Problem
A delivery marketplace shows each merchant with one image on the home feed and search feed. Consumers judge an unfamiliar restaurant by that picture first. The first rule showed the best-selling item, which is often a side or a drink and not an entree. Picking the image with the best past conversion fails too: new images have a conversion rate of 0 and never get shown, tastes change, and a consumer who ignored an image keeps seeing it.
Example from the article: a popular side such as fries or a soda was featured instead of an entree that would better represent the merchant.
Solution in brief
Collect data with an image rotation experiment, then score each image by its conversion rate plus an exploration term that shrinks with the number of impressions a given consumer has already had. Show the top-scoring image. The weight of the exploration term is chosen by an offline simulation on the rotation data and then tested online.
Questions this case answers
- How to choose one item from a small pool of visual options per entity so that conversion rises when new options have no history and tastes drift
- How to set the exploration weight of a bandit-style score offline, before an online test with limited traffic
- How to choose which image represents each merchant on a discovery feed so that it raises conversion, not just clicks?
- How to let new or stale images get exposure without locking into the historically best one?
Tools and methods
- Multi-armed bandit (UCB-style) — Image EnE (exploitation and exploration) score per consumer, merchant and image.
- Image rotation experiment — A/B test that rotates images and produces the conversion data for the first scores.
- Offline simulation of image changes — estimates how often the image changes after X views for each exploration multiplier.
- A/B testing — compares multiplier values and the contextual search experience.
Google or a one-time email code. No password.