David M. Raab
DM Review
February 2007
– traffic analysis. These systems track the movement of visitors from page to page within a site. They are what systems most people think of when they hear “web site analysis” and were the primary focus of last months’ column. Products include Coremetrics, story Omniture, seek Google Analytics (formerly Urchin), Unica NetTracker (formerly Sane Solutions), WebSideStory, and WebTrends.
The underlying data for traffic analysis was originally derived from log files and often still is. But today many systems use an alternative approach, applying page tags that notify an external server when a site page is loaded. The primary input is the URL string that requested each page, which provides details such as visitor source and search terms in addition to the actual page name. The systems also use cookies to identify repeat visitors and to link visitor behavior across sessions. Most of these products are hosted services, particularly the ones using page tags. Some also offer installed options.
The primary outputs of traffic analysis are reports showing how often each page is visited, where the visitors came from, and where they went next. This resembles a traditional sales funnel report. It is often used in a similar way to determine where customers are being lost before completing a purchase. But Web behavior is considerably less linear than a basic sales sequence. In addition, customers often make multiple visits for different purposes. The systems address these issues by letting users report separately on different customer segments: people who abandoned a shopping cart, first-time vs. repeat visitors, people who made a service request, people from a particular source or marketing campaign (including pay-per-click campaigns), people who bought a particular product, and so on. It’s important to recognize that segmentation functions vary considerably among different products. Areas of differentiation include what data is available for segmentation, how hard it is to define a new segment, and whether new segments can be applied to previous visits.
– version testing. These systems move beyond observing visitor behavior to actually testing their response to alternative versions of a site. Vendors include Memetrics, Offermatica, Optimost, SiteSpect, and Vertster. All take control of Web pages, deliver different versions to different visitors, and measure the results. Simple “A/B” tests compare one page against another, while the more complex “multivariate” tests break each page into elements and test changes in each element independently. The trick with multivariate testing is that clever test designs—often but not always based on the Taguchi method—can test relatively few combinations of elements but still estimate the impact of each element by itself. This lets marketers identify the most effective combination of elements, even thought that particular combination may never have been tested.
Version testing systems enable marketers to handle test set-up and execution with a minimum of help from Web site technicians. Most systems modify the test pages by embedding Javascript snippets that call the vendor’s server when the page is loaded. The server returns the correct element versions to display. These systems use cookies, sometimes in combination with server-hosted visitor profiles, to ensure each visitor receives the versions appropriate to their segment and past activity. This is important because many tests rely on consistent visitor treatment over time or across multiple pages. All the version testing systems are hosted, although some also have on-premise options.
The version testing systems can also customize content outside of test situations, for example to send different treatments to different visitor segments. Some products can automatically increase the proportion of visitors who receive the most productive options.
Web version testing resembles traditional champion/challenger tests in other media–except that it’s vastly faster, cheaper and more flexible. This opens up new possibilities for refining customer treatments and quickly improving results.
– behavioral targeting. These systems use self-adjusting predictive models to select visitor treatments. Certona, Kefta, Touch Clarity, and [x+1] specialize in on-site targeting, while Accipiter, MediaPlex, and RevenueScience work mostly with off-site ad networks. Some vendors do both. All are hosted services.
These products observe visitor behavior and build statistical models that predict likely next actions. Once a model is ready, they can use its predictions to select related contents, which could be ads, offers, search results, or replies to service questions. Sometimes the systems run structured tests which show different choices to different visitors, but usually they present the same options to all visitors and build their models from the results.
Like the other Web analysis tools, the behavioral targeting systems rely on a combination of current behavior and persistent cookies to gather visitor data. They can collect pretty much any information available to the system, from visitor source to browser type to the sequence of pages just visited. Because their modeling is fully automated, they can consider hundreds of variables for their models rather than the handful that can be managed manually in traffic analysis or version testing. The behavioral targeting models continue to learn over time, so predictions adjust automatically as user behaviors change.
Behavioral targeting systems use similar statistical methods to other types of automated modeling systems. Again, though, the data volume and acquisition speed of the Web requires specialized adaptations and makes possible new uses that are unavailable in traditional channels.
* * *
Leave a Reply
You must be logged in to post a comment.