Personalisation, said Mottershead, is “delivering tailored, relevant communications to the right user at the right time”.
The benefits of personalisation include improving KPI metrics and improving the brand experience for consumers. Customers want to be inspired, said Mottershead, and shown relevant content that they might not have thought of themselves.
This makes personalisation an invaluable tool for retailers, who can use it to introduce customers to relevant products from across an online store that they otherwise might not have encountered.
Nick Mottershead illustrates the importance of personalisation with a statistic from Econsultancy’s Conversion Rate Optimization Report
Enter recommendation engines: a type of information filtering system that uses machine learning algorithms to provide the most relevant items, or content, to a user. Recommendation engines can incorporate a variety of data, including user purchasing behaviour, user browsing behaviour, user demographics, and real-time triggers.
How does this work in practice? Mottershead used a case study from a large online retailer, with whom Lynchpin Analytics had worked to implement a recommendation engine, to illustrate how recommendation engines can benefit retailers, lift conversion rates and improve revenue.
He walked the audience through the process from start to finish, explaining how Lynchpin developed and executed its recommendation engine, what it learned from the process, and how it plans to improve on the system in future.
The Power of Personalisation: Simple Steps to Provide a Unique Customer Experience
A retail personalisation case study
Who, what, why?
Lynchpin’s client was a large online retailer that wanted to improve conversion rates and average order value by offering more relevant and personalised product listings on its website.
In data terms, said Mottershead, Lynchpin wanted to analyse user behaviour to determine the products that a consumer was most likely to purchase, and use those to provide recommendations.
Data collection & understanding
Initial data collection involved accessing historic transactional data across all markets, then using a cross-sell tool to uncover insights.
The tool analysed crossovers at a category and SKU (stock-keeping unit) level, determining the product categories and product codes purchased by each consumer in order to draw associations between them.
Market Basket Analysis & product relations
Mottershead’s team discovered an unexpected quality issue with their data in that the User ID allocated to customers who checked out as guests was not consistent over time, making it difficult to link repeat purchases.
In order to counteract this, the team used a type of modelling called Market Basket Analysis, which looks for combinations of items that occur together frequently in transactions.
These were then used to generate a set of “if this then that” association rules:
- Customers who bought this, also bought that
- Users who listened to this also listened to that
- Subscribers who read this also read that… and so on.
From this data, the team created a Network of Product Relations which visually mapped out the relationships between products that customers frequently bought together.
Mottershead explained that these could be used by the training team to understand how to inspire customers by showing products that were related to one another – products which might otherwise have a complicated path between them due to website structure.
Results & evaluation
How did these recommendations appear on the site? Users with a product in their basket, or with a purchase history on the site, would be presented with a banner of recommended items on the basket page and homepage.
All product pages were also updated with banners of items that users were most likely to purchase based on the product being viewed – much like Amazon’s “Customers who viewed this item also viewed…” or “What do customers buy after viewing this item?” recommendation banners.
Mottershead described the algorithms as optimisation that you could “sit and place once”, and then leave them to their own devices. There was no need to update the engine in real-time – weekly or monthly updates were enough to keep the recommendations fresh and current.
He noted that there were some products customers would buy over and over, leading to Lynchpin applying some specific business rules based on the client and the type of products they sell.
The purchasing rules for the recommendation engine were output as a simple .csv file. This format meant that the information could be shared across the business, giving teams from different parts of the company information that they could use – instead of the insights being a “black box that sits on the website”, as Mottershead put it.
What were the results of the case study? The pilot of Lynchpin’s UK engine achieved a 32% conversion rate uplift, as well as a 23% uplift in revenue.
Mottershead’s team are still perfecting the engine, working to develop more sophisticated algorithms that make use of information like time and on-site behaviour. At the moment, he said, the personalisation is very “one-to-many” – the ideal, eventually, is one-to-one personalisation.
Five tips for personalisation
Mottershead concluded his talk with five tips for carrying out effective personalisation:
1. Take action on the data you have today – even if it’s not 360 degrees
Mottershead emphasised at multiple points that you don’t need a complete, 360-degree customer view in order to use your data for personalisation.
In fact, he stated that having a more limited scope – and being able to action that right away – is more effective than having vast amounts of data.
2. Get access to data you have across the organisation
Data is very often siloed and spread across various business units: CRM data, transactional data, web interaction data – the list goes on. Unifying these, said Mottershead, is the key to a powerful personalisation project.
3. A more advanced model is not necessarily a better model
Complexity is not necessarily better. Understanding what’s being fed into the model, what the output means and why can be more important than the output itself.
4. Determine how to measure success
What are the key KPIs that your personalisation should be affecting?
Keeping it simple and quantifiable, advised Mottershead, is probably the best approach: for example, “I want to increase conversion rate for first-time buyers”. This kind of goal is shareable, and very simple to A/B test.
5. Adopt a “test and learn” mindset
Speaking of testing: personalisation should be an evolving process, not a one-time solution provided by one model.
Mottershead advised the audience to start simple and slowly work their way up as they become more confident and start to better understand their customer and their data.
6. You actually want several models, because the amount of data available varies. For example when a completely new shopper arrives on site, you know very little about them, but you still need to personalise the site – so your first model is based on what little you can deduce from things like the time of day and their IP address. Then after a few minutes, you know what products they have been looking at. So you can activate a second model based on the network of product relations, as in the picture above. And so on: as you know more data, you can activate more accurate models.