Amazon has been using algorithms to try to sell you extra stuff for years.
But the technology to personalise merchandising, much further than recommendations, is advancing rapidly across ecommerce.
I caught up with Sören Meelby, VP Marketing at Apptus, to get an introduction to the technology (Apptus eSales), and to pose some questions about the user experience in online retail.
Econsultancy: Many ecommerce sites allow the user to filter by ‘most popular’? How much further does automated merchandising go?
Sören Meelby: Each and every sort order typically follows a logic or business rules and ‘most popular’ is fairly straight forward. It is the most popular products over a given or configurable time period.
Using ‘most popular’ as an example, it has several aspects that can be controlled and therefore optimised and automated, such as:
- Time period to decide what is most popular.
- Metric to determine what is popular = is it # sold units, # of views, # of sales to unique customers.
- Type of filter applied to determine product set.
- Visitors’ historic behaviour on the site in general and in ‘most popular’ as a specific type of list.
- What type of product attributes to look at when building the list of most popular.
Each parameter will affect the output of the list and thus the type of products that will be exposed.
If a long time period is used as input then the list of products shown in ‘most popular’ may seldom change and the retailer will sell the same small set of products over and over at least from the real estate where ‘most popular’ is in play.
If a retailer wants to get specific results (typically uplift of a KPI) from having an area of their site showing ‘most popular’ they will track the performance and adjust (if possible) the parameters (e.g. 1-5, above) and measure the effect and then iterate until satisfied.
If humans need to be involved in these multi-step optimisation iterations the process will be slow and error prone (especially if the goal is to optimise an entire site where a multitude of different areas should play in concert towards an overarching business goal).
With an automated merchandising system the optimisation process described above is automated by means of using algorithms (driven by AI and machine learning principles) to continuously adjust the parameters 1-5.
With an automated system no human input is strictly needed to get the optimisation process to happen but in our case a human can enrich the automated system by telling it what goal to optimise towards, typically a specific business KPI (conversion, revenue, profit).
One example of Apptus’s solution
E: If facets and sorts are adjusted based on user behaviours, doesn’t this remove important predictability from the user experience?
SM: First, a brief history lesson:
- Lists were originally in alphabetical order. This quickly becomes cumbersome to use.
- Then came using result counts in combination with a category structure. This quickly becomes cumbersome.
- Then came exposing more detailed attributes from the result set, such as color, size, price range, language etc. – and this is, at present, the industry standard for how to facilitate help for a user to find what they are looking for when starting out at a broad product set.
Our system offers the industry standard described in C above but with an important twist – we make a relevant selection of what facets and filters to expose to the user.
This does give up a tiny bit of predictability but you make significant gains in usability and usefulness for the user experience.
Firstly, we look not merely at the counts of results to determine how important they are, we look at the aggregate sales performance for the underlying product set in each filter or facet.
Secondly, we apply machine learning and AI to select the most relevant facets and filters and what order to present them in (implied here is that order has an impact on the performance, which we have tested).
So the end result is a more useful list of filters and facets that doesn’t add to the cognitive load for the user when they consume the UI.
E: How do you deal with customers who browse on mobile and shop on desktop?
SM: We follow the user, cross-channel. Anonymous users still get a record in our system.
Anonymous users on multiple devices can to a certain degree successfully be merged into one, but we have a dependency on the site ‘owner’ to facilitate this in their cookie handling. A user is best identified when he/she makes a purchase or signs in – then we get the entire landscape of sessions merged into one and our ability to perform increases.
To summarize – as long as the site owner is not preventing anything we can deal gracefully with multi-device interactions.
E: What’s next in merchandising?
SM: One important mission for us is to marry automation with control in our system, to offer merchandisers the freedom to act without being dependent on IT staff or data scientists.
We see some of our richest clients being completely locked down due to systematic failure of platforms and systems to offer merchandisers any freedom to act.
Also, despite the best intentions, the end customer experience simply does not delight customers in all cases – the proof is in the metrics. At Apptus, we believe that reversing this trend requires a radical new approach: an approach that is already proven in solving the same problems in other markets.
Predictive machine-learning is being used, successfully, in financial markets, and marketing automation is moving over to using AI. These are trends driven by the same issues facing retailers – masses of information and not enough people or time to act intelligently on it instantly.
With computers becoming the dominant force of retail, we believe that in five years’ time over 90% of the virtual shopping experience will be automated by computers with AI-powered ecommerce optimisation.