Gannett Is Using Deep Learning to Determine Why Certain Ad Designs Work

The newspaper giant wants to beef up its local marketing services

Gannett is doubling down on local marketing with AI. Getty Images

Gannett is turning to a form of artificial intelligence to design better online ads.

The USA Today publisher recently rolled out a new internal platform that uses deep learning and computer vision to determine which images, colors and other design aspects work best in online ads across its dozens of local news sites. The company says the tool has boosted click-through rate by between 10 percent and 33 percent in four out of five A/B tests it conducted (the difference remained flat in the fifth test).

Advertisers that tested the tool included a local realtor, an auto dealership and a large regional grocery chain, all based in the Indianapolis area. None were available for comment.

The software, called Grandstand, examines the creative of each ad and predicts how well it will perform on a scale from “very poor” to “very well” across four categories: color, text area, call to action and format. The forecast is also specific to the type of product or service advertised, although Jeff Burkett, Gannett’s VP of display advertising, said some insights tended to hold true across the board. The color white and large blocks of text were found to be generally unappealing, for instance. The company also used the program to generate a set of principles for its designers to reference.

The rollout is the latest step in Gannett’s ongoing effort to build what is essentially a marketing agency arm to serve brands that advertise on its more than 100 local properties, including The Des Moines Register, the Detroit Free Press and The Indianapolis Star.

To that end, the publisher has also acquired three ad-tech firms over the past two years—ReachLocal, SweetIQ and WordStream—and integrated them into a marketing platform called LocalIQ, which was unveiled in September. LocalIQ, which Grandstand is now a part of, is meant to give local businesses tools to do everything from planning marketing strategies to designing websites, even encompassing media properties beyond Gannett’s own.

Gannett is hoping that the move into marketing can help it squeeze more out of digital ad revenue, which has grown modestly but not enough to offset plummeting print revenue. In its most recent earnings quarter, the publisher reported a slight 3.2 percent uptick in digital dollars as compared to the same period of the previous year, while print fell a precipitous 16.7 percent. Operating revenue fell to $711.7 million from $744.3 million in the third quarter of 2017.

Its biggest threats on this front are the duopoly powers of Facebook and Google, which have drawn local businesses with a combination of scale and granular targeting.

“It’s one thing to have the scale that we have, but I think the whole idea here is that you still have to be better competitively,” Burkett said. “We think that data—particularly the areas of deep learning, computer vision, using those for optimization across all channels—is really the future of how we’re going to be better than everyone else.”

Burkett said that while the tool hasn’t been introduced to the majority of advertisers yet, the few businesses it did approach about the trial runs were receptive to the idea of allowing Gannett to rework their creative material.

“We’re using this data to go back to the advertisers and say, ‘Digital’s different. Here’s why. Here’s how you properly transition from print to digital,'” Burkett said. “That’s something we’ve always tried to do, and now we have the data to prove it.”

Gannett didn’t initially have AI in mind when it set out to build a better performance measurement system to inform design last year. The company first tried to sort the flood of data it was receiving by having human designers classify each ad according to a set of design principles. But Burkett said the team quickly realized the method was too subjective and couldn’t cover anywhere near the scale needed.

That’s why starting earlier this year, Gannett tapped its data science team to find a computer vision model that could identify objects and text within ads and a deep learning model to produce insights into what worked best. Burkett said he has a lot of ideas of where to take the tool going forward, including potentially incorporating demographic targeting and adjusting for the type of content on which ads appear.


@patrickkulp patrick.kulp@adweek.com Patrick Kulp is an emerging tech reporter at Adweek.
{"taxonomy":"","sortby":"","label":"","shouldShow":""}