A large retailer with a significant high-street and online presence in the UK.
Performance marketing is an important driver of new customers and sales to this retailer’s business, with an annual budget for online ‘Pay Per Click’ alone of £11m. This marketing tool helps to direct potential customers who are looking for a particular product on a search engine to the retailer’s landing page, via prioritised advertising. Typically a retailer with successful landing page optimisations could expect an average increase in conversion of 15–20%.
The retailer in question was not using this marketing tool effectively. Often the landing pages presented were not specific enough, with customers being directed to a generic product rather than one similar to their specific search. The impact was that the company was spending a lot on these advertisements without obtaining the desired return on investment.
Faculty was charged with building a model that could map search terms entered by customers, onto specific landing pages on the retailer’s website. In order to do this effectively, we approached the problem in two ways.
First, we built a supervised learning algorithm that was able to manually match a given search term to a URL in the retailer’s website. This was done by using a large volume of historical data (where the retailer had manually linked search terms to landing pages), and then matching the retailer’s numeric product code to the specific search term.
This approach worked well for high-level categories such as gender and product, but for the more complex search terms this was more challenging. This was because of the presence of a larger number of categories, with fewer training examples available (e.g. brand and colour). We addressed this by applying ‘fuzzy matching’ to look for approximate correspondences between the search terms and the retailers’ product code.
Misspelled or ambiguous search terms can be mapped to the right URL using fuzzy matching, to optimise customer traffic.
‘Fuzzy matching’ was effective at capturing specific information such as brand and colour, while the supervised classifiers were more effective at capturing basic categories such as gender and product. Together the two approaches were able to effectively match search terms to complex landing page URLs, increasing the likelihood of conversion.
The output of the model was an application that allowed the retailer to upload a list of search terms, returning URLs for those search terms immediately. This had a huge benefit for the retailer through the creation of better landing pages, allowing it to tag landing pages based on past data and to redeploy resources for more cognitively difficult tasks.