Earlier this week I had the opportunity to post on the m6d blog about some exciting technology we’re developing to ensure the quality exchange inventory we buy. I re-posted it below, but you can (and should) read the original here.
Ad exchanges are like a garden bed. Just as soil doesn’t know the difference between a flower and a weed, ad exchanges do not always distinguish between good and bad inventory. It takes a vigilant gardener to weed out any undesirable elements. Lately, Media6Degrees has been doing a lot of gardening.
We are, and intend to remain, the most aggressive company in the exchange ecosystem in rooting out the bad stuff so that we buy only the best impressions for our marketers. We’ve been told repeatedly by our supply partners that nobody takes inventory quality more seriously than we do.
Our approach manifests itself in four areas:
- Brand Safety
- Ad Collision (also known as Unintended Roadblocks) Prevention
- Fraud Protection
First, we are one of the original subscribers to third-party brand safety services that allow us to turn off sites that do not meet the brand standards of clients like Disney and Allstate. When we see a site that does not pass muster for these marketers, we block it. This is an important step, but fundamentally a black-list approach; we turn off stuff known to be bad. We recently took the much more aggressive step of turning off all sites until we are reasonably certain they are good. Historically, we assumed all sites were innocent until proven guilty. Now we assume guilt until we establish innocence. While we don’t claim to be perfect, we are confident that we work with the best, most brand safe publishers in the vast exchange ecosystem.
Second, while “Roadblocks” or “Page Takeovers” are sometimes desirable in premium buys, these types of ad collisions (two or more ads from the same marketer on the same web page) are typically not something marketers want with their programmatic buys. Earlier this year, we introduced a feature that ensures we serve only one impression per page per campaign. This was a very challenging engineering problem because a publisher can put multiple placements from the same page up for auction at the same time, often through multiple exchange partners. Our solution requires geographically dispersed servers to be aware of the actions of all other servers, so that a no-bid decision is rendered when a neighboring server claims inventory on the same page. This bid-locking happens billions of times each day in near real-time. Our Ad Collision Prevention (popularly known here as the “Roadblock Buster”) is a major technological feat, and something for which we have patents pending.
Third, we recently noticed a distinct rise in the click-through rates for several of our campaigns. It quickly became clear that these CTRs were artificially inflated — greater than 1% in some cases. We simultaneously uncovered a very high overlap of user visits between certain sites. We were finding some pairs of sites, seemingly unrelated, with co-visitation rates of 80% or greater. For example, an English language site about women’s health shared 80% viewership with a site about Chinese movies. In comparison, the maximum overlap for large and related sites like NYTimes.com and CNN is only 20%. We surmised that the high CTR sites with huge overlaps in viewership were very likely fraudulent, and suspended buying impressions on these sites for our campaigns.
Identifying and turning off sites with very high click-through rates and very high rates of user overlap is easy. The more difficult challenge was how to handle the users. We expect that in most cases, a user stumbles upon a bad site and that their behavior is suspicious for a short period of time, after which they behave normally. We did not want to block these users indefinitely, but we wanted to avoid them while they were “infected.” So we came up with the concept of the Penalty Box. When a user visits a site that we consider to be suspicious, they enter the Penalty Box and we suspend bidding on them for a period of time. Once they haven’t exhibited suspicious behavior for an acceptable period, we treat them like a normal user.
Our Fraud Protection work is cutting edge and we have filed patents both for our approach to detecting suspicious sites and our techniques for blocking suspicious users.
The next challenge we see on the horizon is viewability. This one is complicated, as we have observed that viewability is sometimes inversely related to site quality. As an example, most of the link farm sites (which we do our best to avoid) place all of their ads above the fold. They rate high for viewability, but otherwise are undesirable for high quality display campaigns. Our approach is to be cautious in implementing a viewability-based quality score into our bidding system. We have a major initiative on this front, and will be rolling out solutions over the coming months.
We are extremely proud of the work we do to make sure our inventory is clean and meets the standards and performance goals of our clients. We remain diligent in weeding out the bad stuff and delivering a fertile terrain for major marketers to find new customers.