The Victory Lab (21 page)

Gage and Seaborn traveled to Little Rock to meet with officials at Acxiom, one of only a few data vendors that claimed to have built a file covering the whole American population.
In 1969, an Arkansas school bus manufacturer named
Charles Ward had decided to start the company as a way to help the Democratic National Committee use computers to manage its fund-raising lists. As Ward’s staff got better at gathering personal information, their manila punch cards became valuable to customers outside of politics, such as the American Bible Society, which Ward was pleased to learn was a more reliable client than parties and campaigns. Over the years, the company, which
took the Acxiom name in 1988, acquired
smaller list vendors and cut deals to bring in data from a variety of businesses about their customers: magazine publishers like Rodale, mail-order retailers like Lands’ End, financial institutions like Charles Schwab, pharmaceutical manufacturers like Pfizer, plus hundreds of boutique lists that compiled things as arcane as the types of motors purchased by boat owners.
Between 1983 and 2004, with advances in computing, the amount of data Axciom was able to store increased a millionfold, and the company used every byte to fill out one of its personal portraits of an American with a new brushstroke of data. While civil libertarians chafed at companies buying and selling personal consumer information, some of the most valuable details came from government itself. Bureaucratic applications, for gun licenses and construction permits, could say a lot about how much money someone had and how he or she spent it.

With Acxiom’s individual dossiers, Gage realized, he could finally design clusters that were unmoored from geography. The Republican National Committee had assembled the country’s first national voter file in 1990, and had compiled IDs conducted by local parties and campaigns, information shared by coalition partners, and some so-called enhancements from commercial vendors, such as individual phone numbers. (Because that data would come from a magazine subscription record or completed rebate form, it included many numbers unlisted in the phone book.) No one had ever proposed that the RNC buy a list of pool owners, because it was extremely expensive and no one knew how it could possibly be useful in a political campaign.

If Gage merged Acxiom’s personal dossiers with the RNC voter file, he could use that as the base for polling calls, picking names off the file instead of dialing random digits. Then he wouldn’t have to waste polling time asking people how much money they made or what job they had—Axciom already had categories that knew the answers to those questions, or at least predicted them based on information it did have. The polling questions could stick generally to political matters, taking in respondents’ views of issues and personalities in the news. Gage could come away with
nearly one thousand variables for each of his respondents, most of them Acxiom’s consumer categories. Gage knew that a lot of those variables would never have any bearing on politics, but it wasn’t possible to tell for sure in advance which ones he would need. He would let the computers find out.

In Little Rock, Gage and Seaborn realized that the logic of political targeting was foreign to Acxiom officials. They were used to dealing with corporate clients who knew their intended audience, and so Acxiom could offer them things like a roster of truck owners, or of Kansas City truck owners who also buy religious reading materials. Bewilderingly, though, Acxiom didn’t have a way to order a list of truck owners who did not read the Bible—or any other combination based on subtraction as opposed to just addition. Perhaps more important, the company’s pricing was structured for commercial customers who wanted to string together a few variables on a limited group of people. Gage would need to cover a whole state’s voting-age population, and demanded many more pieces of data on each of them; he needed all the variables so he could isolate the ones that were helpful in predicting political behavior. Gage and Seaborn negotiated a deal to get access to all the variables Acxiom had available at a flat rate, and left Little Rock with hundreds of them for every Michigan adult so they could start hunting for patterns that would pull them together into targetable groups.

Because his groupings would not be defined by geography, Gage preferred to call them “segments” instead of clusters. He viewed their creation as something of a passive process. It wouldn’t be for political consultants to divine that segments should be formed around socioeconomic status or religious views or participation in party primaries. Algorithms could find the variables that were pulling people together in ways that informed their likelihood of backing Bush’s reelection, whatever they may be. The segments would be only as large as they needed to be to ensure that everybody within one of them belonged there equally. And when the campaign wanted to speak to a segment of voters, there would be no doubt on how to
find them. Gage could print out a list for a mail vendor or canvasser, with enough information on every single member of a segment to ensure that he or she was only a phone call, postcard, or door knock away.

IN EARLY 2003
, Gage returned to Dowd, this time with PowerPoint slides that referred to the method as “MicroTargeting.” Gage thought this was an improvement over “super-segmentation” (even if a search of news archives revealed that the new term was used elsewhere to describe a medical technique for removing cancerous tumors). Gage had emerged from the 2002 elections with a mixed record in gubernatorial campaigns, but he thought he had a good story to tell—especially as Dowd began plotting an electoral-college strategy for Bush that emphasized some traditionally Democratic states, like West Virginia, Oregon, and Wisconsin. “We’ve got to find out who is more likely to be a Republican,” says Gage. “We know they’re in there, somewhere.”

The previous fall, Michigan Republicans lost the governor’s mansion but won the attorney general’s office for the first time in four decades, while expanding their footprint in both houses of the state legislature. “The down-ticket performance of the party that year was incredible,” says Gage, who credits it to an advanced ability to rouse Republican voters living in Democratic strongholds. In Massachusetts, Romney had entered the last week of his campaign for governor lagging Democrat Shannon O’Brien by five points, according to his internal polls. He ultimately beat her by that margin, helped by major gains among the independents and conservative Democrats who had been Gage’s targets. Gage had not been given a lot of time to develop a new data-driven strategy for Romney before the September primary. He merely wanted to rank-order
the state’s nearly two million independents based on their openness to Romney’s appeals, so that the campaign could devote its resources to speaking to its friendliest targets first. Gage invented an index he called “consideration,”
a ten-point scale predicting how likely a voter would be to “consider voting for Mitt Romney.” At the same time, Gage modeled those voters’ issue priorities to see which ones should be approached with mail and phone calls emphasizing Romney’s tax plan and which ones should get an education pitch.

As he ran his numbers, one variable popped out. Those who ranked highly on Gage’s consideration index were very likely to be premium cable-TV subscribers. Gage suspected HBO subscriptions were a proxy for other variables—something that neatly packaged the well-to-do and highly educated suburban independents who would warm to Romney’s technocratic approach—but the reason didn’t matter as much as the result. Now instead of trying to pay or recruit the manpower to canvass more than one million potential targets by phone and to judge whether they should receive Romney’s mail, Gage could just send brochures to everyone shown in Acxiom files to be a premium-cable subscriber. Meanwhile, the campaign used the targeting to bring new value to its volunteer operations. Romney, struggling to overcome the almost complete lack of Republican organization in Massachusetts, was able to assign the campaign’s five thousand volunteers to speak to those persuasion targets who lived within their own neighborhoods. “We felt like it was a pretty powerful way to do outreach in these communities, because you had someone who was calling from two blocks away and could talk about the local school,” says Dunn.

Dowd found the results from the Michigan and Massachusetts efforts to be promising enough that he persuaded Bush’s White House advisers to back a trial run of Gage’s approach in Pennsylvania, which had a series of judicial elections in 2003 and would be a key presidential battleground a year later. Gage again attached Acxiom consumer data to the RNC’s voter records, and commissioned a survey of five thousand voters. Then he used algorithms to divide the state’s electorate into more than twenty-five segments based on concentric patterns in voters’ lifestyles and beliefs. Sometimes the Acxiom variables that formed Gage’s segments were apparent: many of the 446,698 “Bible Believers” had shopped at a Christian bookstore,
had told a consumer survey that they had religious material in their home, or otherwise resembled those who did. Regardless of how they got into the group, Gage wrote, “despite their higher than average scores on other conservative indicators, social conservative messaging is a must to maximize the vote in this segment.” Other segments—like Pennsylvania’s 243,517 “Dining Room Debaters,” or the 139,586 members of the “Republican Intelligentsia”—were a little less intuitive or self-explanatory. “That’s where we saw the data dance a little for us,” says Adrian Gray, who became the campaign’s voter contact director.

Still, it wasn’t clear that Gage’s computer models were any more effective than traditional targeting methods. After Pennsylvania’s election, the RNC hired Dave Sackett, a pollster for the Tarrance Group, to sample voters and see how accurately the microtargeting had predicted voter behavior. Sackett’s memo argued that Gage’s microtargeting had actually been less efficient than traditional methods at predicting turnout, but had succeeded at finding friendly Democrats and independents and determining which were likely to be pro-life. “That was the piece that was most radical—with microtargeting we’re talking about things that are not necessarily absolutes. You have to trust the statistical inferences in the data set, and that’s a bigger leap of faith,” says Terry Nelson, the campaign’s political director. Because Republicans were used to mailing issue-specific messages to those whose names had been gathered by coalition partners, there was some certainty about why people were on the lists. This was particularly important on cultural and social themes where there was a constant fear of backlash; Gage’s segments appeared, in some ways, to be too refined for the Manichean moral conflict waged on the glossy surface of direct mail. “With a segment said to be eighty percent likely to be pro-life, was it really eighty percent likely?” Nelson asks. “If it was sixty percent likely to be pro-life we’d probably target it differently—because you might not want to mail them with these messages because it would turn them off.”

Furthermore, the price tag for expanding the technique used in Pennsylvania to the entire battleground nationwide was staggering. Gage
estimated it could total $3 million, which would cover the cost of acquiring consumer data from Acxiom and other vendors, the lengthy large-scale surveys to benchmark the electorate, and the statistical analysis that would bring them together. In 2000, the RNC and Bush campaign had spent little of their election-year budgets on data, relying on the party’s permanent voter file and the generosity of coalition groups.

Even for a campaign many expected to be the best-funded in history, $3 million was a lot of money, the equivalent of two weeks of very heavy advertising across Pennsylvania. Dowd set out to do something he knew was among the hardest tasks in politics—rewriting a campaign budget to include a line item from a category no one had ever seen before. Dowd had already successfully persuaded Bush’s top political advisers, particularly Karl Rove, to invest in a year of voter contact research. Gage’s technique, as he saw it, was a natural extension of the 72-Hour Task Force’s findings. Now, as Dowd and White House political director Ken Mehlman argued to Rove, the only way to reap the benefits of more efficient field contact was to ruthlessly segment the electorate. “Without the ability to find three people out of ten on a block, we wouldn’t have had the resources,” says Dowd. “We would have had to knock on all ten doors.” Microtargeting, he and Melhman explained, could effectively automate the sort necessary at the beginning of any voter persuasion operation: separating those already on board from those who will never be, and then sifting through the remainder to identify the best candidates to receive mail and phone calls making the case for Bush. This process, Dowd hoped, would help pay for itself.

In many ways, Rove should have been an ideal consumer for microtargeting. He was an unabashed data nerd who had run nonrandomized experiments in 1994 to measure the impact of his mail and phones in Bush’s first gubernatorial campaign. Previous political strategists in the Oval Office had been pollsters, media consultants, or campaign managers; he was the first presidential consigliere to have a background working in voter contact. But Rove’s experience in direct mail also made it hard for him to imagine a contact universe built from anything other than the mix of precinct
targeting, coalition rolls, and paid IDs that had proved so effective at illuminating a latent conservative coalition in Democratic Texas. Those lists had been built through manual assembly—a few hundred names from here, a couple thousand from there—whereas Gage conjured his through the alchemy of computer algorithms. “Karl was against this originally, because it was new and different,” says Dowd. “He’s a direct-mail guy and so he thinks he knows that.”