I should slow down, as that's not entirely accurate. The search engines themselves weren't predicting anything about the elections (though it would be fun to see them try). What some of my colleagues and I tried examining, however, was whether consumer search behavior was a better predictor of election results than polls. It wasn't.
It's worthwhile to explore a situation where consumer search behavior has its shortcomings as a predictive measure. Generally, search behavior is great for indicating and even predicting trends, but the "wisdom of crowds" phenomenon doesn't always work, and it's helpful to explain why.
This research was spearheaded by my colleague Andrew Chang, who will unseat any incumbent for Researcher of the Decade. Now, let's review why search behavior had to concede to the polls in predicting the midterm elections.
1) All politics are local, but searches are national. I searched for Virginia's candidates for the Senate, George Allen and Jim Webb, after they appeared on "Meet the Press." I'm from New York, though, so my search behavior would be irrelevant to election pollsters. I was merely curious to learn more about the candidates, one branded a racist, the other a sexist (the beauty of the democratic system is that we sometimes have a choice between the two). This study (more of a cursory exercise than a deep dive) looked at general search volume trends on a national level; delving into local search data for each race could have provided better comparisons between search behavior and polls. Still, despite national media attention, seven of the top cities where searches for Allen and Webb originated were in Virginia, according to Google Trends.
2) Sex sells searches. If a politician is embroiled in a sex scandal, searches for that politician take off. It doesn't even have to be that salacious, though--or sometimes, even remotely sexual; any scandal will do. Tennessee's Harold Ford Jr. was smeared by his opponent with claims of interracial dating. George Allen gained notoriety for uttering a racist remark on the campaign trail. In a local race, New York's incumbent comptroller Alan Hevesi had his wife chauffeured around on the state's dime. The volume of searches skewed heavily toward Ford, Allen, and Hevesi, but only Hevesi won his election.
3) Search data works better for trends. The volume of searches did predict most of the contested elections we studied. In a separate experiment, we set up a course where rats would run to the finish line, and 60 times out of 100, the rats wearing blue jerseys beat the rats wearing red jerseys, roughly matching how well the search results predicted races. Yes, the rat race is a joke, but predicting races isn't all that impressive a feat; most voters could have probably done a decent job. What we were trying to gauge was how search behavior compared to polls. Predicting the actual results is much more challenging
4) There's a difference between research and entertainment. In races where there was one candidate who provided good media fodder or was just a more interesting, memorable personality, he or she tended to receive far more national attention and also more searches. In the California gubernatorial race, Arnold Schwarzenegger beat Phil Angelides by 17 points, but searches for Ah-nold were over 90 points higher than they were for Phil.
5) Even rocket scientists can't predict every race. BuzzLogic's blog reviewed an article in Congressional Quarterly where Garth Sundem, author of Geek Logic: 50 Foolproof Equations for Everday Life, wrote a formula to predict House and Senate races. Sundem's equation included Google hits when searching for candidates' names, Google hits when searching the Republican's name with "Bush," Google hits when searching for the candidates' names with "apologize," money raised, money accepted from scandalized lobbyist Jack Abramoff, the candidates' tenure, the weather, President Bush's approval rating, and whether the locality voted for Bush or Kerry in 2004.
CQ reported prior to the election, "Applying the formula: GOP Sen. Conrad Burns will lose in Montana because of his high Abramoff co-efficient; GOP Sen. George Allen of Virginia will fall in a squeaker, mainly because of his Bush associations; Democratic Sen. Robert Menendez's dollar advantage should carry the day in New Jersey; but Democratic Sen. Ben Nelson (who is actually a clear favorite) doesn't have a chance in Nebraska, because the state just votes too red." Sundem nailed Burns, Allen, and Menendez but botched Nebraska. And again, to the point of the research Andrew Chang was doing, the CQ report didn't assess how well Sundem compared to the polls.
Sundem's mixed record and the inconclusiveness of Andrew's experiment both speak to the unpredictability of elections. Consumer search behavior was never designed to be an indicator of election results. To make better use of monitoring search behavior, marketers can tap it to gauge trends for product sales, purchase intent, media consumption, and countless other measures.
There are also key differences between searches for commerce or entertainment and searches for politics. With politics, there are generally only two options, and most of the time searchers have already made up their mind about them before even conducting a search. They probably even know months in advance whether they'll vote in an upcoming election.
Most consumer decisions aren't so simple, as consumers generally have more options, more uncertainty, and no set time constraints where they have to act on one day only. For that reason, consumer search behavior's research value should gain every marketer's vote of confidence.