download pdf

Reading “between the lines” of product reviews on a large scale
When people consume in the digital space, they not only click and buy, but often comment on products, brands and services on social media, on platforms or the sites of online stores. The enormous amount of textual data consumers produce online is in fact a treasure box that hasn’t yet been fully opened. But as data volumes grow, so do new algorithms to process and analyze unstructured data. Artificial Intelligence (AI) is one of the domains that can help open this treasure box further to better understand consumer decision-making.
In a GfK research project, we tested how we can learn consumer preferences and predict choices from publicly available social media and review data which are related to sales data. The common AI tool “Word Embeddings” has shown to be a powerful way to analyze the words that people use. It enabled us to reveal consumers’ preferred brands, favorite features and main benefits. Language biases uncovered by the analysis can indicate preferences, and they fit reasonably well to actual brand sales within various categories. Especially when data volumes were large, the method produced very accurate results and it is completely passive (see Box 1). We have been using free, widespread online data without affecting respondents or leading them into ranking or answering questions they would otherwise not have even thought of. The analysis is fast to run, and no fancy processing power is needed.

Predicting the most preferred brands in one category
To test if brand preferences could be derived from online reviews, we first ran the AI-based text analysis for one category (TVs) and different amounts of data and compared the outcome of the analysis to actual sales data.
Specifically, we ran 3 experiments: using data from one online retailer only, encompassing a total of 3,000 reviews; using data from multiple retailers totalling 4,500 reviews (a random subsample of the whole data); and using the entire data set of 53,000 reviews.

The results are displayed in Figure 1. The first column shows the sales ranks of 5 brands in the category. It is important to note that the sales difference between brands C, D and E was quite small, and therefore we had expected some confusion. The 2nd column shows the results of 3,000 reviews scraped from a single online retailer. With this limited amount of data, the rank is clearly wrong, ranking the most-sold brands A and B on ranks 3 and 4 instead of 1 and 2. The 3rd column introduces multiple retailers with a random subsample of 4,500 reviews. In this experiment, Brand A is now in the correct position (1), but we see confusion on Brand B and the others. The 4th column, using the complete dataset of 53,000 reviews, shows the correct ranking for Brand A and B - the major volume drivers in the category - and confusion of Brands C, D, E.

Predicting brand preferences in multiple categories
Next, we ran experiments across 10 categories, again using point-of-sale data as the reference for the ranking. As in the first experiment, more data lead to more accurate results. The categories we tested were Tablets, Hand-held Vacuums, Full-size Vacuums, Smartphones, Hair Dryers, Fridges, Laptops, TVs, Shavers, Dishwashers and Washing Machines. We see that, to get correct brands in the top positions in at least 80% of all cases, a sample of 5,000+ is required, and that accuracy can be enhanced significantly with larger amounts of data (Figure 2). In the Smartphone category (using Twitter data), we made the interesting observation that the correct order prevails bar Alcatel. This brand was characterized by a disproportionate amount of spam, which could be the reason for the incorrect ordering of Alcatel.

Understanding preferences of product features
As an extension to the method described before, we further tested if semantic networks from Word Embeddings can help us understand the bias toward not just brands but product feature preferences. Such an attempt to understand what people desire goes beyond buzz or sentiment analysis and toward mapping consumers’ minds. The outcome of such a brand-plus-feature analysis could be similar to classic surveys on attribute preferences. Our sample consisted of data from 36,000 reviews of TV sets. Picture quality emerged as the number one feature and sound quality came second. Sound quality’s being so high up should not be surprising since the TV experience actually combines vision and sound, despite its categorization as a visual device. The next most-desired feature was ease of use, and then came the ability to connect with services such as YouTube or Netflix. We learned that unless the process for connecting the TV to the Wi-Fi router and internet is easy, people will not use the TV services for streaming. Not only does getting connected need to be self-explanatory, but the way in which third-party app providers link to the TV platform must be very user-friendly. There seems to be a double-barrelled problem here for adoption! Last among the top five features was the look and style of the TV. Given the current focus on other elements of advertising, we can see how using AI to understand the consumer can help inform the marketing and even the innovation strategy of many manufacturers.

Understanding product benefits
Our extended analysis helped us understand not only top features but also purchase motivations and the benefits that consumers seek. Figure 3 lists the main topics that emerged and the roles they play. Family was ranked first, giving us the insight that TVs are not all about entertainment, or simply watching movies, but rather are devices which can be used and enjoyed by the entire family. Although the top features are picture and sound quality, the purchase motivation for TV sets goes beyond features and into their beneficial effects on home life. The communication implication is that TVs have a connection to social constructs such as families, and should be spoken about in a far more socially-connected manner. Simply stating their ability to provide the best picture and the most engaging movies and gaming experiences does not sufficiently address purchase motives.