacted with it, they made good progress and developed ideas about identifying what clusters of words went together. Then, using machine learning, they trained the system to find more clusters and develop rules. “Google had about ten or fifteen thousand servers then, so we had about two thousand to play with,” says Harik. They were using about 15 percent of Google’s computers on their project.
They named the project Phil, because it sounded friendly. (For those who required an acronym, they had one handy: Probabilistic Hierarchical Inferential Learner.) That was bad news for a Google engineer named Phil who kept getting emails about the system. He begged Harik to change the name, but Phil it was.
In February 2003, spurred by the success of AdWords, Susan Wojcicki wondered whether it might make sense to apply the same auction-based, pay-per-click model to a system that involved publishers other than Google. “The advertisers kept demanding more clicks, more clicks, more clicks!” she says. “The idea of putting ads on nonsearch pages had been floating around here for a long time. If we did this, we could go to AOL and offer to put ads not just on the search pages but the content pages, too.” AOL would be only a start. The potential exposure of Google ads on the web would go from the 5 percent or so they currently served to virtually all of the web. And it certainly was no coincidence that Google had bought the world’s most popular blogging service—named, appropriately, Blogger—that same month. Though Google explained the purchase with only a cryptic statement about “many synergies and future opportunities between our two companies,” this new project showed that buying Blogger could reap material benefit, and quickly. It put millions of blog pages, as yet bare of advertising, under Google’s control—a perfect outlet to satisfy the demands of advertisers for more inventory.
And Google already had the crucial technology to anchor a system that matched ad keywords to web pages: Phil.
Sergey Brin thought this was a terrific idea and became Wojcicki’s most powerful benefactor in pushing the program. It took only a week for Harik and Shazeer to make Phil into a system that would match keywords to web pages. (If a page was full of information about winter sports, for instance, Phil would extract keywords like “skis,” “ice skates,” and “hockey pucks.”) Jeff Dean pitched in to merge Phil with the AdWords technology, while another team tried to build all of this into a complete self-service system for advertisers.
As it turned out, Harik and Shazeer were not the only Google engineers working on a project that analyzed content and extracted keywords that could be used for ads. Paul Buchheit, one of the first twenty-five hires at Google, was creating a web-based email system, and he had an idea for analyzing the text of emails so Google could run ads alongside them. By early 2003, he already had a pilot project working that served ads alongside email. Buchheit’s technology wasn’t used in the Google publisher project, but “it was a great proof of concept,” says Wojcicki. (Buchheit’s name appears on the patent, along with Harik’s and Jeff Dean’s.)
Brin wanted to launch a pilot version quickly and have the full program running by May. Google didn’t even have a payment system in place to distribute the commissions to publishers. The only thing close to such an in-house scheme was the method used in a search backwater called Google Answers, an ill-fated experiment that let users bypass algorithms for tough queries and instead solicit answers from anonymous fellow users, who would be paid small sums for satisfactory responses. The new project used that payment system.
In March 2003, Google announced the pilot product, saddled with the awkward moniker Google content-targeted advertising. The blog post trumpeting the program didn’t garner much attention in the general press, but some sharp industry observers grasped the implications. Danny Sullivan, the editor of the website Search Engine Watch, noted that Google—with more than 2 billion web pages in its index by then—had advantages in the field of contextual advertising that no one else had. “The potential exists for the entire web to be Google’s ad canvas,” he wrote. “Everything could become Google’s indirect content.” This was a sentiment that Google itself explicitly endorsed. “We could change the economics of the web,” Susan Wojcicki said not long after the program launched. “You do the content and leave the selling of the ads to Google.”
The idea of analyzing web pages and selling ads that matched with their information was not original to Google. One person who’d had that idea was Bill Gross of GoTo. His brainstorm had come in 1999. “Our product was called LinkAds,” he says. “We did content analysis and then placed our ads on someone’s site. The revenue took off like a rocket ship. But our CEO said it was too complicated for advertisers, and we canceled the product. It kills me that I didn’t fight harder for that.”
In 2003, a Santa Monica–based start-up called Applied Semantics posed a threat to Google in contextual advertising. Founded by Caltech graduates Adam Weissman and Gil Elbaz, it had patented technology that, according to its own description, “understands, organizes, and extracts knowledge from websites and information repositories in a way that mimics human thought.” It used its system in a product called AdSense (with the intentional pun on the “cents” that people would be paid for links) that analyzed web pages and tried to extract the key themes in order to place relevant ads on the page. It sounded awfully similar to what Google wanted to do—and its patent could have been a problem.
Google was in luck, though—Applied Semantics’ exclusive contract with Overture was due to expire that year. Also, Elbaz was friendly with Brin. When Brin asked Elbaz what was happening with the contract, he said that Overture was dangling “a more strategic” arrangement, meaning an actual stake in the company. Brin asked Elbaz to bring his team to Mountain View to discuss whether Google should work with Applied Semantics.
The two teams met in a Mountain View conference room. Jason Liebman, an Applied Semantics executive, showed slides touting the company’s business. Liebman concluded by claiming that AdSense was “a billion-dollar opportunity.” It was basically the same presentation he had given to Overture the day before. In that meeting, Liebman’s billion-dollar parting shot had been met with derision. “Maybe a hundred million,” snorted an Overture
exec. Brin’s reaction was different. “We actually think it’s a two-billion-dollar opportunity,” he said. If you were negotiating a contract, maybe that wasn’t the cagiest approach. But Brin, who had been obsessed with selling ads on other sites for months at that point, had something “more strategic” in mind. He shooed the Google engineers out of the room, and the Applied Semantics people were left with Brin and Google’s business development team. Soon, Google had agreed to buy Applied Semantics. It was its biggest acquisition to that point. Google paid $42 million cash and 1 percent of its stock.
Google changed the name of its content-targeted advertising program to the catchier AdSense. But the product technology was still Google’s, based on the Phil system. (Years later the confusion caused by adopting the name used by the acquired company would lead some to inaccurately accuse Google, and Wojcicki in particular, of claiming credit for Applied Semantics’ achievement.)
Google identified its first AdSense customers as large publishers such as web portals and big newspapers and quickly did what it could to get those accounts. “Google felt that they had a window to be the only game in town,” says Liebman, who came over from Applied Semantics. Just to show that the system could help advertisers and publishers, Google assumed all the costs while it proved its point. Normally, the process began when a publisher signed up for the program and assigned space on a page for relevant AdWords ads Google would find. Then, when visitors to the page clicked on the ads, Google would split the revenues with the publisher. But Google was so eager to get things rolling that it didn’t wait for publishers to sign up and speculatively assign part of their pages to Google. It bought the ad space itself, paying the publishers retail rates. “We’d call them up and say, ‘We’d like to buy some media,’ and then we would run our ads,” says Wojcicki. Google wouldn’t charge the AdWords advertisers for the clicks either, so they were more than happy to have their ads show up on those nonsearch pages. Essentially, Google was paying the costs from both sides, just to launch AdSense.