Strategies for Personalization – Delivering an Extra, Unexpected Treat


Share on LinkedIn

Personalization is the new frontier for digital analytics. It’s the single most powerful, analytically driven tool in the marketing (and customer service) arsenal. Digital is fundamentally a direct response medium. The companies that succeed big in digital have almost all found ways to create unique and highly personalized experiences. But while market leaders have successfully deployed deep personalization, most of the rest of the market hasn’t even scratched the surface of the opportunity. Challenges in technology, understanding, and methodology have limited and often crippled attempts to drive effective personalization. In this series, I’m exploring how analytics can be used to solve personalization challenges and how today’s enterprise can finally solve personalization.

Before I dig deeper into the challenges around personalization (and the solutions we’ve tried and found), it’s worth walking through the different types of personalization strategies and scenario’s that commonly exist. This walk-through isn’t meant to be comprehensive, just a quick survey of some of the more common approaches.

Content/Product Related Personalization

You looked at X, so there’s a good chance you might be interested in Y. Whether it’s media content, ecommerce products, or stocks on the NYSE, viewing behavior is a powerful predictor of immediate current interest. Note the emphasis on current. Content-based personalization is unique and powerful precisely because the single most important variable is what you just looked at (or bought). There are two main approaches to content-based personalization: atomic and categorical. Atomic approaches require little if any meta-data about the content. In most cases, they simply match what other visitors who viewed this page also looked at (note that this is at the “content” or “product” level not the HTML page level). One of the advantages of atomic approaches to content personalization is that they can be implemented easily on top of any system. Since you don’t need to know much about the content involved, you don’t need access to product databases or content categorizations. Categorical approaches rely on levels of meta-data about products both to identify potential suggestions and to categorize the output. Knowing, for example, that I just looked at a hose, a categorical approach might look at other garden supplies. One advantage to categorical approaches is that they allow for personalization options that aren’t necessarily strong but have considerable business value. Just as site structure often dominates causality, product type often dominates this type of market basket correlation. Without any knowledge of product, an atomic personalization engine will likely identify other TVs as the most associated product with any TV view. A categorical engine can isolate the top TVs, but it can also then search separate categories like DVD players or home theatre systems for the best matches in those categories. Another advantage to categorical approaches is that they function well in high-change environments like media portals where new content is constantly being circulated. Learning systems that work on a purely atomic level often don’t have enough behavior around new content to drive effective personalization suggestions.

Content-History Strategies

The content-based personalization just described works at the content level. It doesn’t matter which visitor views Story X, the engine will suggest Story Y. But intuitively, we know that who the visitor is should probably play some part in the personalized experience we provide. Content-History strategies can be thought of as forms of weighted average of content-based strategies. The idea is that you weight what the visitor just did alongside all their previous behavior to determine the next piece of content to suggest. With each new content view, the weightings will shift. Obviously, the weighting strategies can be simple or vastly complex – but the underlying principle is the same. Just like pure content-based strategies, content history strategies can work at the atomic or categorical level.

Next Best Offer (NBO) Strategies

NBO strategies can be redundant on the content and content-history strategies described above, but there are usually two new twists involved. First, NBO strategies tend to focus on a limited range of possible suggestions. Content and Content-History are deployed most commonly in very high content cardinality environments and are largely neutral about what to suggest. NBO techniques are usually focused on marketing drives where the key is to choose the best offer from a limited set of alternatives. What’s more, there may be a clear prioritization of the alternatives (A is more valuable than B if the visitor is equally likely to do either).  In addition, most offer strategies include some data about the visitor and their existing relationships. If, for instance, we want to decide between suggesting a new commission plan or paperless servicing to a brokerage customer, it’s important to know whether the visitor already has one or both of those options. These two differences turn out to be pretty significant in that NBO techniques are often fundamentally different that content or content-history approaches at the analytic level.

Trigger-based Strategies

Trigger-based strategies are one of the easiest types of personalization to implement. As a consequence, they are quite common. Trigger strategies work by identifying key milestones or strong behavioral cues and then attaching logical actions to those triggers. If, for example, someone views a page on the Website about big data, sending them a corresponding Whitepaper might be make sense. Trigger strategies MAY be analytically driven, but most often they are created based on intuitions about the logical next step or content association. Trigger-based strategies are baked into many marketing automation and CRM systems, so they are often thought of as being tied to those systems, but there is not necessity for such a tie. Many Website personalization strategies developed in tools like T&T or Optimizely are trigger-based strategies.  

Last Behavior Strategies

Last behavior strategies are a form of content-based personalization but I’ve called them out because they form a distinct and potentially interesting class of personalization strategy. The idea behind last behavior is that you direct the NEXT touch based on the behavior exhibited in the PREVIOUS touch. In other words, if someone went to the Website and looked at Product Y and then left, we’d target a message or offer around Product Y to them. This message might be delivered via email, display re-targeting, on the Homepage when they next visit, or embedded in a call-center script. Last Behavior strategies are often easier to implement because they don’t require real-time decisioning. As long as there is some gap between touches (and with channels like email you control that gap), then you have time to process behavior in a traditional fashion and generate personalization suggestions without the pressure of serving content sometime in the next 1.5 seconds.   

Threshold Strategies

One of the common challenges in digital is the amount of behavioral noise exhibited by visitors. Paths are diverse and often include pages, product or content that seem largely unrelated. Threshold strategies are similar to trigger strategies in that they attempt to define milestone points where an interest is clearly demonstrated and can, therefore, serve as the basis for personalization. The main difference is that they look for an amount of behavior rather than a particular point or event. Typical threshold strategies are things like a visitor viewing X pages of content, spending a high amount of time on a page, or visiting repeatedly. If you think about it, nearly all Paywall strategies are threshold personalization strategies. Threshold strategies are often a little more complicated to implement than trigger-based strategies. That’s partly because most Web analytics solutions don’t keep track of behavioral counters – so you don’t have readymade variables to drive the personalization. eVar counters are an underused weapon in the SiteCat arsenal if you’re interested in driving personalization via thresholds.

Time & Event Strategies

Some types of personalization involve time and external cues to work effectively. If you’re an online gift or flower buyer, for example, that behavior is nearly always centered on a holiday or a date (like a birthday or an anniversary). Where strong seasonal patterns drive behavior, traditional content weightings will deliver disastrous performance. What a leisure traveler did during ski season will have little relation to their interests in the summertime. Most business have some type of time & event personalization opportunity, and these can be the most impactful triggers you have. By isolating them, not only do you have a great set of powerful personalization options, but you clean up the rest of your signal so that your non-special event personalization will be more effective.

Offer Matching Strategies

Offer matching strategies are about matching up a visitor to the best possible offer from a large inventory. Listing NBO and Offer Matching as two separate types of strategies may seem redundant. There’s no doubt that NBO is, almost by definition, a form of offer matching. The reason I’ve chosen to split these two is that in actual practice, there tend to be two different situations – I’ve used NBO to cover the set of situations where the list of offers is relatively small. In most NBO situations, the list of offers is typically a choice between a small set of 4-12 static marketing drives. When the list of available offers is much larger, however, the type of techniques used in the analysis changes. It’s also significant when the offers themselves are dynamic. With a small number of static offers, you can use regression type analysis to create a fixed model that helps you understand each visitor’s proclivity for an offer. But when the number of offers is very large and changes frequently, those techniques won’t work. An additional complication comes when the offer matching model has to factor in limited inventory – as frequently happens in ecommerce, travel/hospitality, and couponing sites.

Filter-Based Strategies

Websites have become far more dynamic than they used to be. Sites are stuffed chock-full of interactive gadgets ranging from faceted search to DHTML popups, rotators, zooms, sorts, and filters. When it comes to gathering personalization cues, this is a very good thing. Filter-based strategies take advantage of the intelligence gleaned from visitor’s interactive behavior to understand their preferences and personalize. Does somebody always sort by price? Makes sense to surface up low (or high) priced items for their immediate purview. Do they click on the red color swath when they look at purses? Might make sense to show more red choices by default. When they facet, do they favor a specific brand? That’s valuable personalization. There’s hardly a click that doesn’t contain some interesting targeting material. Filter-based strategies capture those interactive preferences and then build into subsequent views – even views which are fundamentally different in style (you can take price preferences expressed in faceted search and use them to personalize an email offer). Filter-based strategies are easy to execute but are under-utilized, at least in part because the vast majority of filter-data is never captured in Digital Analytics tools.


The idea behind crowd-sourcing is simple – let other users of content guide visitors to the best, most interesting content (or products). Many media sites use variations of the “most-popular content” technique; it’s easy to do and it nearly always drives additional content engagement. Ecommerce sites often use the popularity of products to guide search order within pages. Most crowd-sourcing techniques are analytically very simple, being nothing more than counting or summing functions. But though simple, they are, nevertheless, an effective means of identifying potentially interesting content. That being said, this type of crowd-sourcing isn’t particularly personal. To get to true personalization with crowd-sourcing, you need the next technique.

People-matching and Nearest Audience Strategies

Targeted marketers and conversion analysts have been using people-matching strategies forever. You find the group of people most nearly like your best shoppers (or readers) and you target them. That’s people-matching. Not surprisingly, people-matching has also proven to be a very powerful personalization tool. By identifying similar users based on previous behavior or profile, it’s often possible to find content or products that are highly personalized to the target visitor. People-matching strategies work well across a wide range of problems and are effective both as background personalization tools (strategies you use to personalize someone’s experience) and as user-controlled personalization mechanisms (friend-based playlists for example).

VoC Based Strategies

If you want to know what someone wants, why not ask? That’s the idea behind VoC based strategies that morph Voice of Customer from a research tool to a targeting tool. Sites with higher engagement may have unique opportunities to garner deep user interest information through the simple mechanism of asking. While few enterprises have seen or taken advantage of the potential for VoC-based personalization, profile building isn’t that uncommon. We also got a sterling example of the power of VoC strategies to drive personalization in the 2012 Obama campaign. Massive field work took advantage of integrated VoC to collect data not simply to drive polling but to tailor messaging back to individuals and micro-regions. What works for campaigns can work for business too. And while it the massive volunteer field forces of a presidential campaign may be outside the realm of the possible for an enterprise, digital provides a low-cost, technologically enabled platform for similar kinds of low-impact collection and personalization.

User-Directed Strategies

Sometimes, you can let your users do the work when it comes to personalization. Go to most portals and you’ll find a significant amount of user configurability. This is personalization driven by user selection. It’s a mechanism that can be challenging technically and that requires a committed and regular audience, but when done well it delivers an undeniably personal experience. One place you don’t see much use of user-directed strategies is mobile apps – and that’s a shame because the limited screen space in mobile makes it a natural platform for user-directed personalization.

It’s a long list and, like a baker’s dozen of donuts, perhaps too much to swallow in one sitting. As is often the case, the very fact of so many different choices creates its own kind of complexity. Not every method is appropriate for every site or situation. Mapping your business needs to the appropriate strategies is critical to success and will be a big part of any good personalization strategy. But that’s a topic for another day…

Republished with author's permission from original post.

Gary Angel
Gary is the CEO of Digital Mortar. DM is the leading platform for in-store customer journey analytics. It provides near real-time reporting and analysis of how stores performed including full in-store funnel analysis, segmented customer journey analysis, staff evaluation and optimization, and compliance reporting. Prior to founding Digital Mortar, Gary led Ernst & Young's Digital Analytics practice. His previous company, Semphonic, was acquired by EY in 2013.


Please use comments to add value to the discussion. Maximum one link to an educational blog post or article. We will NOT PUBLISH brief comments like "good post," comments that mainly promote links, or comments with links to companies, products, or services.

Please enter your comment!
Please enter your name here