What do 65K customer reviews say about the commercial success in the plant-based meat industry?

Plant-based alternative protein (PBA) products have been gaining popularity and shelf space in supermarkets across the globe. The variety of PBA products is also increasing substantially, both at the brand and product-type level, as more companies join the market and innovation increases across PBA producers. While the industry as a whole competes against animal-based protein products, the competition among PBA producers to become the market leader continues. From either perspective, the PBA market operates as a standard differentiated-products market, competing with an almost fully commoditized animal-based product counterpart.¹

In this emerging market environment, information is one of the most valuable but scarce resources. Consumers want to know which products to purchase based on their preferences while producers want to know the product aspects that are most desirable to consumers and where their products lie in the spectra of these aspects. Policymakers and advocates, on the other hand, want to know how to increase consumer adoption. Yet, there does not exist a comprehensive resource available to the public, regarding how various product aspects contribute to observed commercial success in the marketplace, even though these aspects are identified as limiting factors for the growth of the PBA industry.² To address this knowledge gap, we took on this research project to study customers’ perceptions regarding different aspects of currently commercialized PBA products using customer reviews and analyze the relationship between these aspects and the products’ commercial success.³

Customer reviews provide a very valuable (voluntary disclosure of opinions) window into how consumers evaluate different aspects of a product, both relative to each other and in absolute terms. Our dataset for this project consisted of 64,872 publicly available reviews for PBA meat products on Amazon.com, Walmart.com, Target.com, and Kroger.com.

Collected sample of customer reviews in our sample contained 289,501 sentences which would form our unit of analysis. We randomly selected 2,882 reviews, containing 10,181 sentences, to form our training and validation sample. These sentences were manually labeled by which aspect of the PBA meat product they mentioned and whether the sentence was positive or negative. For example “Tasty.” would be assigned a positive value for the “taste” category, while “extremely chewy” would be assigned a negative value for the “texture” category.

Using OpenAI’s text-davinci-002 model from GPT 3.5, a model designed for natural language understanding and generation, we trained 8 separate models to classify positive or negative sentiment for each aspect of the product. 

The predicted sentiments across the full set of reviews show us not only the relative importance of each aspect but also whether it is more likely to be mentioned in a positive or negative context. That is to say, whether it is something that a product would gain by excelling in, or would suffer from a worse-than-expected experience. For example, versatility was only mentioned in a positive context, meaning that no one left reviews saying the product was not versatile enough, potentially implying that versatility is seen as a bonus, not as a requirement. On the other hand, nearly all of the reviews mentioning how salty the product was were negative, suggesting that the salt level goes mostly unnoticed unless something fails to meet expectations.

Our product level analysis also allowed us to compare various brands to each other:

To estimate the relationship between the reviews and the customer’s rating score, we estimated the effect of a positive or negative comment on each aspect separately, allowing us to combine the effects for reviews that cover more than one aspect, using a Poisson regression:

to predict the score of a review based on the text in that review, estimating the effect on the review score of a good or bad comment in each category by maximizing the log-likelihood function:

is the parameter set.

Our results include both expected and surprising results. Taste being the most impactful aspect of a customer’s perception on their evaluation is not surprising, nor the fact that a negative comment on taste would overpower all of the expected rating improvements from all other aspects. Another implication is that negative sentiment is much more influential than positive sentiment. Quantitatively, we find that with everything else being equal, a 0.1 points increase in average product ratings (say from 4.2 to 4.3 average rating) is associated with a 7.4% increase in its sales volume in 2022 (and 5.5% in 2021).