If I were to summarize. Incentivizing your visitors to log into their account with you is the only clean permission based way to track them across devices.

This quote from Avinash Kaushik nicely summarizes what I am going to discuss in more detail in the post. So if you’re short on time, please just remember Avinash’s statement.

Cross-Device tracking means being able to track a user’s digital marketing contacts on different devices. The aim is to merge them into one user journey. Imagine you run an e-commerce shop and advertise on Facebook and Google AdWords. One of your customers clicks on your ad in the Facebook mobile app while commuting to work. Later on that same customer researches for your product on Google on her laptop. She clicks one of your AdWords ads and eventually orders the product on the laptop. The “cross-device” challenge is merging the Facebook ad click on one device with the Google AdWords click and sale on the other. The aim is to account for the advertising impact the mobile Facebook campaign had on this particular customer journey and purchase decision.

If the customer would have clicked the ad campaigns on the same device, ordinary cookie-tracking would allow you to associate the two ad-interactions with the orders. But setting one cookie across different devices isn’t possible, since cookies are per definition tied to one specific device and even browser.

There are basically two different approaches to track users across devices. One is deterministic, the other probabilistic.

For Google, Facebook and Amazon it’s easy to “determine” which users use their apps on different devices. Users are usually logged in while using these apps (Gmail, Facebook). Facebook for example can figure out which sites you visited on device A (for example looking for flights to Cape Town on your desktop computer) and target you with related ads (cheap flights to Cape Town) within the Facebook mobile app on device B (e.g. your smartphone).

This of course is not only possible for the internet’s behemoths but potentially for every advertiser whose ads incentivizes users to login to the respective advertiser’s app.

This deterministic method is what Avinash Kaushik describes as the “only clean permission based way to track […] across devices” in his statement above.

Providers: Google, Facebook, advertisers with login

Advantages: most accurate solution for cross-device tracking regarding false positives (very low rate of matching devices not belonging together), no black-box in terms of matching algorithm.

Disadvantages: Walled-Garden if Facebook or Google are used, platform specific cross-device information can only be used for targeting within these platforms, which doesn’t make sense from an attribution modelling perspective.

Through tracking IP addresses, location data, browser type, operating systems and ad requests, amongst others, it’s possible to develop statistical models which allow with some accuracy an inference to link different devices to a single user. For developing these probabilistic models, deterministic cross-device data is needed in the first place as training data. Hence many of the ad tech companies offering these probabilistic approaches use a sort of mixture between deterministic and probabilistic cross-device tracking approaches.

Before advertisers decide to work with probabilistic cross-data vendors, they should make sure, that they fully understand what accuracy means in this context and what these vendors are capable of delivering. As described in an interesting adexchanger article, “accuracy is calculated as the number of matches correctly identified, as well as the number of non-matches correctly identified. In other words, it’s the number of times a probabilistic prediction was correct, but also includes ‘non-match’ predictions from the total pool of predictions it made.” The match rate measures how many times two or more devices could be correctly connected, whereas the accuracy rate also includes correctly identified non-matches. In one study with Nielse, Drawbridge’s (a probabilistic cross-data vendor) accuracy was 97.3%, which sounds awesome. But taking into account, that the match rate was “only” 10.3% puts things into perspective: Out of 100 actual matches between devices advertiser working with Drawbridge will only be able to identify 10.

The problem of course isn’t these lower than expected matching rates, but the way vendors might manipulate their clients with a selective communication of metrics.

The way the statistical models for cross-device matching work is that there is a tradeoff between precision and match rate (sensitivity). For example, if 6 out of 8 probabilistically calculated matches are correct matches, i.e. 2 matches were incorrect, precision would be 75%. Sensitivity or match rate would be, how many of the actual joined devices are being identified correctly. To achieve a high precision, one could chose to only match devices if the calculated probability for a match is 99.99%. This would lead to a low sensitivity and a lot of missed matches. On the other hand, if you want to increase match rate and detect as many matches between devices as possible, you could just call a “match” for any possible device combination, which would lead to a disastrous precision.

As you can probably tell by now, there’s a lot of room for vendors to be creative in communicating their success metrics.

Providers: Drawbridge, Tapad, Crosswise, Adelphic, Adbrain, Roq.Ad

Advantages: Broadens reach of deterministic cross-device data, no walled-garden since cross-device data can be used to target across ad platforms

Disadvantages: Either low matching-rate or high miss-matching-rate, problematic accuracy, privacy concerns since matching algorithms are a black-box.

Probabilistic cross-device tracking for ad targeting and for marketing attribution modelling is problematic. Ad targeting algorithms and user segmentation systems are mostly based on statistical models, i.e. they only work with probabilities on how likely it is that a user is in a specific customer segment or how likely a user is interested in a specific product category. Most of the times the specific ad tech companies won’t publish the exact algorithms nor the targeting accuracy, leaving the client with another black-box. Thus working with probabilistic cross-device tracking data and probabilistic ad targeting systems combines two non-deterministic black-boxes. This will make it really difficult to understand what part of the systems worked or and which didn’t.

Especially for Marketing Attribution probabilistic cross-device data is problematic, since most sophisticated attribution modelling algorithms also rely on statistical models. The problem again would be to combine two probabilistic systems in a case where highly accurate sales attribution numbers are expected. On top of that, integrating with a probabilistic cross-device id vendor isn’t for free and would result in additional costs.

From our experience, the best digital marketing experts use deterministic cross-device data to measure an overall, macro-level cross-device effect (e.g. 10% of all desktop conversions have had a mobile ad interaction before). They know, that deterministic cross-device data can only underestimate the real cross-device effect and adjust their marketing attribution reports accordingly.

At Adtriba our approach is that what can be tracked (deterministically) should be tracked. Therefore we track deterministic cross-device data by offering our clients user-id-tracking for example at log in, sign up, check out, newsletter sign up or email-click. User journeys across device are then merged together and processed by our machine learning based attribution modelling algorithms. Other than that, we would encourage our clients and all other advertisers to increase incentives for their users to login. Being able to deterministically detect cross-device journeys with a sufficient match rate is a huge competitive advantage, especially when combined with a marketing attribution report. Mobile campaigns that previously seemed unprofitable, suddenly show their real influence on the customer journeys.