Problem
DoorDash restaurants set a monthly ad budget, split evenly by day. Serving stops when the daily spend goal is hit, but a consumer who clicked earlier may order later, and that order is still billed, up to twice the daily budget. For enterprise campaigns this meant about 11.5% overspend in the last week of May 2024 and nearly 25% in the first week of June. Separately, customers submit the same photo, or a scaled or rotated copy, as evidence for refunds on different deliveries, and agents cannot spot repeats by hand.
Example from the article: a campaign switched to fast finish in the evening, spent sharply, and kept being charged after serving stopped.
Solution in brief
For pacing, replace the abrupt switch to fast finish with a transition window that lowers throttling gradually, and start fast finish later for campaigns with high past overspend. For photo fraud, embed each refund photo with a vision transformer and flag pairs above a cosine-similarity threshold; this catches edits that hashes miss.
Questions this case answers
- How to cut spend that lands after a daily budget goal is reached when conversions are attributed with a delay and billing is capped at twice the budget
- How to find reused or slightly edited images submitted as evidence when exact file hashes fail and manual search is impractical
- How to stop an advertiser's budget from being overspent after the daily spend goal is reached?
- How to keep ad campaigns live longer in the day without raising overspend?
Tools and methods
- Fast finish (FF) with a transition window — last pacing stage, softened by a gradual throttle drop
- Dynamic FF start time — per-campaign start hour from history
- A/B test — one-week online check
- SHA-256 and perceptual hash (P-hash) — baselines for exact and near duplicates
- ViT embeddings with cosine similarity — duplicates despite rotation and scaling
Google or a one-time email code. No password.