Learning From Every Visit: Behavioural Optimisation for Online Stores

Search helps a customer who knows what they want. Behavioural optimisation is for everyone else. It has been in online stores since 2016, built by PureNet, our sister company in the PureNet Group, and in 2016 it was one reason PureNet won two awards in one night: Business Innovation of the Year and Technology Business of the Year at the York Press Business Awards.

The Challenge

Most visitors to an online store don’t search. They browse, compare, leave and come back, and every click says something about what they want. Most stores ignore that, and show the same products to every visitor.

Our Approach

The team built behavioural optimisation that learns from what visitors do. It records what each visitor looks at, how they move through the store and what they buy, then uses that to choose which products, ranges and offers to show next.

From the start it reported its results in the client’s terms: the revenue it brought in, its return and average order value. It did not report how many recommendations it had shown. The client saw the outcome on the same screen as the tool. CONVERGE, chapter 8, makes the same rule part of the method.

It ran on personal data, so it ran inside the group’s risk management, which is certified to ISO 27001 for information security and ISO 9001 for quality.

What We Delivered

• TW Wholesale: a BigCommerce store, built by Agency51 in eight weeks over Christmas, with AI merchandising that learns from how customers browse and buy across a very large range of tools and building supplies.
• Duffells: a B2B store for replacement door and window parts, joined to its Merlin ERP, with merchandising and search based on each customer’s segment and activity, and cross-selling to lift order values.
• Pepper Pot Studios: a Shopify store for illustrated map prints, with behavioural optimisation that shows what draws customers to a print and what leads them to order.

The Research

In 2021 the research team set out to predict whether a visitor would buy, at each step of their visit:

• 2021: the first model was a Markov chain, which predicts the next step from the current one only. It couldn’t use the visitor’s history, and a shopper’s next click depends on everything they have done before.
• 2022: the second project used recurrent neural networks, which carry memory forward, trained on years of visitor behaviour across several clients’ stores. It worked better.

Most of the work was getting the data into a shape the models could learn from. CONVERGE, chapter 5, tells the story, because the lesson applies well beyond ecommerce: data first, then the clever part.

The Outcome

• The stores: they kept paying for it because it worked.
• TW Wholesale: the store went live inside its eight-week window, and AI merchandising shows customers the products they’re most likely to want. Michael Tichauer, Systems and Operations Manager, said: “Our BigCommerce ecommerce site developed by Agency51 has provided a first-class platform to keep delivering this commitment online.”
• Pepper Pot Studios: Anna Lawtin, MD, said: “They understood my business inside out almost straight away and provided approaches and solutions to maximise my revenue.”

Why It Matters Beyond Ecommerce

Every business has behaviour data: what customers order, when they stop, which quotes turn into work. That record is often the best AI asset a business already has, and it’s usually in no shape to use. Our data mining and requirements and forecasting work starts there.