How Ad Ranking Works at Avito

Aleksei Kozlov23 May 2026

Problem

Avito's young ad platform (under a year old) runs a separate first-price auction for each ad slot. To pick the ad with the highest expected revenue per impression, it needs a click probability that is accurate and comparable across ads. The launch baseline, the campaign's CTR over the last 7 days, ignores the user, context and banner, and is noisy for new campaigns. Heuristic patches do not scale.

Example from the article: the first calibrated ranker went to an A/B test. Revenue grew sharply, but vCTR dropped, because scores were so close that the bid alone decided the winner.

Solution in brief

Train a CatBoostRanker, turn its score into a probability with Platt scaling, and retune the sigmoid coefficients on a grid so scores differ enough to matter next to the bid. Add the cosine between a user embedding and an ad embedding (built from text and image) as a feature. Result: +14.8% vCTR and +6.4% revenue over the CTR baseline.

Questions this case answers
  • How to turn a ranker score into a calibrated click probability so that it can be multiplied by a bid in a first-price auction
  • How to keep calibrated predictions informative when the bid would otherwise dominate the product of prediction and bid
  • How to rank ads in a first-price auction by expected revenue per impression
  • How to calibrate a click-probability model so that multiplying it by the bid makes sense

Tools and methods

  • CatBoostRanker — click-probability model, trained as a ranker
  • Platt scaling and post-calibration — sigmoid with temperature and offset, retuned for score spread
  • Two-tower embeddings — frozen user tower and a trained ad tower, cosine as a feature
  • SoftMarginTripletLoss — ad tower loss
  • ViT / DINOv3, rubert-tiny2 — pretrained image and text encoders for banners
  • Spark, Redis — batch features and creative vectors
Continue reading — it's free

Google or a one-time email code. No password.