Targeting The Web's Guilty Pleasure: Images

"My name is Steve and I am a slide-a-holic."

"Hello, Steve!"

I admit it. I am a sucker for a good online slide show. On my nightly tablet crawl I hit the USAToday Day in Celebrities swipe-a-thon or’s latest pile of star snaps. For all of the effort we put into targeting and formatting ads for the other mainstream porn of the Web, social media and video, it is remarkable that only recently are we seeing more done with the stacks and stacks of static images that engage people for minutes on end in a site. Take a look at some of the media brands that pop up on image-sharing networks like Pinterest and Instagram. Craft ideas and recipes kill on Pinterest. Some people just love wallowing in food images. And on Instagram you see otherwise niche brands -- like extreme-sports sites and magazines you didn't know still existed like High Times -- get inordinate followings for their image posts. Dude, you gotta see these primo buds! Yeah -- stoner porn. 



And so I was fascinated to speak recently with one of the providers in the burgeoning space of in-image advertising, GumGum, about how one goes about segmenting and targeting images on a site. The in-image ad is a unit that usually pops into the image's bottom edge or in a corner. At its best it look a bit like the bottom-third ads now common for cable TV house ads.

GumGum’s SVP of Marketing Tony Winders explains that understanding the content of an image usually involves multiple kinds of technology. The metadata provided by the site and the article content are usually just a starting point. “We use technology to understand what the photo is about, like pixel-based image recognition," says Winders. "We can detect what a make and model of a car might be. We use image clustering to understand if a picture appears on more than one site, so we can pair their data. We understand what the image is about and what the page is about to target on keywords down to celebrity name or make and model of a car.” Interestingly,. However, Winders says that channel targeting performs better than keyword targeting.

The challenge of serving ads into images is daunting. Sizing is never certain, for instance, nor is position on the page. So GumGum does what it calls “asset sequencing,” an analysis to determine where on the page is the most effective place for the ad.

During the Summer Olympics, for instance, sponsor BP had an ad triggering its TV campaign targeting images related to the Olympics. Winders says that in-image ads, which often lead to video views, are especially well-suited to targeting events. In BP’s case, the online buy was amplifying an on-air buy. Conceivably a sponsor could target images related to the run-up to the Oscars awards ceremony, say, to get some of the effect of an Awards show buy without incurring the same cost as TV.

Indeed, the model is best suited overall for entertainment, sports, auto and other “benign” image content that does not involve news. GumGum has over 900 publishers among its partners, including Tribune, Hearst and Gannett, often working in the photo galleries around sports and entertainment.

The idea, of course, is to overcome banner blindness by putting the unit in the user’s line of sight within the material that is most engaging to them. The method boasts an average click-through rate of .45%, but in cases like BP targeting a hot topic like the Olympics, the CTR on the video lightbox the ad popped up was 1.2%.

Winders says that in-image advertising also can address issues of frequency and viewability. Three-quarters of the ads are coming above the fold, and they are only served when the image is visible in the browser. Since they control all of the image space on the sites they represent, GumGum and the publisher can control the frequency with which ads appear. In a recent post-campaign study, they found that three to four impressions got the optimal lift. But this format may have even greater potential on devices. Winders says that of the 1 billion images into which GumGum is serving ads, 21% of the traffic is coming from mobile. “We are seeing comparable and some higher engagement on mobile compared to the Web,” he says.

I’ll say. Images are not just mainstream porn for daytime office-bound folks. They are the natural prime-time complement to TV viewing. Almost every publisher I know was flabbergasted at the hunger for images on touch-screen devices. Couple that with the mode of use, prime time in front of a TV, and the situation is perfect for image swiping: no-text, low-focus eye candy that multitasks perfectly with a TV screen. Prime time is slide time. 

1 comment about "Targeting The Web's Guilty Pleasure: Images".
Check to receive email when comments are posted.
  1. Sione Palu from Feynmance, December 2, 2012 at 9:38 p.m.

    Image segmentation is compute intensive. Also text extraction from text embedded in an image has high computation costs. Despite this high computation cost, numerical methods that are being applied to image understanding has been advancing in the last few years, such as wavelet, tensor, independent component analysis, etc,... Microsoft was working on something similar a few years ago. Their paper is available for free download, which was published in the computer vision journal. The title is "TextonBoost for Image Understanding: Multi-Class Object Recognition and Segmentation by Jointly Modeling Texture, Layout, and Context" and download is :

Next story loading loading..