[Hiring Note: We are not only far busier here at Semphonic than we’ve ever been but growing so fast I’m having a hard time keeping track of every one new. The first trend is still ahead of the second though, which means we are looking for people; if you are a very sharp Jr. Analyst in the SF Bay area (or are willing to come to our fair city) with aspirations to be more than a reporting jockey or an Omniture tagger, drop me a line! We have multiple entry-level positions open. Yes, there are body-shops and jobs everywhere in our industry, but there are hardly any places where you can learn real Digital Analytics practice at a deep level.]
For almost the whole of this year I’ve been working on an extended series about the gradual convergence of Database Marketing techniques and Digital Analytics. Over the past four or five posts, I’ve delved into a topic I consider a pivotal part of that discussion: Semphonic’s Two-Tiered Segmentation approach.
Segmentation is pivotal because it’s a – really the – fundamental technique in database marketing. Traditional DBM shops have scores of professional statisticians who do nothing but build segmentations for targeting purposes. On the other hand, Segmentation in Digital Analytics has always been a lot like bipartisanship in politics – much talked about but rarely practiced.
This isn’t a simple matter of neglect. Traditional segmentation variables just don’t exist in many Digital contexts – and analysts have struggled to find a good alternative. This series is really about recovering those techniques in a digital world. I’ve talked extensively about the role of Meta-data in Digital Analytics. Meta-data provides a means of inferring segmentation like variables from page views; it’s a critical bridge to developing better segmentation techniques. Even more fundamental, however, is the Two-Tiered Segmentation.
The main idea of Two-Tiered Segmentation is that Digital Segmentation needs to include more than the Visitor Type – the traditional persona/business relationship type of segmentation that exists in traditional marketing. The 2nd Tier is the type of interaction – what the visitor was trying to accomplish in a Visit.
At Semphonic, we believe that every aspect of Web analytics needs to be framed inside this type of segmentation; a segmentation that answers the “Who” and the “What” for every metric. In the past several posts, I’ve laid out some typical Two-Tiered Segmentation schemes across various industries.
These aren’t meant to be definitive. Good segmentations are necessarily business not just industry specific. They are meant, however, to give a sense of what’s involved in a Two-Tiered Digital Segmentation.
When experience analysts look at these segmentations, their most common reaction is “That’s cool, but how do you build the Visit Intent Segmentation?”
It’s a great question.
There is no simple, foolproof way to classify visit intent (what a visitor was trying to accomplish) using only behavior.
In this post, and in the next several upcoming ones, I hope to show how it can be done and some of the tricks and methods we use to make it as robust as possible.
Here’s the basic process we follow for creating the 2nd Tier (Visit Type) Segmentation:
I’ve shown the process as a cycle because it’s inherently iterative.
We start by postulating the existence of various Visit Types. These may be based on existing Site Design Use Cases (which is great) or they may simply be the result of a site-walkthrough and discussions with the site owners / designers. Once we’ve created a conceptual set of Visit Types, we try and construct filter definitions that might isolate the type. This is tool specific and involves a variety of techniques.In most cases, we’ll try and make the definitions hierarchical and mutually exclusive. In other words, if you’ve been classified as having a “Job Seeking” visit type, we won’t also classify you as having a “Product Purchasing” visit type for that same session.
The next step, of course, is to see what happens when we actually apply those filters. Believe me, they aren’t always successful. Sometimes, when we apply a filter, we simply don’t isolate ANY visits. At other times, we can see that the filter isn’t exclusive enough; when we look at the actual behaviors we see things that suggest other visit types. If possible (meaning if the enterprise is running and has integrated their VoC data into their Web analytics data), we’ll also look at how closely our behavioral filter matches the VoC task descriptions and whether these map well to our conceptual Visit Type. The use of VoC data in this type of analysis is a topic unto itself; it’s certainly a key method not only of creating but of validating your definitions and building organizational consensus that the segmentation approach is valid.
Finally, once you’ve created your filter set, the last step is to look at how large the set of visits is that aren’t classified. Typically, this involves a last segmentation to create the “NOT” set. With that “Not Classified” visit population, you go back into the process and try and unearth visit types that might apply.
The whole process keeps going until you’ve classified almost very visit type. We find that as a practical matter, no more than 2 or at most 3 iterations are ever necessary.
I’m going to jump into what I take to be the most challenging step – Filter Creation.
Here’s the Visit-Level Segmentation for one Visitor Type (Consumers) on a Technology Site with various components including ecommerce, customer-support, supply-chain, branding, and marketing operations.
We did a separate more detailed break-out of the Customer Support visit types, but here we were primarily concerned with excluding those Support visits so they didn’t muck up reporting and analysis our other visit types.
The first step in creating a set of filter definitions is to create a mutually exclusive hierarchy. There are some basic principles for creating a hierarchy:
1. Classify all definitive behaviors: where a behavior is definitive or changes your perception of all other behaviors, it should be at the top of the hierarchy
2. Classify the most common and important behaviors: where a behavior is more important or much more common than the alternative interpretations, it should be above those alternatives in the hierarchy
3. Classify by precision of cues: where a cue is more precise (such as Search Term vs. Page View), classify using the most precise cue
4. Use the Firstest or the Mostest: What a visitor does first or most is the best indication of intent
So let’s start building a Visit Intent hierarchy for the segments above.
Typically, I’d organize them something like this when starting my filter building:
- Job Seekers
- Warranty Registrants
- Support Seekers
- Enterprise Researchers
- Discount Shopper
- Potential Buyer
- Early-Stage Shopper
- Background Researcher
Why?
Job Seekers, Warranty Registrants, and Support Seekers are all Visit Types with strong and overriding behavioral cues. If you visit the jobs page, you’re a job seeker – almost no matter what else you do. Indeed, it’s typical of job seekers to look at all sorts of random content on the Web site. None of that matters. If they look at jobs content, they are a Job Seeker.
Warranty Registrants are a slightly different case. The vast majority of Warranty or Product Registration visitors come to the site with that specific visit purpose in mind and do nothing else. Typically, this results in a pretty clear behavioral signal. On the other hand, not every site makes the path to Registration crystal clear. You might go into Product or Customer Support Detail to get to product registration. But if you end up in Warranty or Product Registration, we’re likely to classify you with that visit intent.
Support is my third case. We’ll likely classify any visit with significant support content as a Customer Support visit – but we have to be a little careful with this. Prospects will often check out support content, and support content often hides additional types of visit functionality (such as Enterprise Research content). Nevertheless, the general principle is that Support intent overrides shopping or research intent as a likely indicator.
I put Enterprise Researchers next for this site. The enterprise products from this company aren’t purchasable on the Web and are highly unlikely to be of interest to any other class of consumer. Because of that, I felt they comprised a distinct and fairly definitive behavioral signature. If you looked at enterprise content, we felt you were highly unlikely to be in the traditional buying funnels.
Discount Shopper came next. Like most consumer technology companies, most sales for this client are in-channel not direct via the web. That means that discounting of remaindered items is a significant online function. In this case, there was significant discount specific content and most discount visitors arrived via alert emails. We felt anyone arriving on that type of campaign had clear Discount visit intent. In addition, if there first action off the home page was into a discount section, we felt this indicated specific intent. On the other hand, a visitor that went first into product detail and then to discount we decided to classify as a Potential Buyer (the next segment down).
This illustrates Principle #4 from our list – what a visitor does first is highly significant for classification.
Probably the hardest rules to define involve trying to separate out where in the buying stage a visitor might be. Are they doing early product research, background research or are they a potential buyer? What makes this particularly challenging is that it’s hard to separate out a failed potential buyer from a successful early-stage researchers – especially for a multi-channel enterprise. This is where meta-data comes in; the classification of content by buying stage and then using counter variables (e.g. Omniture eVar counters) to measure consumption by stage is one of the best techniques for parsing out these visit types and it’s a perfect example of where understanding your segmentation requirements is a must when creating an implementation design.
I’ve chosen to categorize Potential Buyer as the highest level because it’s better to capture ambiguous visits as belonging to the most important type.
Now it’s time for Filter Creation. Here’s a quick look at some of the filters build using Omniture Discover. Starting at the top of the hierarchy, here’s our Jobs Visits filter:
Note that the 1st Tier Segmentation (Consumers) is present in every 2nd Tier Segmentation. One very unusual aspect of this particular filter is that we opt to use a Visitor scope for the “jobs” filter – that’s extremely rare for a Visit Type Segmentation (and we don’t always do it even for jobs). However, our thought in this case was that ANY visit by a known Job Seeker in the period is probably Jobs related. So this is an unusual case of mapping a Visitor-level segmentation across a Visit intent filter.
For each subsequent segment, we start building out Exclusions to make sure that our segments are mutually exclusive:
With each new Segment, we have to add criteria to the exclusion list:
For the segments at the bottom of the hierarchy, this means a list of Exclusions that is often 10-20 blocks long. It’s cumbersome; particularly in tools that don’t allow you to simply drop in a previously defined segment. But however long it takes, it’s a one-time activity and it’s worth it to create a powerful Visit-Type Segmentation.
The segmentation shown here for Discount Shopper takes advantage of one of our implementation tricks:
We like to create a variable that includes Page Depth – sometimes on its own and sometimes as part of PageName. This allows us to identify in the Segmentation engine when something happened as the first click (or at any point in session we choose). Here, we are looking for Discount Shoppers – and defining it as any visit that starts on a Discount Page or that goes to the Discount Page on the first click (using our Page w Depth variable). This is a perfect example of how proper implementation enables segmentation.
We also like to setup counter eVars based on classifications of content type. This allows us to do much crisper segmentation of sales stage populations:
This type of segmentation allows us to target visits with heavy Early Stage content and exclude visits with heavy Late Stage sales content.
There is no one method for building a Two-Tier Segmentation. The techniques I’ve shown here: mutual exclusivity, Visitor and Visit scoping, entry page, first click, and content meta-data with counters are some of the more common and interesting techniques. Internal and external search terms, product order types, visit number, campaign type and whole host of other variables are equally relevant and often used. The goal is to be as precise as the behavioral cues on the site will allow, and, of course, what the site will allow is a function of the quality of your measurement implementation.
These segments, once created, will be used constantly for everything from Management Reporting to Analysis to Targeted Marketing. Spend the necessary time on them – they are worth doing right.
In my next post, I’ll discuss some of the techniques for validating your Visit Segmentation – both to yourself and the rest of the enterprise.
