Predictive Modeling: Benefits to a Pixel-Free Approach 2The Use of Tracking Pixels 3. Maintaining pixels remains troublesome: In theory, pixels can be very simple, but in practice, they providein Online Advertising several notable challenges. They’re a cumbersome and inefficient way of managing ad campaigns – choicesSince the first e-commerce marketing professional must be made about which pages to pixel, leadinglooked to analyze shopper behaviors in order to drive to an inevitable loss of user behavior capture. Oncemore transactions on his web site, the only way to gather they’re there, they’re often difficult for the IT staffthat information has been by implementing tracking pixels. to maintain. All of this adds up to the fact that mostThese are small, transparent images embedded in websites. companies get a fairly small coverage of their websiteWhen users ultimately browse such a site, information with pixels. That decreases the effectiveness of theis gathered about what they are doing – how they’re information advertisers have and limits their optionscoming to the website, what they’re looking at, what as to what to do with that data.they’ve browsed, put in a shopping cart, abandonedor bought. Tracking pixels provide both a high-level and(anonymously) individual view of certain behaviors on Understanding Predictivea given web site, and they lead to an ability to optimizebased on those user behaviors. In the case of advertising, Analyticstargeted advertisements based on those behaviors can Predictive analytics have been used by different indus-then be served to users. tries for years to solve difficult problems that range from detecting credit card fraud to determining patient riskHowever, there are three key issues surrounding pixels levels for medical conditions. It combines data miningand online advertising for e-commerce marketers. and machine-learning technologies to create statistical models based on historical data. It then uses these1. Site performance directly affects e-commerce models to predict future events. Extracting the power conversion rates: In-house research shows that from the data requires powerful algorithms behind companies who use Akamai’s Dynamic Site Accelerator, predictive analytics. a site performance enhancer, enjoy conversion rates that are fully 23% higher than those who do not. One increasingly popular way to apply predictive (3.69% conversion rate using DSA; 2.88% without). analytics to online advertising is to look for patterns For an online retailer with five million monthly unique in user shopping behaviors that indicate likelihood to visitors and an average order value of $100, each buy particular products or services. Powerful relational percentage point improvement in conversion rate databases seek out correlative purchase activities; these would result in incremental revenue of $5 million can be intuitive (a blouse purchase leading to a jewelry a month. And according to Steve Souders, Google’s purchase) or non-intuitive (auto parts leading to music web performance guru, when online retailers shave downloads). The more data that gets fed into the algo- up to 5 seconds off each website transaction, they rithms, the stronger those algorithms are likely to be. In can boost their conversion rate up to 12%, double the case of a shopping data co-operative, each company the number of engagements from search engine is incented to ‘co-operate’, sharing their data in return marketing, and slash by half the number of servers for use of everyone else’s. required.2. Adding tracking pixels hinders e-commerce site performance: In reviewing 15 different companies who currently advertise online and use pixels to track visitors, the average home page load time was slowed down by 39% – comparing the time it took the home page to load with pixels and without, with a range of an 8% to a 99% slowdown in home page load times.
Predictive Modeling: Benefits to a Pixel-Free Approach 3By analyzing the historical data, these models learn – across thousands ofdifferent shopping behavioral variables – how consumers act when they arenearing the target transaction. The models determine which variables are mostimportant to the particular transaction in question, and how to combine thesevariables into a scoring calculation that achieves the best predictor of behavior.Every model is unique.What should be noted is that, typically, no single variable ends up beingweighted more than about one-tenth of the overall score – meaning there is Did you know?no single data point that predicts shopping behavior with consistency. Generally, Typically, no single variablethe best predictors include a long tail of variables, reflecting the inherent complex- in a predictive model endsity of modeling human behavior. So, for example, knowing that someone put a up being weighted more thansurfboard in their shopping cart in the last 24 hours is not enough to accurately about one-tenth of the model’spredict that they are likely to purchase that surfboard. On the other hand, know- overall score–meaning thereing that they had also recently booked a trip to Hawaii and purchased a wetsuit is no single data point that predicts shopping behaviorsubstantially increases the likelihood of a near-term surfboard purchase. with consistency.The predictive engine gives each user (that is, each anonymous cookie) a scorethat reflects the probability that the user is in-market for the given product. Itdoes this by looking at the user’s most recent behaviors and computing a scorebased on the variables indicated by the model. The more behaviors that are fedinto the predictive engine, the better the prediction tends to be. Scoring shouldbe done continuously to incorporate users’ most recent actions. This is critical,as consumers continually move in and out of market, often within a span ofjust a few days.Maximizing the Potentialof Predictive AnalyticsWhile it’s true that the sheer number of user behavior data points increaseswhen employing a pixel-free approach, the real value to such an approach isin the improved richness and variety of user behavior data.Take, for example, a company in the auto parts business. Initially, they placed atracking pixel on their home page, and our predictive models would understandsomeone coming to that site as an auto enthusiast. Correspondingly, our modelswould be able to draw some inferences from that visit, but miss out on manymore. As they’ve added pixels, we’re now able to see if someone is interestedin brakes, floor pads, spark plugs, etc. However, in contemplating a pixel-freeapproach, the predictive models can become exponentially better, and we caninfer (anonymously, of course) that someone is looking at brake pads, and infact by the combination of pages and products he’s looking at, that this personis a do-it-yourselfer. Now, with many incremental data points, the models nowreally have something to work with. Instead of simply understanding purchasesat a product category level, the models now understand product details, com-binations of products, pages browsed, etc to arrive at a much more in-depthunderstanding of someone’s purchase intent. Combining this level of detailacross many different companies within Akamai’s data co-op can make ourpredictive algorithms much stronger and more effective.
Predictive Modeling: Benefits to a Pixel-Free Approach 4Measuring Pixel-Free Site Visitor CaptureAdvertisers generally have to decide the proper number of pixels to place on their site to generate the most usefuldata about what their site visitors are doing. Placing pixels on even the most obvious pages captures only a smallpercentage of unique users. Our analysis across four advertiser verticals shows that pixeling the home page capturesas little as 18% of all unique users to a site, and most pixel strategies in their entirety only capture about half of a site’stotal visitors. Keep in mind that this level of pixeling a site can take months of effort to complete, and can be difficultto optimize in a meaningful timeframe. Key to Charts Below: • The 30-day estimate reflects the estimated total • Pixelling Strategy indicates the particular set number of unique visitors captured in the 30-day of pages on a Web site on which tracking pixels period in which the report was run. might be placed. • Lost Users indicates the total number of site visitors not captured by the respective pixelling strategy.Office Supplies 30 Day Estimate Lost Users Pixelling Strategy 6,254,592 0% All Users (Pixel-Free) 2,610,176 58% Product pages, Self-Fulfillment Pages, Category Pages, Checkout Process and Home Page 2,088,960 67% Self-Fulfillment Pages, Category Pages, Checkout Process and Home Page 2,051,072 67% Category Pages, Checkout Process and Home Page 1,428,480 77% Checkout Process and Home Page 1,229,824 80% Home PageThis office supply company uses a rather robust pixeling strategy as they’re interested in a large variety of userbehaviors across the site. However, their current pixeling strategy misses 58% of visitors coming to their site.And their home page pixel only captures 20% of all user activity, indicating that most people coming to thesite are driving directly to specific actions, or to purchase specific products.
Predictive Modeling: Benefits to a Pixel-Free Approach 5Perishable Gifts 30 Day Estimate Lost Users Pixelling Strategy 273,408 0% All Users (Pixel-Free) 141,312 48% Main Departments, Checkout Process, and Home Page 126,976 54% Checkout Process and Home Page 87,040 68% Home PageThis company also employs a large-scale pixeling strategy as they have 47 separate departments which are taggedin addition to the checkout process and home page. However, with this strategy they’re missing more than halfof all visitors to the site.Home Goods 30 Day Estimate Lost Users Pixelling Strategy 3,230,730 0% All Users (Pixel-Free) 2,606,080 19% Product pages, Search, Department pages, Sub-dept pages, Checkout Process and Home Page 2,470,912 24% Search, Department pages, Sub-department pages, Checkout Process and Home Page 2,382,848 26% Department pages, Sub-department pages, Checkout Process and Home Page 1,465,344 55% Department pages, Checkout Process and Home Page 1,184,569 63% Home PageThis home goods company puts tracking pixels on more than 250 different pages on their site, and as a consequencecaptures more than 80% of all users coming to the site. But as pages change (sales, seasonal offerings, new products,discontinued items, etc), their IT department must stay abreast of a particularly challenging pixelling initiativeand they suffer potentially slower load times on those pixelled pages.Auto Parts 30 Day Estimate Lost Users Pixelling Strategy 1,651,712 0% All Users (Pixel-Free) 1,244,544 25% Product pages, Department Pages, Checkout process and Home page 695,296 58% Checkout process and home page 295,936 82% Home PageThis auto parts company is a sophisticated user of online display ad technologies, and many of their site visitors cometo the site via dynamically generated display ads taking into account what auto parts they’ve shown interest in. Theyalso drive traffic and corresponding behaviors with these ad strategies, more so than many other online retailers. Yet,their pixelling strategy still manages to miss 1 of every 4 visitors to the site.