Swiggy has built an internal machine-learning model that estimates a new customer’s long-term value before that person places a first order. The Indian delivery company uses the ranking with Google’s target-return-on-ad-spend bidding, so acquisition spending can reflect expected future value instead of only an immediate conversion.
The model draws on more than 350 pre-order features, including acquisition channel, device and fraud signals, geography, payment patterns and category affinity. A shared multilayer perceptron produces separate predictions for Swiggy’s Food and Instamart businesses. Adding order count as an auxiliary task reduced the model from 363,000 to 135,000 parameters, a 63% cut, while marginally improving the company’s results.
Because many new users initially generate no value while a small group spends much more, Swiggy evaluates ranking quality rather than relying only on average error. It reports Spearman correlation above 0.75 and 70% to 80% coverage along matching predicted and actual value groups. In production tests, the in-house model improved retention and gross order value per acquired user versus a third-party system, though that outside system acquired more users per dollar.