Smart Merchandising – Inspired by Artificial Intelligence

“If artificial intelligence has its way, discounting could disappear, thanks to software that tells retailers exactly what and how many products to buy, and when to put them on sale to sell them at full price. Online shopping could become a conversation, where the shopper describes the dress of their dreams, and, in seconds, an AI-powered search engine tracks down the closest match. Designers, merchandisers and buyers could all work alongside AI, to predict what customers want to wear, before they even know themselves.” - Business of Fashion

What I like about this vision is that it characterizes three interesting aspects of information access and search – namely, that the software will more accurately predict demand, that “search is a conversation” – a premise that we have described over the years during our typical search strategies and initiatives –  and that merchandizers will work alongside AI,  rather than being superseded by it.

New call-to-action

Design is About Prediction, AI Requires Creativity

The idea of predicting what customers want before they know it is actually what the fashion industry does today.  Designers don’t ask customers what they want; they create apparel they believe customers will want once they see it. The more astute the understanding of the customer and their needs, the more accurate the prediction.  In fact, any good sales person, consultant, or stylist gets to know the wants and needs of their customers and does their job by responding to needs that are clearly defined as well as those that are not well defined – or even well understood – by the customer. This applies in any industry B2C or B2B.

Last evening I watched the movie “Joy,” which stars Bradley Cooper and Jennifer Lawrence.  In this story, based on the real life of entrepreneur Joy Mangano, the title character invents a product that meets the latent needs of consumers.  That inherently creative human process of understanding and predicting needs will still be the job of people.   

Despite today’s crop of AI-driven applications, creativity will still be the job of designers, marketers, merchandizers, product managers or, of course inventors.  An AI-driven merchandising system may assemble bundles of products into solutions or recommend particular styles and choices given one’s wants and latent needs, but the human part of this is coming up with those choices and understanding the factors that drive preferences.  What then, is AI-driven merchandising or, as some are calling it, Smart Merchandising?

AI Applied to Customer Experience

One merchandising concept that leverages intelligent technology is to assemble a series of components into a user experience dynamically when a response (“success”) leads to the survival of certain traits and selection of the next generation of traits.  An experiment in the UK, for example, measured the response of passersby to a poster that was composed of dynamically selected components of pictures and copy. The combination eliciting the best response evolved into the next generation of the poster.  (Response was measured by detecting whether people were looking at the ad using a Kinect device that “read” body language.)  

Segmentation and Hidden Attributes, but No Magic

These variations cannot be managed without automation; algorithms use continual iterative A/B testing to pick “winners” for audience segments.  The segments might be based on latent characteristics – that is, they may not be obvious or easily described.  Perhaps a segment is based on nuances of demographics combined with recent purchases.  Dozens of factors may be hidden in the data set.  Humans don’t need to make the specific determination of those factors – unsupervised machine learning algorithms make the determination and create the segments.  It sounds like magic, but there are limitations and constraints. 

  1. Large data sets and large numbers of transactions are required.  The algorithm is essentially running experiments with customers.  There needs to be a control group against which results can be compared. A great number of customers and purchases are needed for the algorithms to operate.  This process will not work with one hundred or even one thousand transactions.
  2. The components need to be defined.  For an algorithm to work in an experiment, it needs the variables with which to run the experiment.  This means an agency, marketer or merchandizer needs to define the variables around offerings that can be recombined, which is not a trivial task.
  3. Good quality data is required.  Customer data needs to be integrated across platforms and systems, and a meaningful number of elements that describe the customer (such as demographics, purchase history and behavioral data). Most organizations are working with platforms that require integration, so data needs to be cleansed and mapped across systems.  

Ecommerce = Recommendation Engine  

At its core, most ecommerce platforms boil down to a simple concept – that of a recommendation engine.  The recommendation may be explicit, as in “you want products in this price range, style, color and type – here they are,” and this retrieval of products is based on straightforward product attributes.  The list may be long or short, depending on the depth of the catalog, and the filter; choices may be extensive or limited, depending on the detail and quality of the product data and structure. 

Language plays an important role – customers describe what they want using varying terminology, and the overall look, feel, layout and structure of the site may either facilitate or limit customer wayfinding.  In fact, beyond price and excluding house brands or exclusive merchandise, the curation, organization, and selection that a brand provides is the key differentiator in many situations. 

Target and Walmart have overlapping product assortments and similar prices.  The character of the web sites, brand feel, structure, navigation, and merchandising attributes are what gives the shopper a unique experience.  On one level this may seem obvious, but on another level, these organizations are largely competing on their data and data structures. That is certainly an over-simplification, and does not negate all the efforts of the hard working product managers and merchandizers, but at the end of the day, the online customer experience is data driven. 

Mental Models and Data Models

But even when data driven merchandising is precisely modeled to meet the needs of users, different terminology, different navigational choices and different product collections will appeal to different users at different times in their journey.  There is no “one size fits all,” because our mental models – the ways we look at the world – are different and depend on a variety of factors. 

These not only vary across individuals, but may also vary depending on factors such as mood, location, and what we are doing at the time or are trying to do.  Consider the experience you want when you are in a hurry or under stress versus when you have a leisurely morning over a cup of coffee to browse and wander through a web site. In the first case, you want to get the task completed quickly, in the second you want to learn, take your time, and enjoy the experience.

Nuanced Signals Trigger Varying Experiences

These nuanced experiences are not only difficult to construct; it is also difficult to interpret the signals that a shopper provides in order move the process in the right direction. A salesperson in a store can read body language and facial expression and adjust how they interact.  Perhaps one day a web site will be able to detect these subtle signals and vary the experience. But even considering the less nuanced challenge of presenting products according to task and objective, the process becomes very difficult without mechanisms that can automatically group products according to need and problem. 

Product Information at the Core of the Ecommerce Experience

Going back to today’s web site differentiators, product information is at the core of every commerce experience.  Not having high quality, high fidelity data means that even basic selections and filtering according to user preferences will be difficult to manage and automate.  If the data is not there, then the ecommerce teams are left to manually assemble landing pages that cater to user scenarios and tasks.  This process becomes slow, costly, and generally unmanageable over time, in addition to not adequately addressing the user’s needs.

Mapping Solution Families Based on Attributes

But products can be dynamically grouped by mapping attributes into families of solutions and even using mined content to provide the contextual signals that an algorithm can interpret and use to group products.  This is where human knowledge needs to be collected – whether through expert interviews or by processing explicit knowledge contained in articles, manuals, FAQ’s, support materials, and other artifacts.  An article about building a deck will contain lists of tools and materials.  That article might be part of curated collection of enriched product data that can then be mined to understand product relationships.

 A certain amount of this work is manual – the development of product relationships, solution classes, problem types, content types, and the relationships between concepts – such as products that support a solution. These structures and relationships map the knowledge domain – the ontology.  Over time, these relationships can be catalogued and curated to drive AI engines that interpret content and associate that content with products and solutions.  Those solutions are then used to create dynamic solution pages.  These approaches are already in use today at a small number of highly progressive, leading edge companies.

AI engines will continue to evolve and become “super recommendation engines” that will put ecommerce sites on steroids.  The organizations that invest wisely in these applications of AI will understand and serve their customers more effectively, gain market share and become formidable competitors in their space.

Seth Earley

Seth Earley is the Founder & CEO of Earley Information Science and the author of the award winning book The AI-Powered Enterprise: Harness the Power of Ontologies to Make Your Business Smarter, Faster, and More Profitable. An expert with 20+ years experience in Knowledge Strategy, Data and Information Architecture, Search-based Applications and Information Findability solutions. He has worked with a diverse roster of Fortune 1000 companies helping them to achieve higher levels of operating performance.