At the highest level, marketing attribution is about estimating the relative effectiveness of different advertising channels. While the term is used quite broadly to cover lots of different techniques and tools, the goal is the same: determine which marketing dollars are driving business value and which aren’t.
“Half the money I spend on advertising is wasted; the trouble is I don’t know which half.” John Wanamaker
In particular, when we talk about marketing attribution, we’re doing it because we want to estimate the Return on Ad Spend (ROAS) for our different advertising and marketing channels. ROAS is a common acronym and is really a business measure about understanding the return-on-investment of dollars put into advertising. The business wants to make sure that the dollars being invested in marketing channels are 1) driving positive ROI and 2) couldn’t be better spent in other parts of the business.
While the concept is straightforward enough, practically it’s a pretty tough question to answer. And before the advent of the internet and e-commerce, it was even more challenging: imagine you are a large national brand (e.g., Coca-Cola) running ads in the 1990s — you are probably spending a lot of money on television ads, radio ads, newspaper and magazine ads, and smattering of other out-of-home (e.g., billboard) advertisements. Since there was no way to tell which customers saw which ads (or how much each ad impacted their purchasing decision), it was very difficult to tell the marginal impact of ad spend in these different channels. If you’re the chief marketing officer, how do you know which of these channels is actually driving sales? How will you know if you should move spend from television to billboards?
With the advent of the internet (and the digital advertising tracking pixel), we thought we were close to solving this problem. It seems so easy! Now we can see exactly who clicks on an ad and what they eventually purchased (sale: attributed!). But it turns out that we still have the same problems — while we can track the internet ads that customers clicked on, it is still very difficult to know how much that advertisement impacted the purchase decision (if they’ve seen multiple different ads) and we still do not have the ability to easily measure the impact of offline advertising.
When you sell products that require a longer consideration period (like, buying a house) it’s even more difficult to measure attribution accurately since people see many more advertisements on their path to conversion.
The main problem with standard last-click attribution for digital marketing channels are:
- Bottom-of-funnel channels (e.g., branded search) get over-credited compared to top-of-funnel channels (e.g., social media engagements)
- Offline channels (e.g., radio or TV ads) don’t have a digital trail and so can’t be easily tracked
- There’s no simple way to understand the interactions between channels — maybe a customer needs to see both a TV and a facebook ad to convert.
What are the different types of attribution models companies use?
To solve this problem, people have developed many different types of models and approaches (that are not mutually exclusive!) in order to try to understand the attribution problem. These different types of models fall into three different buckets:
- Order-level attribution (attributing a specific order to a specific channel or set of channels)
- Incrementality via experimentation (making a change to spend and measuring what happens)
- Media mix modeling (using statistical approaches to unpack the relationship between time series of spend and revenue)
I’ll quickly cover the basics of these different types of models so that you’re oriented
This is the most common type of attribution models for digital companies that are just getting started, and we briefly discussed this flavor of attribution in the opening section above. The notion is that you can use the digital trail from customers visiting your site to tell which ads they saw or clicked on before they visited your website and you can use that trail to assign attribution. The most common form of this is known as “last click” attribution where you assign credit to whichever channel a customer last clicked on (e.g., if a click on a facebook ad led immediately to a purchase, facebook would get full credit for that conversion even if they saw and / or clicked on other ads in the past).
There are lots of different flavors of this that use different weightings than 100% last-click. You could use a first-click algorithm, a 50/50 between first- and last-click, or any other assignment of weights across the different digital touchpoints you might have.
In concert with these methods, many companies also implement the use of coupon-codes to track customers obtained through offline channels (think about all of the podcast ads you hear) or the use of post-checkout surveys to understand what other channels might have driven the conversion.
Algorithmic multi-touch attribution is a generalization of the digital attribution models that use statistical models to determine the optimal weighting algorithm. So rather than choosing attribution weights for the different touches upfront, you use a statistical model to estimate the appropriate weights based on their predictive ability.
“Cross-device” attribution and tracking
Cross-device tracking is an add-on feature that improves the effectiveness of the two types of digital attribution models listed above. Because users may use different devices at different times during their path to conversion it can be difficult (or impossible) to know that customer’s entire conversion path if you do not know that the same user who was browsing on their mobile device later logged in to their desktop computer to actually make the purchase.
Cross-device models and vendors use other information to provide a unified view of a customer’s journey so that the digital attribution methods are even more effective.
Incrementality via experimentation
Rather than trying to assign each user a “channel” (or a number of different channels with different weights), incrementality via experimentation attempts to directly measure the ROI of advertising spend in a channel. The easiest way to implement this type of experiment is to run a geographic hold-out test. So if you run a national radio campaign, you might try to hold out all of the state in Colorado in order to measure the difference in revenue between the non-Colorado states (that had the radio advertisements) with the performance in Colorado (which didn’t). If you structure the test correctly, this can give you good evidence about the effectiveness of that advertising spend even when you cannot track which users or customers may have been exposed to the ad.
There are lots of ways of to get sophisticated in exactly how you run and analyze the results of these experiments, but the idea is generally the same. One issue that many people face when running these experiments is that the operational overhead of running a geography test can be quite high, and in some channels it may actually be impossible (e.g., for Podcasts).
Media mix modeling
“Media mix modeling” is an old technique that broadly refers to the practice of using a statistical model to estimate the relationship between spend in an advertising channel and overall performance. The idea is similar to estimating incrementality via experimentation, but instead of intentionally “holding out” spend in certain markets, the model attempts to use natural variation in the spend and response over time in order to estimate the marginal impact of spending extra dollars in a channel.
This type of model does not associate a specific user with a specific channel, but rather attempts to estimate the marginal effectiveness of the channel at a high level without requiring the business to manage lots of experiments.