The two great challenges in digital analytics right now are tracking customers over time and across channel. Naturally, they are deeply related. Digital channels are very measurable. But the ability to track specific interactions at a very high-level of detail doesn’t always translate into a broader capability to track a sequence of actions at the customer level.
It’s no mystery why digital analytics techniques often struggle to bridge time and channel. The high-level of detail typically available in digital masks the often paltry information we have about who the customer is. A huge amount of digital tracking relies on cookies (small files dropped on a machine) to understand and connect behavior over time. Unfortunately, cookies are notoriously non-persistent. Problems ranging from software cookie blockers to upgrading machines to switching devices (cookies are device specific not visitor specific) to periodic cookie deletion all make tracking events over time and across channel difficult or impossible. Every business is different, but we tend to use 3-4 months as a general rule of thumb for the half-life of a cookie – meaning that there is a 50% chance that the cookie will be lost in that period. Where analytics is reliant on cookies, a great deal of traditional customer analytics is simply impossible. Techniques around churn and lifetime value, for example, simply won’t work. Worse still, these gaps in tracking often doom efforts to measure the impact of digital activities on the broader customer journey.
Of course, digital analytics isn’t solely reliant on cookies. Registration and logged-in sites can provide relatively stable views of the customer over time and even across device. Making this work isn’t as simple as it sounds. Connecting keys from digital to non-digital is critical for effective analytics, but often ignored in basic infrastructure tasks. Digital registration systems don’t always (or even often) tie back to true customer keys in back-office systems. Particularly in B2B, it can be very challenging to navigate the hierarchy of keys from individual users to organizations to actual relationships. Making sure, for example, that your systems aren’t being used by shared logins, that you can understand roles for individual users, and that you have a standardized client organizational hierarchy are all widely neglected aspects of creating a clean key-matching between online systems and actual customer accounts.
Even in this realm of logged-in sites, cookies can present significant challenges. Particularly where there is a mixture of logged-in and public site behavior (quite common on Financial Services sites and Media sites), this can be problematic. A visitor can go from known, to new (after deleting a cookie) and then go back to known either mid-session or after multiple visits incorrectly identified as belonging to a new user. Cleaning this up requires significant batch re-processing of data and fairly sophisticated key-joining.
Where sites don’t have consistent log-ins, it’s especially challenging to study any form of cross-channel behavior. It’s not unusual for interactions (the online interaction and the offline interaction) to be anonymous. A manufacturer, for example, will generally have no idea about the actual identify of a Website visitor or a retail purchaser. Given that, it’s no surprise that connecting the two can’t happen. You need to de-anonymize both channels with a consistent identifier before behavioral cross-channel analytics are possible.
In my post from two weeks back (a really important one and well worth circling back to if you missed it), I laid out a case-study where we used re-survey techniques to answer a set of critically important questions about the connection between online behavior and downstream value for a manufacturing client. In that post, I described how we answered questions like ‘”What’s the value of improving satisfaction on the Website?”, “What’s the value of customers in specific online use-cases?”, and “What are the best measures of online success for any given online user and use-case?”. These are fundamental questions but outside of pure ecommerce plays they can be nearly impossible to answer with traditional behavioral analytics because of the gaps in over-time, cross-device and cross-channel tracking that usually exist.
I thought if might be worth laying out the alternatives for doing cross-channel and over-time analytics and highlighting when each might be the most sensible approach.
Behavioral Approaches
When it’s possible, behavioral analysis is by far the most powerful approach. It provides a deeper, better, more customer-specific analysis path. Not only are the results of behavioral analysis stronger, they also lend themselves to operationalizing via personalization in ways that don’t obtain with other methods. The challenge to behavioral analysis isn’t, generally, the analysis itself. The hard part is creating the connections in the data.
There are many techniques for helping building behavioral connections in the data. These include (but are certainly not limited to):
- Loyalty Programs
- Key Joining: welding sessions together by chaining multiple keys and cookies
- Email Joining: use emails to drive site visits to tie cookied behavior to emails
- Registration Programs
- iFrames: drop 3rd party cookies to tie experiences together
- 3rd Party Databases: device databases that tie multiple devices/emails to a single user
- Operational Connections: such as creating appointments with online systems to showrooms
- Payment Connections: bridging cards or electronic payment systems
Most of these behavioral approaches provide only partial coverage. That’s important to understand and work with when you consider these techniques. If, for example, you’re a multi-channel retailer and your goal is to understand 100% of the impact of the digital experience on in-store shopping, no single behavioral approach is going to suffice. You’re almost certainly going to have to deploy an array of these techniques and build toward increasing levels of coverage and connection.
That lack of a 100% (or even close to it) connection doesn’t make behavioral analysis impossible. But it does mean you have to be careful how you apply it. A loyalty program, for example, might provide a powerful means of tracking online to offline impacts. But the sample set within a loyalty program is quite different than the broader customer population. That means you won’t be able to infer impact on the non-loyalty population from analysis of the loyalty population.
Sometimes, it isn’t obvious whether a sample set introduces bias or not. Suppose, for example, you identified a population of persistent cookie users over 12 months. You may not know when a cookie has been deleted, but you do when it’s persisted. Such a sample would be useless for understanding attrition (by definition, the population is all retained), but might be a good enough sample for understanding cohort behavior. You might, for example, be able to use this group to understand the health and performance of all customers based on cohorts of when they were first acquired. Is the sample of cookie persisters somehow biased? Hard to know.
Re-Survey Approaches
In the work I described previously, we used Voice of Customer (VoC) techniques. We re-surveyed target populations 3 to 6 months after initial contact. By combining the initial survey data points with the subsequent behavior and the re-survey results, we were able to make connections between behavior and outcomes across a wide variety of problems. This technique nicely bridges both time and device.
The time bridge is inherent. The device bridge can be trickier. We’re dependent on the user’s description of their touchpoints (outside of the Website) to understand what actually happened.
As powerful as this approach is, however, it still has its weaknesses and challenges.
For example, the path to re-survey can be challenging. There are three basic scenarios:
- Re-survey based on known identity email
- Re-survey based on persistent cookie
- Re-survey based on permission/opt-in
Re-survey using known email addresses is trivial. Where you a population of known users, it is simply inexcusable not to be using re-survey techniques as a critical part of your VoC program. Many organizations re-survey without ever realizing it and never take advantage of the fact in their analysis.
Re-survey using cookies requires subsequent behavior onsite and introduces cookie persistent biases discussed above. This type of re-survey will almost certainly introduce some significant biases into the analysis. That doesn’t mean it isn’t worth doing, but it does mean you have to be a little more careful how you interpret and generalize the results.
The same is pretty much true for the last technique, anonymous site permission-based re-surveys. With this technique, you ask for permission to re-survey in the initial survey.
Nearly all online surveys bias toward engaged populations. This bias tends to be pretty strongly reinforced with a permission based re-survey approach. As with the persistent cookie mechanism, this can be managed but it does limit the broad generalization of findings. One advantage you have with re-survey techniques is that you can actually measure the extent to which your opt-in population differs from the general population of online visitors – giving you reasonable methods to stratify your sample and interpret the results more intelligently.
Private Panel Approaches
Panels are one of the more neglected tools in the broader scheme of research tools and I’m not exactly sure why. A panel is a sizable, opt-in based group of consumers who have agreed to provide feedback on a continuing basis. Because they can be carefully sampled, panel members can be incentivized (which is extremely dangerous in broader research settings). But the two big advantages of panels are that they provide a way to execute fairly deep research – you can ask more of a panel member than you can of site visitor – and they provide a way to track changes over time.
Both aspects of panel research are important. For some of the techniques I talked about last week, panels are ideal. Conducting means-end or conjoint research on a panel can be challenging to execute in random samples of online visitors. Panels provide a more committed audience – including research into pure prospects. Nevertheless, it’s in the over-time aspects that panels really shine. Because you’re tracking the same extended group of people, you have the ability to do carefully controlled research into time-based impacts. You can measure evolving brand awareness based on your mass media campaigns. You can use control group invites to measure the impact of Facebook on long term engagement and behavior. You can determine how digital usage impacts broader relationships by segments. It’s important stuff that can be hard to get at without some kind of panel mechanism.
Most of our large enterprise clients run some form of offline panel. For some reason, many haven’t migrated those panels into the online world. That’s a big mistake. Migrating a panel into the online world gives you the opportunity to track panel based behavior across the online sphere. That’s incredibly useful for things like targeting display ads and understanding site use cases by customer segments. It also provides much better insight into the role of digital in driving attitudes and behavior than can be extracted via questions. Asking someone if they use your Facebook page isn’t nearly as useful as SEEING them use your Facebook page.
As with every form of marketing research, understanding your sample is an essential part of understanding how and how far you can use your private panel findings. A well designed private panel has fewer essential constraints than most other mechanisms, making it one of the more attractive options for measuring analytic problems that require time-based tracking.
Choosing the best research method starts, unsurprisingly, with understanding your business problem. The target population, the duration of the tracking period, the need for comprehensive coverage, the desire to operationalize findings, your research budget and timing, and the ability to get at the information with behavioral or survey analysis will all play a role in choosing the best solution. Across most enterprises, it would be surprising if all three methods aren’t routinely needed – making it more a question of when to use each than which to have.
[X Change is filling up pretty rapidly. We’re about seven weeks out and only have about 50 slots left. So if you’re thinking about coming out, please do register quickly. We have usually sold out and this year looks to be no exception! Register here.]
