Designing a satisfaction proxy metric for Uber when surveys go unanswered

You are a product manager at Uber. The team wants to understand how satisfied customers are, but so few of them respond to satisfaction surveys that the survey results can't be relied on. How would you measure customer satisfaction in this situation? Explain which signals you would use, how you would turn them into something the team can track, and how you would confirm that it reflects how customers actually feel. ```hint Look past the form Customers who never answer a survey still leave traces of how a trip or an order went. Think about which of those traces you could check against the few surveys that do come back. ``` ### Clarifying Questions - Whose satisfaction matters here: riders, drivers, Uber Eats customers or merchants? Each group leaves very different signals. - Is the problem a low response rate overall, or that the responses come mostly from customers who had an unusually good or bad experience? - Does the post-trip star rating count as a survey response here, or is the concern a separate, longer satisfaction survey? - What will the measure be used for: tracking overall health over time, judging individual launches, or spotting individual bad trips to follow up on? ### What a Strong Answer Covers - Why the surveys go unanswered, and whether to repair collection as well as replace it. - Satisfaction signals drawn from specific moments of the Uber trip or order, and the ways each one can mislead. - Combining the signals into a measure the team can track by market, product and customer segment. - Validating the measure against the survey data that does exist, and correcting for who chooses to respond. - The limits of a behavior-based measure and how to use it without it being gamed. ### Follow-up Questions - In a city where Uber has no real competitor, unhappy riders keep riding. How does that affect your measure, and what would you add? - Your measure improves after a launch, but support contacts about the launched feature also rise. How do you work out what happened? - How would you test whether a change to the survey itself (timing, length or channel) raises the response rate without biasing the answers?