How can we predict product success in this multi-channel and multi-category business, with a fully changing product assortment 52 times a year, within a strict one-way go-to-market process, 40 weeks before launch for thousands of different products? How can we precisely identify the flop products in very heterogeneous assortments? A forecasting system supports Tchibo in the evaluation of assortments. With the help of Artificial Intelligence (AI) it makes predictions about flop probability in order to better align the product range with customer wishes.
"Founded in 1949 by Max Herz, Tchibo has been synonymous with freshness and quality in the coffee market for over 65 years. Originally a coffee mail-order firm, it has evolved into an international company and operates in many more business sectors than the traditional selling of coffee. Over the years, Tchibo has systematically expanded its range and distribution paths, ensuring diversity and quality the world over with strong brands."
Managing this weekly changing assortment with 50-60 curated products across three sales channels and many markets is a high entry barrier, as it requires mastering a high degree of complexity from product development up to logistics. As assortments are (time-)limited and curated, they have to meet consumer need spot-on. Product flops have to be avoided by all means, as they create significant friction and transactional costs in this fine-tuned, high-speed business model. Imagine testing products before placing them in our sales channels in such a way that we could precisely predict whether the product would be a flop or not – channel by channel. Imagine gathering consumer feedback so accurately that we could understand exactly what consumers like and dislike about our intended product. Finally, imagine evaluating consumer feedback in a scalable and automated way to generate high pre-testing ROI.
Alexander Falser (Head of Consumer Insights, Tchibo)
To leverage the power of semantic analytics and create a powerful predictive algorithm for product success, Tchibo and Cauliflower have developed a four-step solution.
1. A baseline of the success potential of each product is created by analysing vast amounts of historical product data. From this analysis, general or overarching product characteristics can be derived that contribute to sales success or failure.
2. A set of quantitative survey questions defines, such as purchase intention and open-ended questions about the impression of the respondents of individual products. Without any predefined structures – such as scales or survey items to follow – participants can subvert our expectations of how products are evaluated. Instead of solely relying on the usual closed survey questions, new aspects that may be particularly liked or disliked about a certain product emerge from the unfiltered feedback of survey participants.
3. Natural Language Processing and Deep Learning are used to leverage open-ended survey questions for prediction as well as sales figures and closed survey questions. In this way, all relevant aspects are extracted from the unstructured feedback and prepared for the prediction model.
4. A self-learning neural network dynamically recognises the patterns and relationships between the various data and can use new survey data to make predictions about the flop probability of products.
Using semantic analytics to understand the mental concept behind a consumer’s spontaneous reactions, even to early sketches or product prototypes, turned out to significantly boost product pre-testing accuracy. As a result, Tchibo today identifies 80% of product flops early and cost-effectively.
The team with Alexander Falser and Marco Walter from Tchibo and Gianluca-Daniele Speranza and Lukas Waidelich from Cauliflower has written a paper about the innovative approach and submitted it to the ESOMAR Best Paper Award 2020. With their submission, the team reached the final round alongside technology giants like Microsoft, Facebook and Google and presented the approach at the ESOMAR Insights Festival 2020 on the main stage. The presentation was reported in ResearchWorld and planung&analyse. The team also presented the approach at Pforzheim University.
Schedule a demo with a consultant and learn how to start analyzing open-ended responses.Schedule Call with consultant