Problem
Canva's performance marketing team bids on tens of thousands of keywords per Apple Search Ads campaign, across many regions. Apple provides no automated bidding, so someone has to keep every bid current. A SQL rules system multiplied each bid up or down by keyword quality, but it learned much faster from winning bids than from losing ones. Bids drifted down, only a few keywords kept winning impressions, and marketers began overriding the rules by hand.
Example from the article: a keyword that a competitor starts outbidding gets zero impressions, so the rules have no signal on how far to raise the bid.
Solution in brief
Replace the rules with three parts: a per-keyword model of spend as a function of bid, a per-keyword model of return as a function of spend, and a budget optimizer. Thompson sampling over the return curves makes uncertain keywords explore, and a pooled LightGBM model supplies a prior for keywords with little history.
Questions this case answers
- How to set per-item bids under a budget when only winning outcomes are observable and the unobserved competitor price is censored
- How to keep exploring items that currently look unprofitable while spending close to a fixed budget
- How to automate bids on tens of thousands of ad keywords under a fixed weekly budget?
- How to estimate what a higher bid would have won when the ad platform reports only won auctions?
Tools and methods
- Turnbull estimator (survival analysis) — win probability and expected spend per bid from censored daily data
- Hill function — diminishing-returns curve from spend to return for each keyword
- Thompson sampling — a different sampled curve each day, so uncertain keywords explore
- LightGBM — one model over all keywords, used to add synthetic points to sparse keywords
- L-BFGS — numerical solver that splits the daily budget across keywords
Google or a one-time email code. No password.