In Online Advertising Auctions, DataXu’s Rocket Science Turns the Tables in Buyers’ Favor

12/15/09Follow @wroush

The introduction of pay-per-click advertising in the early 2000s by Overture, Google, Yahoo, and others made the Web a lot more appealing to advertisers, since it meant they only had to pay for ads when Web surfers actually clicked through to their sites. A more recent innovation—real-time bidding for individual display-ad impressions—also helps advertisers, by allowing them to spend only on ads that will be shown to qualified leads.

But much of the power in online advertising markets still rests with the sellers of ad space rather than the buyers. To guess which display ads will have the highest click-through rates, for example, buyers in real-time auctions are still dependent on just a few sparse pieces of data from sellers, such as the names of the publications where the ads will appear.

Now a startup in Boston, DataXu, is working to change all that. DataXu’s bidding engine, based on optimization software originally developed by aerospace engineers at MIT, analyzes the ad slots available to a buyer at any given moment and predicts which are most likely to induce clicks or conversions, based on dozens of parameters such as the location of the viewer, the day of the week and the time of day, and the content of the ad itself. It does this 100,000 times per second or more, learning as it goes by measuring actual click-through rates for ads the engine has placed.

“The Internet advertising industry has never really created powerful tools for buyers of advertising,” says Michael Baker, DataXu’s CEO and president. “Yield management and analytics—all that stuff was invented by the sellers, the Yahoos and the Googles and the Microsofts. For the first time, we are going to have those kinds of powerful next-generation tools available for the buyers.”

DataXu’s system could mark the start of a major shift in power in the Web advertising world. Imagine, for instance that an auto company is trying to attract prospective customers to a website where they can sign up for test drives. DataXu’s software would examine the performance of all the prior ads the company has run for the site. “We might learn that from Tuesday to Thursday, from 2:00 to 5:00 in the afternoon, in the Northeast and the South Central regions of the country, there is a much higher likelihood of that ad actually causing a consumer to fill out the test drive form,” Baker says. The software would advise the car company to bid only on ad slots that fit those criteria, and avoid slots that don’t.

Not only does such a system allow companies to figure out which half of their ad budget they’re wasting—to cite the old saying about advertising—but it does away with the whole concept of a “media plan,” the carefully planned list of outlets where ad buyers would traditionally look for ad inventory (i.e., available impressions). “We’re saying there’s a new paradigm, which is that you can plan your media buy based on empirical results taken from the Internet on the fly,” says Baker. At least one major customer, the Havas network of advertising agencies, has signed up to use DataXu’s system.

DataXu was founded in 2007 but flew under the radar until its coming-out at the TechCrunch50 conference in San Francisco this September. (The “Xu” in the name comes from the Mandarin word for “need,” according to Baker.) A $6 million round of venture funding from Flybridge Capital Partners and Atlas Venture last spring allowed DataXu to move its 20-some employees out of their original space at the Cambridge Innovation Center to funky offices in Boston’s Leather District.

When I visited the location in October, staffers were still hanging blueprints of Apollo-era Saturn V rockets on the walls—an homage to the company’s beginnings, which really are in rocket science. MIT aeronautics and astronautics professor Edward Crawley developed the original technology behind DataXu’s decision-support system with students in his lab as an entry in a NASA competition designed to find an automated tool that could judge the best way to get astronauts to Mars and back.

“It was basically a design tool that looked through 30 billion variables, searching vast combinatorial spaces very quickly, to find the successful combinations,” says Bruce Journey, DataXu’s chief revenue officer, who originally joined the company when Crawley asked him to find the most lucrative commercial application for the system. “We looked at a number of markets like commercial aviation, logistics, financial trading execution, but there were none that were nearly as compelling as the online advertising market,” Journey says. (Full disclosure: Journey was CEO at MIT’s Technology Review magazine when I was a senior editor there.)

The DataXu system connects with eight different auction-based ad exchanges, including Google/Doubleclick, Yahoo/Rightmedia, Microsoft, AdMeld, AppNexus, and PubMatic. It continuously sucks in data about their available inventory—which changes from millisecond to millisecond, since real-time ad bidding is all about serving individual impressions to Web surfers at the moment they request a Web page—and picks the best deals for each of DataXu’s customers, depending on the goals they’ve set for their advertising campaigns.

Explains Baker: “It’s looking at the stream of ads and saying ‘No, no, no, no, yes to this one at 47 cents per click, no, no, no, no, yes to that one at 17 cents per click, in an ongoing dialogue” with the ad providers’ networks. Except that the system is also learning as it goes, and it’s making the recommendations very fast—about half a million ad impressions per second were available in October, a number that’s expected to rise to 2.5 million per second by January.

With all that inventory to sort through, ad buyers need powerful software on their side to find the best prices and the most effective venues, says Jeffrey Bussgang, a general partner at Flybridge who is on DataXu’s board. “The dirty secret of online advertising is that it is incredibly inefficient to purchase and optimize and impossible to do it in real-time,” Bussgang says. “With its highly automated, machine-learning approach, DataXu solves both problems. There’s no reason every advertiser on the planet wouldn’t want to use the technology to improve the performance of their online advertising.”

For the moment, Havas is DataXu’s only announced customer. Whether more advertisers will come to see DataXu’s services as a worthwhile investment, and whether ad networks will release more of their high-quality inventory for sale through auction platforms, are two of the big questions that will determine what kind of year the startup will have in 2010.

Right now, the major online ad networks still sell the bulk of their display ads through direct contracts negotiated by their sales forces; only the unsold inventory goes into the auction-based exchanges. But Baker thinks rising demand will create more supply. “We think a lot of media buyers and brands will invest aggressively in this way of buying,” he says. “So, put simply, that’s where the money will be. If you’d like to participate in those revenue streams as a seller, you will need to make inventory available.”

Wade Roush is a contributing editor at Xconomy. Follow @wroush

By posting a comment, you agree to our terms and conditions.