The use of algorithms in ecommerce is tied up to a certain extent with two similar yet separate concepts, Artificial Intelligence (AI) and Machine Learning (ML). The two terms have been used a lot in recent years, so I think it’s useful to distinguish between them.
First coined in 1956 by John McCarthy, AI involves machines that can perform tasks that are characteristic of human intelligence. Machine learning is a subcategory of AI. Artificial intelligence is the idea of using technology that behaves like a human, while machine learning algorithms are about finding patterns within data.
These self-learning algorithms enable the machines to learn from data sets and can have various applications for ecommerce, which can help customers and retailers alike.
The technology behind these algorithms may seem the domain of the tech giants like Google and Facebook, but there are already lots of applications for retailers large and small, and machine learning algorithms look set to become more influential in retail and marketing in years to come.
For smaller businesses, much of this technology is accessible right now, and many are already using it. For example, there are plenty of third party recommendation engines which are easily affordable for SMEs.
Algorithms in ecommerce are used in several ways:
- To increase understanding of the customer / target customer.
- Segmentation, learning from user behaviour and on-site data.
- Predicting customer preferences. Recommendation engines, which seek to show relevant products at the right time are one example of this.
- Predicting customer intent. Using data such as previous browsing and add to cart behavior, algorithms can predict customer intent, when using site search for example.
In this article, we’ll look at some examples how AI and machine learning algorithms are being used in ecommerce.
Product recommendation engines are now a common feature on ecommerce sites. Using algorithms, they are used to surface products for customers based on various factors.
They benefit retailers by showing visitors the products that, based on the data fed into the engine, they’re most likely to buy. For the shopper, the improved relevance means a better shopping experience.
These engines work well too. Indeed, Netflix says its recommendation engine is worth $1 billion a year to the company. This is because it helps users to find a show or movie they want to watch, and quickly, before a customer decides to give up the search.
The company’s research found that Netflix users lose interest after 60 to 90 seconds of choosing. The algorithms that learn from their behaviour and preferences can surface relevant content more quickly than if they had to search and browse, and keeps them subscribing.
They can be used in various ways. For example, to show trending or best-selling items, or those which visitors are viewing more than others.
Product recommendations can be more advanced, finding relationships between products in different ways.
Some estimates credit recommendation engines for 35% of Amazon’s sales, and this seems very plausible given the number of recommendations Amazon makes, and the different ways it uses data to present products.
A range of data is used to produce these recommendations, from personal shopping history, to the purchase journeys of other site users. In this way, recommendation engines can produce recommendations accurately and at scale, something which would be impossible to do manually.
Site search is a key yet often overlooked feature in online retail. However, when implemented well, it can improve the user experience, and ensure that customers can quickly find the products they’re looking for.
Traditional site search is relatively basic, and relies on finding an item in the product database which matches all or some of the shoppers’ search query. This can produce good results for the user, but it can fail, and often also relies on the shopper spending time tweaking queries and filtering and sorting returned results.
Machine learning algorithms can be applied here, which allow additional data such as add to cart and purchase behavior around products to influence the sorting of results, meaning shoppers will see more relevant results, and those which other shoppers have a higher propensity to purchase.
Often referred to as dynamic pricing, algorithms can be used to control and set pricing levels for online retailers.
For the retailer, it can help them find the right price point for their products and maximise profitability. It can take into account several variables to get, testing for different visitors, before finding the best blend.
For example, product seasonality, competitor prices, product margins and factors like stock levels can be considered by the algorithm.
Dynamic pricing can have negative connotations, when associated with Uber’s surge pricing for example. However, it can actually deliver better deals and more competitive pricing to the individual shopper in some cases.
Machine learning algorithms can also forecast inventory levels retailers will need, using demand estimation engines.
By using various factors to forecast demand, retailers are able to order the right amount of stock at the right time to cope with differing levels of traffic.
Chatbots are designed to simulate online conversations with users, and have the potential to reduce pressure on customer service teams and speed up live chat interactions.
I think it’ll take time before they can replicate human performance, and also before the majority of people are confident interacting in this way, but some retailers are already trying out this technology in other ways.
For example, eBay’s Shopbot is a personal shopping assistant which aims to help customers find the right product for them.
Another example comes from North Face, which uses IBM Watson to help shoppers find the best jacket for them. It asks customers questions about when and where they’ll be using the jacket, and what for, preferred styles, and crunches this data to find the best match.
Email subject line optimisation
Email remains a massively important marketing channel for retailers, and subject lines are a key aspect of this.
A good subject line can make the difference between the recipient opening or not opening an email, and so it pays to test and find the best performing copy.
Through A/B testing and a little creativity, marketers can improve performance, but companies like Phrase are using AI algorithms to beat the results achieved by humans.
Language generation algorithms can use a brand’s language to produce subject lines specific to them and relevant to customers, testing variations to find the best performing subject lines.
Personalisation can be used in various ways on ecommerce sites and in marketing emails, with the aim of improving relevance for the shopper.
When used well, personalisation improves the experience for the shopper, as communications, product recommendations and even whole pages on the site are tailored to them according to their preferences.
It also helps the retailer, as an easier shopping process and better experience means customers are more likely to buy.
Personalisation, as it has been implemented so far, has often been basic. Using a customer’s name in emails, recommend products based on those a customer has bought before. These are tactics that can work, but AI algorithms offer the opportunity to personalise on a different level.
For example, Asda sends a Daily Alert email to each of its subscribers in the 10 days leading up to the end of each month (timed around most peoples’ pay days). These contain personalised product and content recommendations.
The emails are simple enough in design and content, but the process and algorithms are very impressive. Each email is unique, with the offers and products personalised for each user according to their propensity to purchase.
Segmentation has been used for years in ecommerce in various ways, so that marketers can sell to different customer groups in different ways.
Common segmentation methods include:
- Customer demographics – data such as age, gender, income and occupation can be used to target different customer groups in different ways.
- Customer lifetime value. It’s a good tactic to target the most valuable and most growable segments amongst your customers.
- Customer behavior – previous purchase history, browsing behavior on site etc.
- Devices and channels used. Customers can be grouped according to the devices they use to access your site, or the preferred channels.
AI powered algorithms have the potential to analyse the various data sources, making connections that traditional segmentation methods may miss.
For example, a learning algorithm might identify alternative means of segmentation, such as online behaviour or preferences, that can serve as a more accurate predictor of interests or tastes than standard methods.
Machine learning algorithms can help with image recognition, identifying images and finding best matches for those uploaded by customers.
This has a lot of exciting uses for ecommerce sites, allowing for applications which improve shopping experiences or new product discovery. For example, image recognition can be used in site search to find visual matches for products entered and show relevant results for users.
It has been around for several years, but the technology has improved a great deal, and is now used effectively by several retailers. ASOS is an excellent example, with the Style Match feature on its mobile app.
Style Match allows users to use a photo from their phone’s media library to search for visually similar items. The app then returns suggested results that they can buy on ASOS.
It works well for mobile users, and makes perfect sense for ASOS, which pulls in 80% of its UK traffic and 70% of sales from mobile. For retailers, image recognition helps customers to find the item they want to buy, while it means a better and easier product search experience for shoppers.