Why Firms Want to Personalize Prices
The ZipRecruiter study provides a controlled way to examine the business case for personalized pricing: what happens when a firm deliberately varies prices, learns how different customers respond, and uses that information to target offers. This chapter follows the study from randomized evidence to revenue, then asks a harder question about consumers: who gains, who loses, and which definition of welfare is being used. The result is not a universal endorsement of personalization. It is a clear explanation of why a firm may want to do it even when the consumer judgment remains mixed.
A controlled experiment reveals the business case
The study by Jean-Pierre Dubé and Sanjog Misra examined ZipRecruiter, an online service that matches jobseekers with employers. The relevant customers were prospective employers, not household grocery shoppers. They were business customers reaching a paywall for the first time, and the service was sold as a monthly subscription that could be canceled. That setting matters because the study’s results are evidence about a particular enterprise market, not a direct measurement of what happens when a grocery platform shows different prices to households.
The first experiment ran from August 28 to September 29, 2015. The study’s narrative says that 7,867 unique prospective customers reached the paywall and were randomly assigned to ten pricing cells. The control price was $99 per month. The other cells ranged from $19 to $399. The extracted table and model description use 7,866 firms, a small internal count difference that does not change the design but should not be silently erased.
Random assignment is the key methodological choice. If a firm simply observes that customers who paid more were more likely to buy, it cannot tell whether the price caused the purchase difference or whether those customers were already more likely to buy. In the experiment, the price was assigned before the purchase decision. That makes the conversion pattern across cells more informative about the effect of price, at least for this population and time period. It does not make every later prediction causal. The personalized prices were built from a model, and the model still depends on assumptions about demand and the information available about each customer.
The researchers used 12 categories of self-reported business and job features, represented as 133 indicator variables. They estimated not only a general response to price but also how that response varied with the features. Their weighted-likelihood-bootstrap Lasso procedure repeatedly selected and estimated feature relationships while representing uncertainty about the model. In plain language, the method tried to avoid treating every possible feature relationship as reliable simply because it could be fitted to the data. It estimated which observed differences were useful for predicting price response and carried uncertainty into the pricing decision.
The experiment first showed that ZipRecruiter’s $99 price was probably too low for its own short-run revenue objective. Raising the price from $99 to $199 produced only about a 25 percent decline in conversions in the model-free comparison. Because the service had very low marginal cost and available capacity, the lost conversions did not offset the added revenue from the customers who still subscribed. The estimated decision-theoretic optimal uniform price was $327.