Demystifying Causality for Marketers

Causality – what does it even mean in digital advertising? And what does it have to do with attribution?

Granted, causality is a relatively new concept in our space, but it’s one that all marketers should at least know about.

In plain English, causality in digital advertising is measuring the number of conversions that stem from consumers who needed to see a display ad in order purchase a product or service. In other words, it separates the consumers who purchase stuff on their own because they’re both brand-aware and in market for a particular item, from those who need an extra nudge to convert.

Central to causality is this notion that some websites naturally attract consumers who are pre-disposed to buy from particular brand (which means marketers don’t need to spend as much of their budget advertising on those sites), while others attract more brand newbies, so the display ad may actually cause those visitors to convert. If marketers can identify the publishers that deliver the most incremental sales, they could drive campaign performance and get more bang for their marketing buck.

So if the measurement of causality is all about knowing which online publishers deliver the most incremental conversions, isn’t that the same as attribution? “Not exactly,” explains Dan Hill, PhD, and senior data scientist at Integral Ad Science, and the person leading the causality team at Integral. “Attribution takes all of the conversions that result from a campaign and divides up the credit among the publishers that displayed the ads to converters. What it can’t do is tell the marketer how much incremental revenue each of those publishers delivered, or which publishers are most effective in influencing wholly new customers to a brand.”

Sounds intriguing, but is it doable? How does one go about distinguishing between consumers who are natural converters, from those who need to see a display ad in order to move through the awareness, consideration, and purchase cycle?

According to Integral, the A/B tests of the public health studies serve as a useful blueprint for measuring causality (in fact, Dan Hill worked as a neuroscientist doing academic research prior to joining Integral).

Causality measurement begins by establishing a baseline of natural conversions for each publisher. Let’s say that brand A receives two sales from The New York Times each day without any advertising whatsoever. We can say that the baseline for The New York Times is two sales per day. If brand A opts to advertise on The New York Times, the marketer should expect to see additional sales each day above its baseline. If no incremental sales result from the ad spend, then it’s clear that The New York Times isn’t a good fit for the brand. But if dozens of sales start pouring in each day, then the site is truly a winner for that brand.

Of course, selecting sites for a campaign would take forever if an advertiser had to first establish a baseline for each potential website and then measure the incremental lift. Is there a way to do both simultaneously and cost-effectively?

According to Dr. Hill it is indeed possible with A/B tests. In A/B scenarios, for every customer who saw an ad and converted, you need to find a similar person – a twin, so to speak – who didn’t see the ad. If the two people are similar enough, the twin serves as the baseline for what would happen if the original customer did not receive an ad.

Traditionally, when marketers wish to test an ad they show a PSA to the control group and their branded message to the test group, but that’s expensive. If you want a test population of 1 million consumers, that’s a lot of wasted inventory to purchase.

To get around that extra expense and still perform A/B tests at scale, Integral proposes using unviewable ads as the control group. We can assume that, thanks to the campaign criteria, the consumers who are targeted are relatively equal. And we know that more than 50 percent of ads will be unviewable, which means they had absolutely no chance of influencing the consumer. To establish a baseline, one needs to compare the number of conversions among consumers who were targeted with an ad that was unviewable, to those who were presented with an ad that could actually be seen. The difference between the two is the incremental lift, and the causal impact of an ad.

There’s a bit of magic (read: algorithms) required to ensure consumers with viewable ads are really comparable to consumers with unviewable ads, so causality must control for various other factors. However, 90 percent of the bias is removed simply by ensuring you’re always comparing consumers who are served ads.

It’s an interesting concept, and one that we will follow in the months and quarters ahead since digital advertising thrives on accurate measurements.

Related reading