The Attribution Model: Unmasking Bad Behaviors

We’ve previously written about the problems with the existing “last-touch” attribution model in the digital advertising industry in this column, and how this model breeds bad behaviors by encouraging people to game the system. Two of the ways this can occur are cookie bombing and conversion scooping. To understand how these practices undermine effective attribution, let’s imagine a walk down New York City’s Fifth Avenue.

Here we are, strolling along and browsing the store windows. In this scenario, the cookie bomber is the advertiser who plasters flyers on every visible surface – lampposts, windows, mailboxes, walls, etc. – indiscriminately distributing their message. By targeting practically everyone, this advertiser ensures that anyone who enters the advertised store will likely have been exposed to their ad (whether they paid attention to it or not), thus allowing them to potentially take credit for “converting” these shoppers as the last impression seen before entering the store.

Imagine once again walking down Fifth Avenue. You see a store you’ve read about and have been thinking about visiting and decide to go inside. Suddenly, as you reach to open the door, someone waving a flyer with the store name on it pops up out of nowhere and then runs away. This is the conversion scooper – a sneaky bloke who appears immediately before a user converts and takes credit simply for being the last point of contact. You might not even have noticed the flyer in his hand and it certainly didn’t influence your decision to enter the store. This is how conversion scooping works – the ad quite likely had no value to the consumer, but if it is the last touch, it will get credit for the conversion.

Savvy marketers recognize these behaviors are a problem and are looking for ways to measure the real impact of their programs instead of relying on the last-touch model. The most common measurement technique is the A/B lift test – rotating in a control ad, and then comparing the effective conversion rates between the control and exposed groups. But the problem with A/B tests is that each test can only take into account one variable at a time. Good marketing is multi-dimensional and conducting enough A/B tests to effectively cover all the major dimensions of a given campaign is likely to be too cumbersome and time consuming.

How can a marketer run all those A/B tests without creating an unsustainably complex testing phase? The answer is multi-touch attribution.

Multi-touch attribution can measure the impact of multiple touch points and group them across various dimensions such as channel, inventory source, or creative. This allows for analysis based on interactions between different campaign elements instead of just assigning full credit to the last chronological touch. These touch points represent real-world events such as a click, impression, site visit, email, search, etc. The influence of each touch point can then be calculated, based on the probability of conversion when that touch point is present.

Because multi-touch attribution doesn’t simply look at single touch points, this unmasks the techniques such as cookie bombing and conversion scooping. Cookie bombers get penalized because of the high volume of wasted impressions they generate. Conversion scoopers are in trouble as well. Since MTA models take into account all the touch points leading up to a conversion, they can identify a “scooping” impression that is too highly correlated with other – more legitimate – touch points. In other words, if ad “x” can produce a conversion on its own, it should receive the credit for the conversion. But if the impression only produces a conversion when ad “y” is also present, and never on its own, that ad can be factored out – no more last minute scooping!

Beyond automated bidding, granular multi-touch attribution can allow for specific valuation of dimensions to account for cookie bombing and conversion scooping. These valuations can be applied to the campaign in the form of modified CPA payouts (lower for low-influence variables), sequencing exclusions (preventing interactions between variables that show negative interaction rates), and preferential pricing to promote positive variable interactions. For example, using multi-touch attribution-based insights, an advertiser may lower the payout for affiliates or CPA buys that appear to be cookie bombing or scooping, use exclusion pixels to prevent brand search term clickers from being served additional display advertising, and bid at higher rates for email campaign recipients found on display exchanges.

A solid multi-touch attribution model gives marketers an effective means to measure lift without endless A/B testing. With attribution data available across dimensions, optimizing campaigns becomes easier. By integrating multi-touch attribution with a bidding engine, these results can be used to calculate the value of each impression before it is served, allowing advertisers to bid in a forward-looking manner; advertisers can now effectively ask each impression, “what will serving you be worth to me?” rather than the rearward-looking question, “what have impressions like you given me in the past?”

Let’s take that walk down Fifth Avenue again. This time, the walls and lampposts are clear of clutter. This time, the dastardly conversion scooper is nowhere to be seen. This time, sensibly-targeted impressions – delivered at regular intervals and through appropriate channels – will lead you to a store that is ready to meet your needs. Much better.

Related reading