Eternal Horizons  /  interactive lab  /  companion to the note Amazon Ads, Deeply

The Ad Auction, From the Inside

An interactive tour of how sponsored ads on Amazon actually get priced - the auction, the conversion-rate models, and the hourly bidding loop - built from the published literature.

Every search on Amazon runs an auction in the time it takes the page to load. The platform disclosed in 2026 that 92% of placed ads are not won by the highest bid, and that average winning bids fell 50% over six years while conversion rates rose 24%. Those numbers are only possible because the auction multiplies bids by predicted relevance - so the whole game is modeling. Below: three working models you can play with, each one a miniature of a system that exists in production somewhere.

01 · The mechanismWho wins, and what do they pay?

The sponsored-search auction is a generalized second-price auction (Edelman, Ostrovsky & Schwarz 2007; Varian 2007) with a relevance multiplier. Each advertiser i is ranked by score = bidi × qi, where q is the platform's predicted conversion propensity for that exact (query, product, placement). Winners pay the minimum that would have held their rank:

price1 = ( bid2 · q2 ) / q1  +  ε

Drag the sliders. Watch a high-bid / low-relevance advertiser lose to a cheap, well-converting one - and watch its price fall as its own q rises. Toggle bid-only ranking to see the world Amazon explicitly rejected.

Fig 1. A relevance-weighted GSP auction. Score = bid × predicted conversion propensity q; each winner pays the minimum price that holds its rank, so price falls as your own predicted conversion rises. Expected platform revenue assumes 1,000 impressions and q = P(conversion per impression)-style click propensity for illustration.
The platform's disclosed numbers - winning bids down 50% while conversions rose 24% - are this formula compounding. As q-estimates improve, the ranker increasingly prefers relevance over cash, and everyone with a good product pays less per click.

02 · The prediction problemWhat is a keyword's conversion rate?

The multiplier q is not observed; it is estimated, per search term and placement, from clicks and conversions. Head keywords have millions of observations. The tail - most keywords - has almost none, and the tail is where the unexploited auctions live. The working tool is a Beta-Binomial posterior: start from a prior shared across a keyword cluster, then update with every hour of evidence. A weak prior trusts raw data quickly (and whipsaws on noise); a strong prior shrinks hard toward the cluster mean (and is slow to notice a genuinely great keyword).

6.0%
40
40
Fig 2. Learning a conversion rate, hour by hour. The curve is the posterior over the keyword's CVR (prior mean fixed at an 8% category average; dashed line = truth). The raw estimate (conversions ÷ clicks) is jittery for hours; the posterior mean is stable and still converges. A bidder using the raw estimate overbids and underbids with the noise; a Bayesian bidder prices the posterior mean - and knows its own uncertainty, which is the exploration signal.

Production CVR models add three refinements the toy can't show: multi-task training over the whole impression space to fix selection bias (ESMM, Alibaba 2018), explicit correction for conversions that arrive days late (Chapelle 2014; Ktena et al. 2019), and de-biasing for the fact that top-of-search clicks were easier to win (position-based bandit learning-to-rank). The Bayesian skeleton underneath stays the same.

03 · The bidding loopRepricing ten thousand keywords, every hour

Given a CVR estimate and a margin per conversion, value per click is v = pCVR × margin. But no account is unconstrained: a daily budget must be spread across hours and keywords so the marginal return on the last dollar is equal everywhere (the Lagrangian condition behind Amazon's OPTIMUS bidder, +2-6% sales in A/B tests) and so the budget lasts the day (the pacing problem; Balseiro & Gur). Below: two policies on the same simulated day. Fixed sets one bid from a stale average CVR and never touches it. Hourly loop re-estimates the CVR posterior each hour and re-prices the bid toward true value, with a multiplicative pacing controller steering spend toward an even schedule. In this market, evening shoppers convert better - and evening auctions price accordingly.

$500
$30
$1.40
Fig 3. One simulated day, two bidding policies. Shaded curve: impression volume by hour (diurnal). Lines: cumulative spend. The fixed bid, priced from a stale average, never wins enough auctions to spend the budget - it strands most of it while the high-converting evening passes; the hourly loop learns the keyword's real CVR from the first hours of data, paces its spend, and is still bidding when the valuable traffic arrives - same budget, more profit. This is the loop behind "every keyword repriced hourly".

04 · MeasurementWhy you can't just A/B test a bid

In an auction, your treatment leaks: bid more aggressively in the treatment group and prices rise for the control group too. The fixes are their own literature - switchback experiments that randomize time blocks instead of users (with rerandomization to kill carryover), shadow-price corrections for the equilibrium effect, and simulators populated with adversarial-bandit bidders for anything too risky to try with real money (Amazon's AuctionGym won a best-paper award for exactly this). Offline, logged bids carry bandit feedback - you never see what an unsubmitted bid would have done - so evaluation runs through inverse-propensity and doubly robust estimators.

The meta-lesson of all four sections: the durable edge in this market is not a clever bid. It is a faster, more honest loop - better posteriors, margin-based value, disciplined pacing, and measurement that respects the marketplace.


Sources

  1. Edelman, Ostrovsky, Schwarz - Internet Advertising and the Generalized Second-Price Auction, AER 2007 · pdf
  2. Varian - Position Auctions, IJIO 2007 · pdf
  3. Amazon - Response to the FTC's lawsuit regarding Sponsored Ads, Aug 2026 · aboutamazon.com
  4. McMahan et al. - Ad Click Prediction: a View from the Trenches, KDD 2013 · research.google
  5. Ma et al. - Entire Space Multi-Task Model (ESMM), SIGIR 2018 · arXiv
  6. Chapelle - Modeling Delayed Feedback in Display Advertising, KDD 2014 · doi
  7. Ktena et al. - Addressing Delayed Feedback for Continuous Training, RecSys 2019 · arXiv
  8. Learning to Rank in the Position-Based Model with Bandit Feedback, CIKM 2020 · arXiv
  9. Mondal et al. - OPTIMUS: Optimal Offline Bidding Strategy for Manual Targeting Campaigns, 2026 · amazon.science
  10. Balseiro & Gur - Learning in Repeated Auctions with Budgets, Management Science 2019 · repec
  11. Balseiro et al. - A Field Guide for Pacing Budget and ROS Constraints, ICML 2024 · PMLR
  12. Cai et al. - Real-Time Bidding by Reinforcement Learning, WSDM 2017 · doi
  13. He et al. - A Unified Solution to Constrained Bidding (USCB), KDD 2021 · doi
  14. Jeunen, Murphy, Allison - Learning to Bid with AuctionGym, 2022 · amazon.science
  15. Johari et al. - Interference, Bias, and Variance in Two-Sided Marketplace Experimentation, WWW 2022 · doi
  16. Zhang et al. - Optimizing Multiple Performance Metrics with Deep GSP Auctions, WSDM 2021 · arXiv

Full derivations and the complete reference list: Amazon Ads, Deeply.

Eternal Horizons · an interactive companion to an #evergreen note · in the spirit of Transformer Circuits · back to tannhausergate.saisur.info