Attribution modeling isn’t dead, but it isn’t perfect, either
“What is more challenging? Rock climbing or creating an attribution model?”
“Definitely the attribution model!”
When we sat down with Buddy Marshburn, rock climbing enthusiast and data engineering manager at Loom, we had to ask him that million-dollar question.
His answer didn’t surprise us. Attribution modeling is notoriously hard. And LinkedIn posts claiming “attribution is dead” don’t really help.
The takeaway? Stop chasing the elusive “perfect” attribution model, prioritize what works for your business, and focus on insights.
Last-touch isn’t better than multi-touch attribution
There’s a ton of debate on which model is better: last-touch vs. multi-touch.
But there’s no agreed-upon way to build an attribution model. One model isn’t superior to the other. In fact, companies shouldn’t be thinking about multi-touch or single-touch at all. You need to aggregate all these touchpoints to understand your customer and make better business decisions.
As Boris Jabes, CEO at Census, shares: “There’s so much power in getting every touch point aggregated and exposing that to various teams.”
Boris believes that most companies benefit from having product, marketing, and sales activity in one stream or timeline to make better business decisions.
When you know all the touches that have occurred, marketing and sales can comprehend the entire customer journey, use the right messaging, and learn which channels actually work. This data helps them make improvements to their strategy and increase the conversion rate throughout the funnel.
Attribution modeling is complex, with no single source of truth
Anyone who’s tried building an attribution model will attest to its challenges and complexities.
Buddy shares: “I think of all the data modeling work that an analyst has to do, building an attribution model is by far the most complex.”
You have to answer hard questions to avoid wasting your budget on a channel that doesn’t give you a positive ROAS.
For instance, if your customer clicks on both Google and Facebook ads, both these platforms will take credit for that conversion. You have to know which channel you give credit to and why.
To make matters worse, Jarry Ahmad, an analyst at Uber, adds: “There’s no such thing as a clean data set, ever. I think 80-90% of the initial building of the marketing attribution model is cleaning up these data in a sense that works for attribution. Because these data sets are for the purpose of wide use and not specific use.”
But it doesn’t stop there.
Spend data is complex to work with. When you spend millions of dollars on ads, you want to actively use spend data to optimize campaigns. But what happens when the data you’re dealing with isn’t accurate?
Imagine you’re running a campaign, and you want to find out how much you’re spending in a city and where you’re getting conversions from.
“Not all of your spend is allocated on a city level; it’s allocated on a national level. Do you develop a heuristic to allocate that money on a city level, or do you stick to a national level?” asks Jarry.
The answer, of course, is that it depends. What are you trying to achieve with your attribution model? What level of accuracy are you okay with when optimizing your campaigns?
Answering these questions is a collaborative sport.
For instance, most people might default to Google Analytics to define a session, but Buddy believes that’s not necessarily accurate.
“We do this hand-wavey thing where we assume a session is 30 minutes. Until someone asks: why do we have to define it that way?” he asks.
To remove this ambiguity, involve stakeholders — sales, marketing, GTM, and data teams — so everyone gets a say, and you reach a decision together.
This process is how Loom sticks to concrete truths — engagement and activity — and stays away from fuzzy, gray areas.
Campaigns can cost you thousands with no real return
There are many reasons to care about attribution, but Jarry shared a simple way to look at it:
“What are we spending, and what are we getting for it?”
Take SEM, for instance. Cannibalization with non-branded and branded SEM is a real problem.
Say someone is looking for a ride-sharing app online. They are bound to see an ad for Uber and may click on it. But someone specifically searching for Uber on the search engine is going to click on the ad for Uber.
In the second instance, you’re spending thousands of dollars on people who already know your brand. And you might wonder if SEM contributes to your sales at all.
What would happen if you didn’t run any ads?
You can answer that question by performing an incrementality test.
In 2012, eBay’s test showed that search ads only slightly increased the likelihood that someone with no prior knowledge of eBay would make a purchase. But for customers who had purchased from eBay at least three times in the past year, search ads didn’t result in any increase in sales.
While your results may differ, incrementality testing helps you measure an incremental lift in the ROI of your ad spend, ensuring you can evaluate the real impact of your campaigns.
The future is marketing-mix modeling
Jarry predicts that given the current shortcoming of attribution modeling, “the landscape is going to shift towards marketing-mix or media-mix modeling, which is a more holistic, probabilistic way to look back at what happened.”
He adds, “attribution modeling is based off of clicks. But the bigger the company gets, the more you’re spending money on not-click-based marketing.”
Take Uber, for example. The company is spending a lot of money on long-term strategies like brand marketing via TV ads, billboards, and direct mail. You can’t log clicks for those channels.
Media-mix modeling holistically looks at the entire marketing picture and budget and uses data science techniques like multi-linear regression to evaluate how marketing activities, pricing, seasonality, and variable factors impact sales and ROI.