Let’s say you run an ad campaign across desktop, mobile, outdoor, and TV, and you notice that sales jump after the campaign. That’s great. But then you want to know which ads in particular drove the consumer to take action.
That question has bedeviled marketers forever. Outlining the difference between direct-response advertisers and general advertisers, ad icon David Ogilvy said in the late 1980s that “the general advertisers and their agencies know almost nothing for sure because they cannot measure the results of their advertising.”
Ogilvy’s comments were made well before today’s omnichannel media environment came to pass. These days, ad campaign options include email, mobile, desktop, TV, addressable TV, social media, and digital outdoor, not to mention traditional analog media vehicles such as print and billboards.
What’s more, the explosion of media choices has emerged simultaneously with machine learning and a newfound appreciation for data. Now many marketers are looking to close the loop and use machine learning to assign “multitouch” attribution (MTA) to various media to provide a better accounting of how their advertising worked.
According to a 2016 eMarketer survey, 57.6% of marketers said that cross-channel measurement and attribution would occupy their time and resources that year. Dennis Buchheim, SVP of data and ad effectiveness for the IAB, said he believes that about 50% of marketers are using some form of MTA.
“The level of understanding varies quite a bit,” added Ian Dahlman, VP of search and analytics for B2B agency gyro. “We’re going to see it more in our day-to-day in ’18 or ’19.”
Since the advent of web-based advertising, the industry has attempted to divine its ROI. One method has been last-touch attribution, which gives 100% credit for a purchase or desired action to the last ad a consumer has seen. This works especially well for online sales. If you see an ad for a pair of Nikes on Facebook and then click to buy them, then you can make a case that the ad drove the sale.
The reality, though, is that some 91% of sales take place offline. The average consumer also zips back and forth between websites and tabs, desktop and mobile, so even connecting an online sale to a single ad exposure is tricky.
Last-touch’s continued use was illustrated by Google’s announcement in May that its Google Attribution product was designed to do away with last-touch attribution. Digiday recently called last-touch’s use an “Internet mystery.”
For marketers, according to Peter Horst, a consultant and former CMO for Hershey, last-touch was of limited use. “It certainly was the easiest, but that’s certainly not the one you rely on at all because it’s so misleading,” he said.
Like other marketers, he relied on marketing-mix modeling (MMM) and testing instead. Testing would typically involve tens of thousands of consumers and would require various permutations of media. “At the moment, none of them are perfect,” he said.
MMM uses data from past performance to create a model for a media plan. The system then offers an estimate of how successful a particular plan would be. Proponents of media-mix modeling say it’s about 90% accurate. Critics, such as Starcom Mediavest CEO Laura Desmond, have said that MMM is flawed because it is usually based on historical data and isn’t agile enough to adapt to real-time data from new campaigns.
Alison Latimer Lohse, co-founder and chief strategy officer for Conversion Logic, said MMM is useful for some things, such as flagging the seasonal effects of pricing and promotion. “I think media-mix modeling can be very effective at identifying trends and impacting broader business levers,” she said. “I think where it diminishes in its accuracy is where you get into anything tactical.”
MTM and MMM answer two different questions, she said. MMM answers the question of where marketers should make macro investments. “It’s sort of like steering a cruise ship,” she said. “It’s the course you set.” MTM, meanwhile, offers a “hyper-granular view” that lets marketers make micro adjustments to reach that set course.
How MTA Works
Following use of last-touch attribution to assign credit for conversions, cookie-level data became available. Then marketers were able to use rules-based MTA, which assigned credit for all of the ads a consumer saw before a conversion. The next step was linear regression, which gave weighted scores for each touch point based on its perceived ability to prompt a conversion.
A “time-decay” model, for instance, gives less credit to ads that a consumer saw several weeks ago than for one she saw in the day or so before a purchase. So if a consumer saw 12 ads before making a purchase--the average is around 10 or 15 over a period of 60 to 90 days before purchase, said Mike Finnerty, senior director of advisory services for Neustar--then each would get partial credit, but the more recently seen ads would get more credit.
Latimer Lohse said machine learning has allowed for a more sophisticated analysis. Conversion Logic, for instance, uses machine learning to score each algorithm it uses. “We don’t know which one’s going to perform well,” she said. “We want to run every client through the data set and determine which one has the best predictive outcome.” This scoring is all executed via machine learning.
Offline media, meanwhile, is usually factored in as part of an MMM scenario. That is, if a percentage of the budget went to traditional outdoor, then marketers assume it had some impact on the sale.
“The hammer-on-nail part of offline is connecting your credit card transaction at a Macy’s store with exposure that you had in digital,” said Todd Parsons, chief product officer at SocialCode. “Where things like MTA come into play is out-of-home. Technically, it’s a little bit fuzzy, but you can model the number of times people in that fenced area are going to go by a billboard and then do some funny math to see who got exposed.”
What makes this kind of attribution difficult is that the media realm is extremely complicated. For instance, a consumer might have seen 12 ads, but the real catalyst for the purchase was a segment on the “Today” show or a conversation with a friend. That’s why even the staunchest MTA supporters admit that it’s at least partially based on human assumptions and will never be 100% accurate. But MTA can help get marketers closer to figuring out their ROI if they supplement it with other tried-and-true techniques.
“As pragmatic and as dull as it may sound, you’ve got to test this stuff with promotional codes and things that are more direct-response oriented or you run around in circles doing defensive math,” Parsons said.
Though MTA has been presented as a means to illuminate marketing spending and end the jockeying between various departments to claim credit for a sale, Parsons said enough ambiguity still exists with MTA to sustain such infighting. “There’s still a lot of custom math,” he said. “The problem with MTA right now is we’re piecing together this incredible mosaic, but a lot of the underlying tiles are assumptions.”
As of now, the marketing world seems to be split by those sold on or resisting MTA. While the benefits of MTA are clear, the cost of such programs is an obstacle, as is the need for highly skilled employees who understand how to use and manipulate the data necessarily to use it effectively.
The other resistance to MTA is structural. Implementing MTA often prompts some significant budget shifts. “If you work on the TV budget, you may not want to know the digital MTA,” Buchheim said. “There’s a fear of the unknown.”