Authors: Douglas Edwards
Sometimes UI team members disagreed, which made it easier to stay awake. Engineering, research, and marketing were all represented, but alliances shifted. I often found myself in concert with Bay, who had written his doctoral dissertation on human-computer interaction, and with Karen, who intuited what might be useful to "normal" people (that is, users who didn't have an advanced technical degree—or, God forbid, even a BA—from Stanford).
I also agreed with Marissa more often than not. She carefully constructed positions based on user test results or data culled from our logs, the automatically generated internal records of users' interactions with our servers. "When Urs put me in charge of UI," she reminded us, "he said we didn't need opinions. We need facts and research to base good UI decisions on."
The problem was that sometimes we had questions the data couldn't answer. Should color bars be used as section headers or only as page titles? Should we use round radio buttons or a pull-down menu? Should we force browsers to display sans-serif text or allow users to override our choices with their own preferred fonts? Divisive issues that could suddenly turn vicious.
"People want to adjust the number of results they see based on the particular search they're conducting. It's query-specific," a debate might begin. And then, "Any idiot can see we must give them that choice on each results page."
"You kludge-riddled broken brainframe!" would come the reply. "Users want to set their number of results and leave it that way forever. The option to change shouldn't be shoved in their face like a smelly sock every time they search."
"You unbuffered buffoon! You understand nothing about online behavior! Have you even read
User-Centered System Design?
"
"No. But I've read your mama's user-centered system design..."
Okay, so that never happened. Instead, we looked at what other sites did and hypothesized what would be most effective by drawing on our own experience. Truth could not be objectively and rationally derived until we had actually made a decision and implemented it. When we guessed wrong, usage dropped and user email increased.
"Why can't I change my display font?" people demanded to know. "Isn't this still America?"
Arthur C. Clarke once postulated, "Any sufficiently advanced technology is indistinguishable from magic." Our job on the UI team was to set the stage for the technological marvel that happened whenever someone conducted a search—to make Google's interface supernaturally simple to use. The choices we made may look painfully obvious, but that's a mark of the team's success. We didn't try to distract users with anything showy or ostentatious or follow the portal path of flashing colors, dazzling displays, and other hocus pocus. Instead, we presented a well-thought-out, unsullied space with nothing up its sleeve but an incomparable talent for pulling rabbits out of the black box of the Internet. Assuming, that is, you were looking for rabbits.
The UI team grew quickly. Everyone wanted to touch up Google's public face. We spun off a front-end production group to deal with implementation issues so we could stay focused on overall design philosophy. Schwim set up a UI lab for the front-end team with a dozen different computers and browser configurations and even a WebTV unit so they could avoid a repeat of the AOL-Verdana misstep. Even though we added more engineers, designers, and UI specialists, I remained the only resident writer.
Though I would never get under the hood and retool the code base, my aesthetic suggestions occasionally made it onto the site, and my text decrees increasingly carried the weight of authority. I started to feel a role defining itself around me. I couldn't be a tech leader or a design director. But I could be Google's word guy.
L
ARRY AND SERGEY
were not fond of the mainstay of Internet advertising—rectangular graphics called banners that flashed and screamed for attention in annoying ways—and they wouldn't put them on their Google. In fact, they weren't fond of advertising at all, because they felt taking money for ads provided an incentive to bias search results in the advertisers' favor. Still, they had listed ad sales in their original VC presentation as a possible source of income, and they were willing to consider it—if that could be done in a way that was targeted and useful instead of obnoxious.
Cost containment could only carry us so far. At some point Google would need to start generating revenue to survive. We were selling search services to other companies, which put us in competition with well-established players like Inktomi and AltaVista, and the upside potential didn't look fantastic. While Google was gaining a reputation as a search destination, we had no real standing as a provider of services to large enterprises.
That brought Larry and Sergey back to advertising. The logical course would be to outsource development to someone like DoubleClick, whose core competency was placing ads on websites. We were a search company, after all, not an ad network. Craig Silverstein made that case to Larry and Sergey.
"Why spend all the effort? Why get distracted by trying to make an ad platform?" he asked them.
"None of the existing ad platforms work with the kind of advertising we want to do," they replied. "So we'll have to write our own."
"I thought that was a good argument," Craig told me. "But still it made me sad."
The ads Larry and Sergey wanted to serve were all text, not like the intrusive banners that DoubleClick had helped make ubiquitous across the web. No one in early 2000 thought words alone could be effective compared with the glitz of animated GIFs (image files formatted for display online).
"The DoubleClick people we talked to didn't understand," Urs remembers. "They said, 'We can't serve text ads. You'd have to render the text as a bitmap image and we could serve that.'"
And there was another obstacle.
"They had nothing that allowed you to target to a search," Urs added, "and it was obvious that would be important.
*
So the question solved itself." Larry and Sergey took comfort from the fact that their old Stanford friends at the weather website
Wunderground.com
had written their own ad system and found it relatively easy to do.
"Here we were," Susan Wojcicki recalls, "maybe fifty to sixty people, and we were competing already with these huge companies that had much bigger market share in search than we did. And at the same time we were, 'Oh yes, of course, we should build our own advertising system, too.' We wanted to serve ads in every country and have the ads be targeted to every query. And we wanted it to be fast." The company's vision stretched beyond the feasible.
"This was around the time," Susan added, "that Larry decided we should also scan every book in the world." Stretched beyond the feasible by quite a bit.
Jeff Dean began working on systems for managing and serving ad campaigns. He was joined a week later by a recently hired engineer named Howard Gobioff—an iconoclastic thinker and rabid privacy advocate, Howard was a ponytailed, mullet-coiffed Carnegie Mellon PhD and weight-room habitué, equally comfortable sporting a tux, black motorcycle leathers, or a combination of the two. Within days, Jeff and Howard had a working prototype. Then they lent a hand to Marissa Mayer, who had been looking at ways to match ads to searches, and quickly knocked out an ad-targeting system.
To test their prototype, Jeff enrolled Google in Amazon's affiliate program. Every time someone clicked on one of our book ads and then bought that book on Amazon, Google would be paid a commission. Jeff dumped a hundred thousand titles into the new system, which began spewing ads across Google's results pages whenever someone searched for a novel or nonfiction work by name. Even though not every ad earned clicks, the system clearly worked. Suddenly Google was generating revenue, albeit modest revenue, from its ad system. No great huzzah went up. No champagne was uncorked. No expectations were raised. There was still work to be done.
The Amazon affiliate ads proved the program worked. Now we would try selling ads to companies that didn't have affiliate programs. The ads, appearing as text in boxes at the top of our results pages, would be sold on a cost-per-thousand-impressions (CPM) basis. In other words, advertisers would be charged every time Google displayed their ads (delivered impressions), whether anyone clicked on the ads or not, and the rate charged was based on a thousand showings of the ad. To make the program a real product, however, Google needed to be able to tell advertisers how much inventory was available for them to buy. For example, if an advertiser wanted to show an ad every time someone searched for "titanium golf clubs," we had to figure out how often "titanium golf clubs" showed up in our query stream.
Ed Karrels, who joined Google from SGI, built an inventory-estimation tool to harvest that data from our logs system so we could guarantee the number of ads we would show. Even more impressive to his fellow Googlers, Ed established a thirty-five-mile Friday morning round trip to the nearest Krispy Kreme, where he harvested four dozen donuts, loaded them into his silver Mercedes SLK 230, and drove them at dangerously high speeds back to the office.
*
Google now had fried dough Fridays and an actual advertising business.
In January 2000, the Google ad program slid into the open market without so much as a press release to announce it. Cindy thought we should make some waves, but she was overruled, perhaps because Larry and Sergey didn't want to alert the sharks circling startups like ours that fresh meat was swimming in their territorial waters. The percentage of Google results pages that actually displayed our text ads was minuscule, and for years people swore they had never seen an ad on our site. That worried loyal Google users because they didn't understand how we could possibly stay afloat with no revenue stream. Some even offered to send us donations. We politely declined.
Of course, we tested the ads. Marissa, Bay Chang, and Jen McGrath conducted a usability session at a Stanford computer lab.
"I dunno. It seems kind of unethical," one of the participants suggested. "I thought it was a search result. With banner ads, you know that they're paid for."
Well, that was unexpected. Text ads were less ethical than obnoxious blinking banners?
"Shouldn't we test banners and text ads to see which users prefer?" I asked when I heard the feedback. "What's our long-term view on banners anyway? How much are we leaving on the table by not running them?" My assumption was that because image ads were more intrusive, we would be able to charge a higher rate for them. I was on the wrong side of history, but fortunately Larry and Sergey overruled me. They argued that we could always escalate from text to image ads if need be, but it would be much harder to roll back from image ads to just plain text. Text it would remain—at least it would if the ads were successful, about which I had my doubts. We added color backgrounds to distinguish the ads more clearly from search results, but they still looked to me like a throwback to the pre-Windows days of DOS.
Jeff and Howard never expected their hacked-together system to grow into an economic engine. It was just a prototype to prove we could actually serve ads ourselves. Sales gave engineers information about what advertising they'd sold, and the engineers input each ad by hand. The system wrote directly to the database and there was no way to back it up. The gross inefficiency of this manual system bothered Larry. Once it was up, however, no one wanted to take it down.
"The targeting stuff ended up not working so well," Jeff Dean admitted recently, "but we used the core ad-serving systems for many, many years. The nugget of that original design is still in there."
Engineers Ed Karrels, Radhika Malpani, Matt Cutts, Howard Gobioff, and Chad Lester took over as keepers of the advertising flame. They would develop the features demanded by clients and keep the limping system from collapsing under its own weight. Jane Manning joined the group as a project manager and became the port through which the sales team channeled customer requirements. All of the sales team, that is, except for Bart.
Bart Woytowicz, a six-foot-six retired semi-pro international basketball player and bon vivant, headed sales operations. He liked to talk directly to engineers, and the engineers liked talking to him since he was the end user of the code they were writing. Besides, Bart always could be counted on to lighten the mood, as he once did at Halloween by wearing a leather S&M mask and a baseball uniform and claiming to be San Francisco Giants player "Barry Bondage." Bart didn't feel obligated to beat up the engineers about deadlines and deliverables. That was the project manager's job. He did, however, assure them of something they already suspected: the flat-rate CPM pricing we offered advertisers was a dying business model.
The market was shifting to "cost per click" (CPC) pricing. With CPC, advertisers paid for ads only when users actually clicked on them. It was the way our competitor GoTo worked. In GoTo's model, the price paid for each click was determined by a real-time auction, and those advertisers willing to pay the highest CPC won the most prominent positions on the page.
The GoTo model was innovative but flawed. It encouraged advertisers to bid high, but not to target their advertising only to relevant keywords. An attorney representing asbestos victims in a class-action lawsuit, for example, for whom each new client might be worth tens of thousands of dollars, could bid high for dozens of unrelated keywords to blanket results pages with his asbestos ads. ("Mesothelioma" was for a time the most expensive keyword you could buy on Google.) Ninety-nine percent of the ads would be ignored, but the lawyer wouldn't care, because most of that exposure cost nothing. Those off-target, high-bid ads, however, bumped other, more relevant ads to less prominent positions where they'd be clicked on less often. Since GoTo only got paid when the ads were clicked, that would be money out of their pocket. GoTo instituted a manual review process to ensure that advertisers only bought keywords relevant to what they were selling, but they were fighting a losing battle.