“Data-Driven Thinking” is written by members of the media community and contains fresh ideas on the digital revolution in media.
Today’s column is written by Seph Zdarko, director of data, modeling, partnerships and attribution initiatives at Quantcast.
The Oxford English Dictionary and Google tell us that the word “fraud” has Latin origin. Cheating people through criminal deception has a long history, going back millennia. But in the advertising industry, ad fraud – deliberately attempting to serve an ad that has no potential to be viewed by a human – is somewhat newer.
Ad fraud is real, representing about 10% to 20% of the digital ecosystem as a whole, although exactly how bad it is depends on where and how you buy. Suffice it to say that it is a serious problem for programmatic advertising, and the industry has started to take action.
Another type of programmatic fraud is particularly pernicious, but hasn’t received nearly enough attention. It’s called attribution fraud, and it refers to the deliberate practice of attempting to serve ads outside of an attribution model’s measurable capabilities.
Attribution fraud can affect 50% or more of marketers’ ad spend, depending on how and where they buy, and you probably have no idea that it’s happening. Like other kinds of ad fraud, it has been around for quite some time but it didn’t explode until real-time exchanges, big data and programmatic started to take off.
To understand why, you need to look at how attribution has evolved. Back in the early days of online advertising, buying targeted ads was a simple process. You bought sites directly and maybe made buys on a network or two. Simple, bottom-of-the-funnel measurement was all that was needed to determine basic ad effectiveness. Last-touch/click attribution worked well, and the use of more complex algorithmic attribution modeling was mostly for multichannel allocation. However, that was back when marketers bought ads 1,000 at a time across known inventory sources.
Today, precision targeting instruments can slice and dice individuals and impressions down to the millisecond. These new tools can dance circles around the outdated bottom-funnel attribution models. Unfortunately, it is often easier for sophisticated demand-side platforms (bidders) to game the bottom-funnel attribution models rather than work within them. Fraudsters can easily elude even the most complex regression and algorithmic modeling offered today.
Attribution fraud hasn’t become a hot topic like ad fraud is today simply because we can’t easily see it – yet. Attribution gaming is happening, but we’re not sure where and how much.
An open-source initiative called “split-funnel attribution” allows providers to augment their existing attribution models and share specific metrics about where fraud and manipulation are happening. Used by Ensighten, Abakus and other vendors, it adds an additional point of measurement into an existing attribution model at the point where the consumer first visits the advertiser’s site. That first site visit is also the handoff between the ad delivery platform’s third-party data and the advertiser’s first-party data.
Adding this delineation point allows attribution models to run separately on both the upper funnel (the prospecting phase, prior to the first site visit) and the lower funnel (retargeting phase, between the first site visit and the eventual conversion). It results in useful metrics such as split-funnel mix, which shows how much marketers are spending on retargeting vs. prospecting.
Splitting the funnel in this way enhances understanding of the causal relationships between the upper and lower funnel activities. It lets marketers figure out exactly how much each contributes to the eventual conversion event. It shows them the individual incremental gains from prospecting and retargeting separately and, at the same time, exposes the majority of attribution fraud.