Ecommerce ops · Production

DoorDash builds an explore-exploit ML model to select the best merchant image for each consumer

The problem

DoorDash's image selection was static — every consumer saw the same single image per merchant, often the best-selling item's photo, which could be an unrepresentative side dish rather than an entree and did not adapt to individual consumer preferences.

First attempt

An image rotation experiment improved clicks and new restaurant trials but hurt overall homepage conversion — fresh images attracted interest and click-throughs, but consumers were confused when previously-rejected merchants reappeared differently, generating friction that reduced conversion.

Workflow diagram · grounded in source

Image pool assembly

trigger

“the explore-exploit model finds the most compelling image for a merchant from a pool of six images (five top selling items + header image) for every user session”

Exploitation CVR scoring

ai_action

“The exploitation score is the conversion rate (CVR score) for the image, aggregated over all consumers for each image”

Exploration impression scoring

ai_action

“The exploration score is based on the number of impressions the consumer had on the image. The more impressions, the lower the score”

Composite score image selection

output

“We then select the image with the highest composite model score to display to the consumer. Both the exploitation term and exploration term contribute to the composite score.”

Consumer engagement feedback

feedback_loop

“As a consumer engages or does not engage with the merchant, the model learns their preferences and adjusts the image for the merchant. For instance, If a consumer does not convert on image A from a merchant, the model surfaces a differen…”

Search context image matching

routing

“Surfacing contextualized images on the search feed that includes the best selling item related to what a consumer has searched with a goal to pique their interest and improve the click-through at the very first glance”

Reported outcome

The Image EnE explore-exploit model selects the most compelling image from a pool of six per merchant per user session, and an A/B test showed improvement in new restaurant trials while maintaining conversion rate and order frequency.

Reported metrics

Homepage clicks from image rotation experimentimprovements in homepage clicks

New restaurant trials from image rotation experimentimprovements in new restaurant trials

Homepage conversions from image rotation experimentnegative impact on homepage conversions

new restaurant trials from EnE A/B testimprovement on new restaurant trials

Show all 6 reported metrics

homepage clicks from image rotation experimentimprovements in homepage clicks

new restaurant trials from image rotation experimentimprovements in new restaurant trials

homepage conversions from image rotation experimentnegative impact on homepage conversions

new restaurant trials from EnE A/B testimprovement on new restaurant trials

conversion rate from EnE A/B testmaintaining conversion rate

search conversion from context-matching A/B testneutral impact on search conversion

Reported stack

multi-arm bandit algorithmUCB

Source

https://careersatdoordash.com/blog/selecting-the-best-image-for-each-merchant-using-exploration-and-machine-learning/

Read source ↗

Frequently asked questions

What did this team achieve with this AI workflow?

The Image EnE explore-exploit model selects the most compelling image from a pool of six per merchant per user session, and an A/B test showed improvement in new restaurant trials while maintaining conversion rate and…

What tools did this team use?

multi-arm bandit algorithm, UCB.

What results were reported?

Homepage clicks from image rotation experiment: improvements in homepage clicks; New restaurant trials from image rotation experiment: improvements in new restaurant trials; Homepage conversions from image rotation experiment: negative impact on homepage conversions; new restaurant trials from EnE A/B test: improvement on new restaurant trials (source-reported, not independently verified).

What failed first in this deployment?

How is this ecommerce ops AI workflow structured?

Image pool assembly → Exploitation CVR scoring → Exploration impression scoring → Composite score image selection → Consumer engagement feedback → Search context image matching.