Debunking cross-device myth

Recently there has been a lot of noise around Device Graph (cross-device), fingerprinting, cookie resurrection not to mention persistent zombie cookies which caused a lot of confusion on the topic. This inspired me to address the fundamental question on how cross-device technology works by laying out the major principles and defining the foundation.

While specific methodologies vary from provider to provider, Device Graphs typically establish links between devices in two primary ways: deterministic matching and probabilistic matching.

Deterministic Matching

A deterministic (or direct) match is established when there is known information that connects different devices together, such as user login information. A consumer, for example, may login to a website to access certain features or content, and may also login to the same publisher’s mobile app to access content on the go. This login data, which is encrypted to strip it of any PII, can then be passed through a Device Graph to establish a direct link between different devices. While deterministic matching is the most accurate way to link different devices together, the methodology has one notable flaw: scale. The number of publishers with login data, or other data points that can trigger a direct match, is limited. Major portals and social networks have a clear advantage here. For publishers looking to compete with these large players – or for marketers looking to extend their reach beyond these players – an alternative solution is needed, which is where probabilistic matching comes in.

Probabilistic Matching

Where known data linking devices are not available, probabilistic matching can be used as a powerful alternative. Probabilistic matching takes PII-free signals that flow from different devices and uses statistical analysis to identify links between the signals. Over time, the algorithm analyzes enough signals, and sees enough patterns between the signals, to tie different devices together with a high degree of confidence. Probabilistic algorithms require significant scale to make them effective and truly actionable.

It is clear, that emerging Device Graph technologies are being developed and refined to address the challenges marketers face as a result of media fragmentation across devices. At their core, Device Graph technologies link different devices together without the need for a universal ID or any personally identifiable information.

As part of the probabilistic match, which relies on advanced statistical modeling to link devices, Device Graphs typically utilize the following **four distinct signal buckets **to ensure the pairing is as accurate as possible:

• Consumption signals – IP Address, Location, Day Parts, Timestamp
• Visitation signals – Frequency of visits, Time between events
• Device signals – Device make, model, operating system, screen size
• Behavioral Signals – Affinities, Interests, Demographics and Intent

These signals are not mutually exclusive and must be analyzed in aggregate to make device matches. Successful probabilistic algorithms must be dynamic in nature, constantly ingesting and analyzing new data points to refine and strengthen matches between different devices.

Any probabilistic solution cannot by nature be 100% accurate, but applying a number of ongoing tests can help improve the confidence level in matches that are being made. At Lotame, we call this a confidence score. We look at a broad range of different factors to establish this confidence score, and only when the score exceeds a minimum benchmark will a match be published. Factors that go into creating this confidence score include:

• IP Matching Between Devices
• Number of Devices Seen on an IP
• Total Volume of Signals from Individual Devices
• Time Gap between Signals
• Number of IPs Seen on a Device

Looked at in aggregate, and over time, these factors help build up confidence and inform the probabilistic algorithm, which is constantly learning and improving.

Probabilistic match in practice

This example illustrates the confidence scoring process Lotame uses to determine matches between devices. In this example, four devices can be detected on a single household IP, but can they be linked to the same user with a high degree of confidence?

To conclude, we live in the era of programmatic and high automation. Enormous computing power is backed up by boundless cloud infrastructure, so I challenge you to be armed with clearly defined questions when you assess cross-device vendors. There are no excuses to fall back on standard cliché responses like “we use advanced statistical analyses to power probabilistic matching with certain degree of accuracy” or “it’s part of propriety technology”, don’t settle on “black boxes” shelf responses, demand profound explanation on how technology really works. Question everything – scale, accuracy, match rates and technology, you deserve to know how tech you are investing in operates.

Lastly, I will leave you with the focal questions, I expected any respected cross-device vendor to elaborate on in detail as part of due diligence process:

1. What methods do you use to link different devices together?
2. How do you handle false positives when linking two mobile devices?
3. What kind of reach can I expect to see with Cross Device offering? Size does matter: mandate cookie reach and device IDs scale… as well as match rate between the two which is GEO specific.
4. How do you handle data protection and opt-out in light of Cross-Device?
5. How your Cross-Device technology different from “device fingerprinting”?
6. Does your technology work without implementing an SDK?
7. Are you compliant with use-based rules for data mining purposes?
8. How do you measure accuracy for probabilistic matching?

… and let the best vendor win!