Problem
Avito's monetization team had to move sellers of daily rentals from a fixed percentage plus a prepaid promotion budget to a single commission on a successful booking. The seller picks the percentage within a range, and a higher percentage lifts the listing in search. The old results page had a separate auction for paying listings, and ranking had to change with the payment model. The risk was lower revenue, or fewer bookings, while sellers were still learning the new product.
Example from the article: in a first internal A/B test lowering the weight of the commission in ranking, revenue fell. The team traced the drop to conversion models trained on data with little short-term rental, plus several pipeline bugs.
Solution in brief
Replace the paying-listings auction with one ranking of all listings by a linear combination of expected revenue and user quality, where expected revenue is . Launch it in a regional A/B test, with fallbacks for missing predictions. When revenue drops, check the models on the new segment and the whole pipeline end to end, then retrain and add segment-specific features.
Questions this case answers
- How to switch a two-sided ranking from fixed promotion budgets to a per-transaction commission model without losing revenue when the platform earns only on completed deals
- How to find which part of a ranking pipeline caused a revenue drop after a change, when the models were trained on data that under-represents the new segment
- How to switch search ranking from prepaid promotion budgets to commission-based payment without losing revenue
- How to compute expected revenue for ranking when sellers pay a commission on a successful deal
Tools and methods
- Expected-revenue ranking — one formula for paid and organic listings
- Regional A/B test — test and control regions, new regions added step by step
- Boosting conversion models, CatBoost — contact and deal probabilities
- Fallbacks and monitoring — protect ranking when models fail
Google or a one-time email code. No password.