What’s wrong with multi-touch attribution?
Posted Oct 24, 2018
Multi-touch attribution is having its moment in mobile.
With the traditional models such as first and last touch seemingly outdated, the industry is looking towards multi-touch to give it a more holistic overview of ad interactions.
The crux of multi-touch attribution is this: rather than reward the last click that drives a user interaction, credit is given across the entire user journey, whether it be ad impressions, clicks and conversions. The goal is to identify ad campaigns that create an uplift of conversions for other ads, and also to give credit to everyone who was involved in an install.
Sound complicated? That’s because it is. No one’s yet solved the multi-touch attribution puzzle - and in this article, we look at all its current limitations, as well as what the real solution should look like.
Before jumping in - and if you haven’t already read it - our introduction to multi-touch defines what exactly it is and explains the most common models used.
What does multi-touch attribution look like on mobile?
In the web world, multi-touch attribution is an established practice of rewarding different ad campaigns and balancing out attribution methods that would otherwise over-value classic "last-click" ads, like Google Adwords.
Think of a user that, after seeing banners and videos for a new product, finally looks it up on Google, and ends up clicking on an Adwords ad. Without multi-touch, Google would get 100% of the credit for that sale, when the video should also get partial credit for introducing the product to the user.
In this example, the uplift is pretty clear, but what about in the mobile ecosystem? What does this look like for mobile UA?
Apple Search Ads certainly function like Google's Search Ads, and so ensuring that they’re not over-credited seems a logical option. It’s also worth investigating if a user that has seen a few interstitials, or even used an instant app ad, is more likely to click on another banner that finally leads to their conversion.
Theoretically, we would expect to see a higher click-through-rate (CTR) for users that have previously seen another ad. The conversion rate from landing on the App Store page, to clicking the download button, is more likely to be driven by the app page content and the quality of the user.
Sounds good so far - so what's the problem?
Problem one: Installs are overrated
The first problem with multi-touch is how to define the conversion event made possible (supposedly) by a series of ads.
Currently, the market's multi-touch solutions focus on app installs as the conversion goal.
However, the user's intent to install should be weighed a lot less than, say, booking a flight or buying some shoes.
After all, assuming that a "low intent" conversion - like downloading a free app - is the same as spending money in-app is quite a stretch.
But that’s far from the biggest problem.
Problem two: No visibility into ad delivery
As mentioned above, we expect the uplift effect of ad "A" to be represented in a cohort of users that have seen ad "A," and are now more likely to click ad "B". However, the issue is that attribution companies only see a tiny fraction of impressions.
This is because several networks don’t share impressions or click data with third-party attribution companies. In fact, only video networks actively send impressions, while the rest almost exclusively send click data.
This means that - unlike on the web where the entire ad delivery process is visible from impression to conversion - mobile has an incomplete, skewed sample of data. And that doesn’t make things reliable.
So, instead of using a series of impressions, some use multi-touch attribution to determine if a user that has clicked ad "A" and ad "B" is more likely to download the app.
By nature, they "measure" uplift as CR increase and not as CTR increase, where you would expect to see the bigger impact.
Problem three: Assisted installs don’t work
Multi-touch attribution was designed to determine if an ad improved conversions further down the line. However, unsurprisingly, that takes a lot of complex mathematics. So, instead, the "assisted install" was born: basically, all touch points until a user's final click is credited with an assisted install.
Assisted installs vastly increase attribution numbers, and reduce effective CPI - so naturally, networks love this metric. Even better, one install can create quite a few assisted installs.
But it means that often, metrics for assisted installs are nothing but bloated numbers over-valuing networks that managed to send the most clicks - regardless of any actual uplift.
Which brings us to the elephant in the room…
Problem four: Fraud (the big one)
As we’ve seen before, Adjust is (sadly) the only attribution provider that actively filters out click spam. And clicks that either show up with high rates of repetition, or have no correlation between click and install, are the most prevalent types of fraud we see today.
However, the combination of multi-touch attribution and click spam is like pouring gasoline onto a fire.
Without using mathematical models that determine if an engagement actually increased the chances of a user converting, fraudulent and fake clicks can create assisted installs like nobody's business. Imagine a user with 10 (fake) clicks in the past few days and one (legit) click five minutes ago. Instead of giving full credit to the only legit click, the fraudulent activity earns a huge chunk of assisted installs.
Fun fact: our distribution modeling distinguishes engagements that, simply, didn't affect the user - versus entirely fake clicks or impressions.
Fraud metrics for click-to-install times, on the other hand, do not account for the assisted installs that these same clicks produced; it's nearly impossible to express that as a simple metric that allows users to understand what's actually going on.
Long story short: as long as an attribution company does not model actual uplift, fraudsters are the real winners of multi-touch.
Problem five: Fractional callbacks
Last, but not least, even if we were to correctly determine the fraction of an attribution - as in the attribution models explained in our previous post - our partners could not digest that information. Networks do not yet have the capability to accept callbacks with credit for "less than one conversion".
It’s not up to attribution providers to call networks and tell them how much of a percentage they were involved in - each would still try to bill for the full amount.
It will take time before these attitudes (and technology) can change.
What’s the solution?
Now, you're probably asking yourself, "how did multi-touch attribution become so popular?" and "is there even a way to do it right?"
We’ve answered the first question already: networks need higher numbers and cannot produce them any other way than claiming every user… twice.
And, while the answer to the second question is a resounding "yes", the bad news is - as usual - that it's complicated.
First off, the industry will have to push networks towards giving us more insight into ad delivery and to share more impressions with us. Then, we’ll have to develop a way to model real uplift to suppress click spam. That’s a long uphill battle, but it’ll be worth it when we get it right.
We’ll be revealing more about this in the next few months, and in the meantime, keep an eye on the blog for more articles about multi-touch attribution.