On December 4, 2008 the WAA presented the Unica sponsored webcast: 5 Do’s and Don’ts for Behavioral Segmentation, Targeting, & Interactive Marketing. The full presentation of this webcast is available here, but you must register or login to view. This was an extremely informative session that drew many questions from the audience - and now here are the answers from our expert panel!

Gary Angel, Semphonic

Q: You mentioned that web analytics tools normally don't have capability to carry out advanced analytics needed for behavioural segmentation. Can you elaborate?

A: The type of segmentation I’m describing is machine-based segmentation. Typically, it uses either cluster, factor or neural network analysis to group visitors based on a complex set of variables. This is as true for traditional (offline) segmentation as for behavioral segmentation. Web analysis tools typically only provide filtering capabilities. They let an analyst define individual segments, but they don’t help the analyst understand the rules that might be appropriate, the related behaviors that are significant, and they often don’t provide enough logic in the filtering mechanism to create many different types of segments. So to do this analysis, we’ll typically use relational database systems (like SQL-Server or Oracle) to pre-process and transform the data and then tools like SPSS, SAS, neural-network tools, or advanced data analysis tools like Clementine to actually create the segmentation.

Q: Is it better to start with the Behavioral Segmentation or with the survey data? Several companies promote starting from the VOC data. How true is this?

A: I disagree with it, but a case can certainly be made for either direction. The fact is that most traditional segmentations have relied on survey data. So VOC vendors have a real point. However, if you start with VOC data, you lose, in my opinion, the ability to use the segmentation across all web visitors. Some people think you can profile behaviorally and map behavior to VOC-based segments – and this is, in theory, possible. However, we’ve had a terrible time making it work in practice. To some extent, this may also depend on your website and audience. If you have a rich website and lots of engaged behavior, I’m a firm believer that a behavioral segmentation is actually far richer and more interesting than a VOC-based segmentation. But we just did an analysis on a very new site that turned out not to have very interesting behaviors at all. And probably the most interesting part of the analysis was the VOC data. So I’m not sure there is one right answer. But I do have a general preference for beginning with behavior analysis.

Q: The majority of visits to all sites are single page visits. Do you include those in your development of segments?

A: Great question. Typically we don’t. All the data transformation and statistical wizardry in the world won’t let you say anything interesting about one-and-done visitors. Every once in a while, these visitors are worth including because their source or landing page may actually segment in some interesting fashion. More typically, however, I’ll exclude these. One benefit to excluding them is that you dramatically reduce the amount of data you have to process.

Q: If I were to do segmentation for the first time for my company, how long does it take, i.e. what can I tell my boss about what they should expect?

A: I’ll give you the mealy-mouth “it depends” answer first. What it depends on is the size and complexity of your site, the cleanliness of your web data, whether you want to include VOC data, and what web analytics tool you’re using. If you want to include VOC data and you’re not already running and integrating VOC data, you’ll have to do that. Then you’ll have to wait while you collect enough data to match to your behavioral time-period (typically at least a month). The actual process of analysis is probably about 6 weeks, but like most such projects it can easily be made much longer if hurdles exist in getting, processing or understanding the data.

Q: Do you ever use statistics to discover segments?

A: Yes. See my answer above regarding web analytics tools.

Q: How to do that segmentation? what tools?

A: I partially elaborated on this in my answer to the first question. However, a little more detail may be helpful. First, the most common technique for doing visitor segmentation analysis is clustering. Clustering is supported by a number of statistical packages. SAS is surely the most common, though we frequently use SPSS as well. In addition to cluster analysis, Factor Analysis is also used – sometimes in conjunction with clustering to reduce and simplify the variable set. We’ve also used Kohonen Maps with proprietary technology – and that’s a technique that’s worked very well for us in some cases. There are also data-mining tools like Infocentricity and Clementine that provide a range of tools to create segments.

Q: How do you apply this type of segmentation so that you can map your web users into the appropriate segment moving forward?

A: This is a real problem given that you can’t do the analysis in the web analytics tool. When you produce the segmentation, part of what you’ll get is a mathematical description that allows you to assign a visitor to a segment based on values within the variable set. If you are warehousing the data, you can probably take this equation and use it to assign segments directly as you update your data or in regular batch runs. If you aren’t warehousing the data, you’ll need to create a processing loop whereby you extract the data, transform and normalize it, set the code based on the equation and then pass the code back to the web analytics solution. It can be a lot of work.

Q: Where can we get more information, reference, training on Behavioral Segmentation? Thanks!

A: Damned if I know! If you’ve done traditional (demographic and psychographic) segmentation, you probably know enough to get started because you probably know almost as much as any of us. If you haven’t, I’d suggest you first consult the extensive literature on visitor segmentation based on survey data. I have been blogging on behavioral segmentation if that’s any help (http://semphonic.blogs.com/semangel/)

Q: On Slide #16, are the variables used to describe the segments the very same ones used to construct the segments?

A: Yes and no. Most of the variables used to describe the segments are also the ones used to build the segments – and that’s definitely the case on Slide #16. These variables, by the way, were constructed from the web analytics data – they aren’t something that existed independently or were fashioned from survey data – hence my point about data transformation. However, I do have to make a couple of caveats. First, many variables that go into the analysis aren’t used in any one profile and some end up not being used in any profile. So what’s shown is a subset of the variables used. In addition, we do sometimes (for instance with VOC data) profile segments using data that isn’t used in the segmentation at all. This particular segmentation didn’t include survey data integration so we didn’t do that. But if we had, then we would also have profile variables that weren’t part of the segmentation scheme. There are other cases besides VOC data where we might do this as well – but it’s not real common.

Q: What package are you using to display the data for the profile of the snake eyes segment?

A: Pretty much all the report examples are in Excel. The profile report is built in Excel 2007 using Conditional Formatting rules. Like all web analysts we are heavy users of Excel. Total non-related vent, but I must say it drives me crazy that we work with many large companies that are still using Excel 2003 when the 2007 version is better in every respect. It’s hard for me to understand how core personal productivity tools like Excel are allowed to lag by five or six years.

Q: Is there an optimum volume (or % of visitors) of survey responses that is necessary to overlay onto web analytics when building a segmentation?

A: As with all statistical analysis, there isn’t one optimum number but more is better. Not only does more improve the likely accuracy of your conclusions, but it allows for much finer grained analysis. In the traditional marketing world, survey sizes were limited by cost and rarely exceed 5-10K with numbers in the hundreds being common. There’s no reason to put such sharp volume limitations on the web assuming your site drives significant volume. Behavioral segmentations will often create (as I mentioned), micro-clusters that involve only 1% or so of your user-base. Given that, if you want to profile them demographically and you want to have a decent set of respondents in that segment, you’ll have to poll quite a few visitors. Fortunately, there’s really no reason not to. I’d certainly see nothing untoward about polling many, many thousands of visitors on a fairly well trafficked web site.

Q: How much daily traffic should your site have to make the survey statistically viable?

A: See above. There are really two questions here. It’s easy to get to some level of statistical viability for use of survey data on your web site provided you aren’t going to cross-tabulate it with behavior or combine too demographics into complicated profiles. Political pollsters do a respectable job predicting an election involving 100 million people with a survey of 500. But they are generally only looking at a single variable or, at most, some basic cross-tabulations like gender by candidate preference. When you cross-tabulate with behavioral data, you’ll almost always be looking at finer grained groups of people and you’ll still need enough survey respondents to profile with reasonable confidence. That can quickly translate into quite substantial survey volumes. A second factor here is how much you want to impact user experience. Many sites are reluctant to ask a high-percentage of visitors to participate in a survey. So if the site has low volume, you’ll have to be more aggressive about asking.

Q: You mentioned that Behavioral Data can be enhanced by Age, gender etc, some of this might cross privacy lines, won't that be a problem?

A: It can be. Every organization has different rules & regs about this. When you are using survey data in this fashion, you aren’t using it at the PI level – in other words, even though you’ve tied the survey data to the analytics data, both are still anonymous and the analysis is at the level we call NFA (nameless, faceless, aggregated). That doesn’t necessarily make it okay for everyone’s privacy policy, however, and before integrating VOC to behavioral streams you definitely should check this and make sure whatever you are doing is either already okay or updated in your privacy policy.

Q: Can we have behavioral targeting without integration with the web analytics tool? Without segmentation? How useful can each option be?

A: You can and often will have behavioral targeting without integration with web analytics or visitor segmentation in the sense that I’ve described it. Much depends on the way you choose to implement behavioral targeting and the type of web site you have. Certain kinds of behavioral targeting demand visitor segmentation. But some behavioral targeting depends on tools that use real-time behaviors to map visitor preferences and respond most appropriately. These tools don’t really depend on either visitor segmentation in the sense I described or on web analytics. However, other kinds of behavioral targeting are driven by web analytics tools and are greatly improved by having a rich visitor segmentation. For sites that can benefit from real-time preference targeting, I don’t think there is any substitute for the type of behavioral targeting tool that essentially does its job on its own. But for broader uses of behavioral targeting and remarketing, I think both web analytics tools and visitor segmentation are essential. I look forward to reading Akin’s answer as well!

Anil Batra, Ascentium

Q: How concerned are you about the new releases of anonymous browsers?

A: Behavioral Targeting depends on ability to identify a visitor. If we are not able to indentify a user uniquely then that’s an issue. As anonymous browsing becomes more prevalent I am sure behavioral targeting vendors with come up with new technology to counter that. On the other hand, if your customer trust you the you should not be affected by anonymous browsing.

Q: What about the cookies? how to be sure about users not deleting their cookies?

A: As of today cookies are very critical to Behavioral Targeting. In most cased, behavioral targeting can not work without cookies. We can not entirely avoid cookie deletion. If you use first-party cookie then cookie deletion should be a small factor. However to use retargeting services your will need a third party cookie which has a higher deletion (or blocking) rate.

Q: Is behavioral targeting based on cookie info? what happens if the visitor deletes the cookie?

A: See above

Q: Regarding getting customers to opt-in to BT, what kinds of incentives or messaging has been shown to ease customers' privacy/anonymity concerns?

A: It is highly dependent on your business and customers. I don’t think incentives like “20% off on your next purchase if you let us target” will work. However, you should provide your users a clear message about how they will benefit from it. You don’t have to use the word behavioral targeting in your messaging. Something like this should work: “By collecting your browsing and purchase data we will be able to provide a personalized experience on the site. This will also allow us to provide you’re the products and offers relevant to your needs. You can always opt-out by clicking on…..,. . We stop collecting data once you opt-out.”

Q: What analytics tools do you recommend

A: Most of the analytics tools in the market today allow you to segment your users. You will need a behavioral targeting, optimization tool or vendor to help you with Behavioral Targeting. Omniture Test& Target, Optimpost, Widemile etc. all let you do behavioral targeting.

Q: If Dell is sharing individual behavioral information with a partner like Parade, what does their privacy policy need to describe - is it too broad for consumer comfort?

A: Privacy policy can be very simple as I mentioned above. You will also need a statement regarding use and share of data with 3^rd party ad networks. A statement like “ We partner with 3^rd party ad networks/partners to provide you a personalized experience and information about our products and offers even when you are not active on our site”. For exact wording get your legal department involved. The statement should be very clear and up to the point.

Q: What type of tool is there in the market that will help me do behavioral targeting on my site?

A: You can do behavioral targeting with few lines of codes yourself (not very scalable) or can use tools like Test&Target, Optimost, Widemile, etc. There are several tools/vendors out there. First, you need to figure out your goal and the segments your want to target before you try to decide what tool to use. Follow the 5 step process that I outlined in my presentation.

Q: Is all of this based on cookies? If so, what about those that use multiple browsers?

A: Yes it is based on cookies (see my previous answers). If they use multiple browsers then just like web analytics will count them as different visitors they will appear as different visitors in Behavioral Targeting. Some companies are trying to find a ways to bridge that gap.

Q: When users share a computer - like a family or SOHO, how do we identify this? Can we?

A: Most of the behavioral targeting solutions will treat them as one users, it is the same issue with web analytics tool.

Q: What level of behavioral segmentation is realistic without capture of Visitor IDs (e.g. through log-in)?

A: Most of the behavioral segmentation is anonymous i.e. through anonymous cookie id, you don’t need a login id to do behavioral segmentation. Collecting PII data will allow you to enhance the behavioral data but it is not required.

Q: You mentioned that Behavioral Data can be enhanced by Age, gender etc, some of this might cross privacy lines, won't that be a problem?

A: Yes any time you try to use PII data it could be a problem. However, if you are clear in your privacy policy that you will use the data for providing a better experience for customers and provide them a choice to opt-in with easy opt-out then that should take care of any privacy concerns.

Q: Hi! I work for a news website in Brazil and we ask for registration in some areas. We think demographic data from the registration will aggregate value for future actions. For you which are the most important information we should ask in subscription forms and what is the better way to obtain this kind of information: in steps, short forms, or you believe that is better forget about registrations and do behaviour targeting only using data from web analytics tools?? Thanks

A: Behavioral Targeting by definition is based on the behavior data i.e. site usage data including referring sources. So I suggest you start from there.

As far as what registration data you need to collect really depends on your needs, there is no right or wrong fields. Some companies might want to collect gender while other don’t care as much about gender but have a need to collect profession because that’s how they segment their users. Start with goals, it will reval what segments make sense and then collect relevant data. I advice to only collect the data that is required.

Collecting the data: I suggest that you try different kind of subscription forms and do A/B or Multivariate testing to figure which type of process works better for you.

Akin Arikan, Unica

Q: How should heat maps and overlays be used efficiently in behaviour segmentation?

A: The most interesting question to explore with the heat map visualization in regards to segments would be to see if usage preferences differ by segment. This question is difficult to answer though. Granted, many web analytics solutions can create heat map labels by segment so that you can compare the behavior of different groups, e.g. people who arrived searching for digital cameras vs. TVs. But the segments that Gary is talking about are at a higher level. A data mining solution was used to mine the web data in order to come up with his segments. So, in order to use the web analytics solution’s heat maps the results of such segmentation analysis would need to first be mapped back to the web analytics solution. In the case of Unica’s NetInsight that is done by providing a mapping table of visitors to segments back into the solution.

Q: Why not do emails regarding abondened shopping carts - isn't that targeted?

A: Absolutely. Also referred to as re-marketing or re-targeting, it is a basic example of behavioral segmentation and targeting. Something to take into account however would be whether the visitors meanwhile completed the purchase by calling the call center (or similar). If they did, then a re-marketing email with a 20% discount voucher would do more harm than good. Another thing to take into account would be customers’ life time values and purchase propensity scores. For example, a loyal customer who is very likely to make repeat purchases doesn’t need a 20% discount voucher to be convinced. But a customer who tends to be on the fence may be pushed over the fence by such a voucher. Apply direct marketing know-how in designing your abandoned shopping cart programs.

Q: What fields do you use to map online and offline data? Can you do this for prospects as well as customers?

A: In general, you’d use a login name whenever available and a cookie as a fallback mechanism. The specifics depend on each site’s business model. For example, in Finance and Telcos you frequently have visitors that login. In retail and travel even your existing customers don’t login until they are ready to make their next purchase. So here it is extra important to draw on cookies. As for prospects, it will depend on the business model again. For example, in most business models prospects are anonymous. But if you think of auto insurance quotes, you usually give some contact info on page one before continuing to subsequent pages for your vehicle details etc. So even if you abandon after the quote, your contact info would be available. In many cases however visitors may enter bogus info especially if they already know that they only want a quote and aren’t ready to purchase. In B2B, on the other hand, you can often derive the company name from the IP address. While you won’t know who the contact since the registration wasn’t completed, knowing the company name of a prospective buyer goes a long way in B2B.

Q: What tools/services are available to gather prior web search behavior for users?

A: There are two methods that come to mind. Method one is to use panels or CompIntel solutions (e.g. Hitwise, Comscore, Nielsen, Compete.com) to study the search behavior of a representative subset of Internet users. To that end one can also make use of tools provided by search engines, e.g. those from Microsoft adCenter Labs. The Search Funnel tool is very nifty. It shows what series of keywords people tend to use when they perform repeated searches.

Method two is that web analytics can build a search history of individual visitors as long as the visitors clicked through from a search results. Web analytics is blind to searches that individuals make if they do not click through.

Q: I'm interested in the report that speaks favorably of Unica's web analytics product. Please direct me to it.

A: The specific report that was mentioned was the 2008 Web Analytics Buyers Guide by JupiterResearch. It contains Jupiter’s thorough research results for the whole landscape of web analytics vendors, both for enterprise users and SMBs. This report can be downloaded from Unica’s web site @ http://UnicaWebAnalytics.com. The download is free but requires registration.

Q: How do you proceed with Behavioural segmentation for B2B websites...which are not going to sell Online. Doesn’t that limit what variables we can look at?

A: The biggest difference with B2B is not so much that people don’t complete their purchase online. After all, B2B web sites are all about enticing visitors to complete lead generation forms in return for valuable information (e.g. you can download the Jupiter Research web analytics buyers guide @ http://UnicaWebAnalytics.com but only after registration). You can do all sorts of segmentation to see what kind of visits lead to completed forms while others don’t.

Granted, the big question is which leads turn into sales and of what order values. Answering that question does require integrating web analytics data with sales data. But there is controversy in how far Marketing should hold themselves liable for sales. After all, so much goes into a sales cycle that isn’t under the control of the Sales team.

In my humble opinion though, the biggest difference between B2B and B2C is that buying decisions by B2Bs are made by multiple people in the organization. One person may research the web site, register, and download things. Yet somebody else may call in and be the contact for a sales cycle. Therefore, crediting the marketing source with the outcomes requires joining the sessions at the account level rather than cookie or contact level.

Many thanks to our panel and our sponsor Unica!

Comments

[web analytics]

Great interview and post. I learned a lot:)