• Choose a category
• All
• Classic-Fiction

It was Paul Bucheit who showed them where to look.

Just for the Hack of It

Paul liked to hack things together. In his free time, he was constructing an email system he called "Caribou"—named for a poorly managed project assigned to the comic strip character Dilbert. We would later call it "Gmail." He had been working on Caribou for a couple of years. Though dozens of Googlers, including Larry and Sergey, were using it, almost everyone he talked to at Google hated the idea of becoming a provider of email to the public. Some opposed it from a technical standpoint, some from a business point of view. But they all agreed on one thing. An email system should never display advertising based on the content of the email messages it sent or received. It was both technically infeasible and unacceptable from a product perspective.

"Before I was even working on Caribou," Paul told me, "Udi Manber* came in for a talk and I remember asking him, 'Do you think you could ever do content-targeted ads on email?' And he said, 'Oh no, we would never do that. That's a gross violation of whatever.' It was like, you should be fired for even asking that question."

Paul knew his mail system would never launch if it couldn't eventually pay for itself. He considered building in premium services or offering six-month free trials, but knew those models would inhibit widespread adoption. And Sanjeev Singh, a colleague helping out with the coding, kept bringing up content-targeted ads.

Paul liked to hack things together.

One night, he and Sanjeev went through his inbox one email at a time and tried to manually match each message to an ad already in the system. It wasn't that hard. Paul decided to put together a simple prototype that would do the matching automatically. He rummaged around his code files and came up with a classifier tool: software that could identify things that were related and group them. He had written it as part of Matt Cutts's porn filter project. Perfect for what he had in mind now. He reconfigured the porn classifier to match ads to the content of emails, flipped it on for all the users of Caribou at three a.m., and went home.

The next day, all the engineers using Caribou saw ads targeted to the content of the messages in their inboxes. They were not pleased. Paul recalls that Marissa, with whom he shared an office, was furious.

"We agreed not to do this," she insisted. "And then you went off and did it anyway."

"I don't remember ever agreeing to that," Paul replied. "Maybe you said not to, but I never agreed to anything. I'm not really that agreeable a person that I would ever agree not to do something."

Larry and Sergey weren't furious. They were impressed. The idea of content targeting had been around for a long time, but only as a theoretical concept. It was even on the engineering project list, but assigned a priority of one—the lowest—on a scale of one to five. Everyone assumed it would be hard to build and not terribly accurate. Yet Paul had built a working prototype in a couple of hours, basing it on a self-described "stupid hack" he had written for another project.

The founders, according to Paul, "were more open-minded about it than most of the rest of the company." They immediately saw the potential for targeting ads, not just to email, but to the content of any web page that contained text. They pulled together a team in September 2002 to start working on a more robust content-targeting system under the code name "Conehead." A month later, Google content-targeted ads started showing up around the posts in the old Deja News Usenet pages that had been incorporated into Google groups. A month after that we paid for low-cost remnant ad space—impressions that had not been sold and would have displayed house ads or PSAs—on small test sites like the ever-agreeable Wunderground.com to see if we could earn more from the ads than it cost us to pay for the media. In March 2003, Google's content-targeted ads publicly launched, expanding the places our ads could potentially appear