The Misconception Most Sellers Have
Ask most Amazon sellers how the ad auction works and they will give you the same answer: highest bid wins. It sounds intuitive. It matches how traditional auctions work. And it is wrong.
Amazon's ad auction is a relevance-weighted system. Your bid is one input. Amazon's prediction of how likely your ad is to generate a click and a purchase is the other. The interaction between these two inputs determines who wins the auction, what placement they get, and what they actually pay.
This distinction matters enormously for how you should think about your advertising strategy. If highest bid wins, then the path to ad success is simply spending more. If relevance matters, then the path to ad success runs through your listing quality, your conversion rate, and your product-market fit — not just your budget.
Understanding the auction mechanism changes the entire optimization framework. You are not just competing on price — you are competing on predicted performance. A lower bid with a better listing can systematically beat a higher bid with a weaker one.
The Enhanced GSP Auction: Bid Times Relevance
Amazon runs a variant of the Generalized Second-Price (GSP) auction, the same mechanism that Google used for years in search advertising. In a standard GSP auction, advertisers are ranked by bid and the winner pays one cent more than the second-highest bid. Amazon's version adds a critical twist: relevance weighting.
The core formula is straightforward:
Ad Rank = Bid × Relevance Score
Amazon confirms this in their own documentation, stating that relevance is "projected to be measured using expected click-through rates and conversion rates." This is not a vague quality signal. It is a quantitative prediction, computed at the query level, for every single auction.
The relevance score is computed per query, per auction. Your listing does not have a single, static relevance score. It has a different predicted CTR and CVR for every search term a customer might type. This means the same listing can have high relevance for one query and low relevance for another — which is exactly why keyword targeting strategy matters so much.
In practice, this means two advertisers bidding the same amount will get different placements based on how well Amazon's models predict their ads will perform. And an advertiser with a lower bid but a significantly higher relevance score will beat a higher bidder with a weaker listing.
How Amazon Predicts Your Click and Conversion Rates
The relevance score is not a simple historical average. Amazon uses deep learning models that predict click-through rate (CTR) and conversion rate (CVR) in real time for every auction. Two published papers describe the production system in detail.
The PACC Model
The primary model is described in "Click-Conversion Multi-Task Model with Position Bias Mitigation", published at SIGIR 2023 (arXiv:2307.16060). This paper describes the PACC model, a production system deployed at Amazon for Sponsored Products prediction.
The key architectural decisions are significant:
- Multi-task architecture: CTR and CVR are predicted simultaneously by a single model, sharing lower layers. This allows the model to learn shared representations while still specializing for each prediction task.
- Position bias correction: The PACC model explicitly debiases predictions to correct for the fact that ads in higher positions get more clicks simply because they are more visible. Without this correction, the model would overestimate the CTR of products that historically appeared in top positions.
- PACC-PE variant: A neural position embedding approach that learns a continuous representation of position effects, rather than treating each position as a discrete category.
The second paper, "Practical Lessons on Optimizing Sponsored Products" (arXiv:2304.09107), provides additional detail on the production constraints. A critical finding: the predicted CTR must be well-calibrated, meaning the model's predicted probability must closely match the actual observed probability. If the model predicts a 5% CTR, approximately 5% of impressions should actually result in clicks. This calibration requirement ensures the auction mechanism functions correctly.
The Practical Implication
Combining these findings, the effective Ad Rank formula becomes:
Ad Rank = Bid × Predicted CTR × Predicted CVR
Seller A bids $2.00 on "stainless steel water bottle." Their listing has a mediocre main image and a 2% predicted CTR with a 6% predicted CVR. Ad Rank: $2.00 × 0.02 × 0.06 = 0.0024.
Seller B bids $1.50 on the same keyword. Their listing has a sharp main image, strong reviews, and a clear value proposition. Predicted CTR: 4%. Predicted CVR: 10%. Ad Rank: $1.50 × 0.04 × 0.10 = 0.0060.
Seller B wins the auction with a 25% lower bid because their predicted performance is substantially better. And in a second-price auction, Seller B pays even less than their bid.
This is not a hypothetical edge case. It is the intended design of the system. Amazon wants to show ads that customers will actually click on and buy from, because that generates more revenue per impression than showing the ad with the highest bid that nobody clicks.
Beyond Binary: How Amazon Selects Which Ads to Show
The prediction models described above produce probability estimates. But turning those probabilities into actual ad placement decisions requires additional machinery. The paper "Ranking and Calibrating Click-Attributed Purchases in Performance Display Advertising" describes Amazon's two-stage approach.
Stage 1: Ranking
The first stage determines the order of candidate ads. A key finding from Amazon's research: ordinal ranking models outperform binary classifiers for this task. Rather than simply predicting "will this ad convert: yes or no," ordinal models predict relative performance across multiple levels (e.g., no click, click but no purchase, purchase of 1 unit, purchase of multiple units). This richer prediction leads to better ranking decisions.
Stage 2: Calibration
The second stage ensures the predicted probabilities are accurate in absolute terms, not just in relative order. Amazon uses a combination of techniques:
- Non-uniform binning with isotonic regression: Groups predictions into bins of varying size and fits a monotonically increasing function to map raw model scores to calibrated probabilities.
- Platt scaling: Fits a logistic function to the model's output scores, providing smooth probability estimates that are well-calibrated across the full range.
Why does calibration matter so much? Because the auction mechanism relies on predicted probabilities to compute Ad Rank and determine payments. If probabilities are miscalibrated — if a predicted 10% CVR actually means 15% or 5% in practice — the auction produces incorrect outcomes and Amazon leaves revenue on the table.
Dynamic Bidding: Amazon Adjusts Your Bids in Real Time
Amazon does not just use your bid as-is. Through its Dynamic Bidding system, Amazon modifies your bid in real time based on the predicted likelihood of conversion for each specific auction. This is a separate mechanism from the relevance weighting in Ad Rank — it adjusts your input bid before the Ad Rank calculation occurs.
Dynamic Bids — Down Only (Default)
This is the default setting for new campaigns. Amazon will reduce your bid when it predicts a lower likelihood of conversion, but it will never increase your bid above what you set. This provides a floor of protection: you will never pay more than your maximum bid, but you may pay significantly less for low-value impressions.
Dynamic Bids — Up and Down (Recommended for New Campaigns)
This setting allows Amazon to both increase and decrease your bid based on conversion likelihood. The adjustments are bounded:
| Placement | Maximum Increase | Maximum Decrease |
|---|---|---|
| Top of Search (first page) | Up to 100% | Up to 100% (to $0) |
| Product Pages | Up to 50% | Up to 100% (to $0) |
| Rest of Search | Up to 50% | Up to 100% (to $0) |
A $1.00 bid with "Up and Down" dynamic bidding can become a $2.00 bid for a top-of-search placement where Amazon predicts high conversion likelihood, or a $0.10 bid for a low-relevance impression deep in search results. Your actual CPC range is much wider than your nominal bid suggests.
Fixed Bids
Fixed bidding uses your exact bid for every auction regardless of predicted conversion likelihood. Amazon will not adjust it up or down. This gives you maximum control but sacrifices the potential efficiency gains from Amazon's real-time conversion predictions.
The Signals Behind Dynamic Bid Adjustments
Amazon evaluates multiple signals when deciding how to adjust your bid:
- Search term: The specific query the customer typed, and its predicted relevance to your product
- Placement: Where on the page the ad would appear (top of search, rest of search, product pages)
- Time of day: Conversion rates vary significantly by hour and day of week
- Historical performance: Your campaign's track record for similar auctions
- Similar ads: Performance data from comparable products in similar auctions
The 25% ROAS Guardrail
Amazon has implemented an automated profitability protection mechanism for Sponsored Products campaigns. If a campaign's return on ad spend (ROAS) falls below a 25% guardrail threshold, Amazon will automatically reduce impressions to protect the advertiser from continued unprofitable spend.
The guardrail only activates when all of the following conditions are met:
The campaign must be at least 30 days old. It must have accumulated at least 30 conversions. The daily budget must be at least $10. And the ROAS must fall below 25% over a 21-day rolling window.
Amazon excludes major shopping events (Prime Day, Black Friday/Cyber Monday) from the rolling window calculation, since these periods have atypical conversion patterns that would distort the measurement.
This guardrail has important strategic implications. If you are running campaigns with intentionally low ROAS — for example, during a product launch where you are prioritizing ranking momentum over immediate profitability — you need to be aware that Amazon may throttle your impressions once the campaign matures past the 30-day/30-conversion threshold.
Four Ways Real Auctions Differ From Theory
The academic model of ad auctions assumes rational bidders with complete information making optimal decisions. The paper "Advancing Ad Auction Realism" (arXiv:2307.11732) from Amazon identifies four critical ways real-world ad auctions diverge from this theoretical ideal.
- Partial bid tuning: Sellers do not continuously optimize their bids. Most set a bid and adjust it infrequently, if at all. The auction must function well even when most participants are bidding suboptimally.
- Incomplete competitive information: Sellers cannot see their competitors' bids, budgets, or relevance scores. They make bidding decisions with severely limited information about the competitive landscape.
- Aggregated feedback only: Amazon provides performance data at the campaign and keyword level, not at the individual auction level. Sellers see average CPCs, not the specific dynamics of each auction they participated in.
- Partially known payment rules: While Amazon discloses that it uses a second-price-style mechanism, the exact computation of actual CPC — including the role of relevance scores, dynamic bidding adjustments, and reserve prices — is not fully transparent.
The paper also identifies the role of soft floor prices (dynamic reserve pricing). Amazon sets a minimum price for each auction below which no ad will be shown. These floor prices are dynamic, varying by query, time, and competitive intensity. They ensure Amazon does not sell ad placements below their estimated value, even if there is only one bidder in the auction.
Amazon is Learning Better Auction Mechanisms
Amazon is not just running a static auction system. The company is actively researching how to design better auction mechanisms using machine learning.
The paper "A Probabilistic Framework to Learn Auction Mechanisms via Gradient Descent", published at AAAI 2023, describes a system that can learn optimal auction rules through gradient-based optimization rather than relying on hand-designed rules. This approach allows the auction mechanism itself to adapt and improve based on observed outcomes.
A second paper, "Optimal Auction Design with Deferred Inspection and Reward", addresses a fundamental challenge in ad auctions: the value of an ad impression is not known at auction time. A click may or may not lead to a purchase, and that purchase may or may not be returned. The paper develops auction mechanisms that account for this deferred and uncertain reward structure.
These papers signal that Amazon's auction mechanism will continue to evolve. Strategies that exploit specific quirks of the current system may stop working as the mechanism is refined. Strategies built on fundamentals — genuine relevance, strong conversion, quality listings — will remain robust because they align with what the auction is optimizing for.
What This Means for Sellers
The research paints a clear picture of how Amazon's ad auction actually functions, and the implications for advertising strategy are substantial.
Listing Quality Is an Advertising Lever
Your main image, bullet points, A+ content, reviews, and pricing are not just "listing optimization." They are advertising inputs. Every element that affects your predicted CTR and CVR directly affects your Ad Rank and your effective CPC. Improving your listing is, in a very real sense, improving your ad performance.
A Better Listing Lowers Your Effective CPC
Because Ad Rank = Bid × Relevance, a higher relevance score means you can achieve the same Ad Rank with a lower bid. Over thousands of auctions, this compounds into a significant cost advantage. Two sellers in the same category, targeting the same keywords, will pay systematically different CPCs if their listings have different predicted performance.
Dynamic Bidding Rewards Strong Converters
When you use "Up and Down" dynamic bidding, Amazon is more aggressive about increasing your bid for high-value placements if your product has strong historical conversion data. Good converters get more impressions in premium placements. Poor converters get throttled. This creates a compounding advantage for products that convert well.
ROAS Targets Need Headroom
With the 25% ROAS guardrail in place, campaigns operating near breakeven need careful monitoring. If your target ROAS is close to 25%, normal performance variance could trigger the guardrail and reduce your impressions. Build headroom into your targets to avoid unexpected throttling.
You Compete Against Amazon's Prediction, Not Just Other Bids
Even if you outbid every competitor, Amazon's model may predict that your ad will perform poorly for a given query — and suppress it or charge you more accordingly. You are always competing against Amazon's prediction engine in addition to competing against other advertisers.
Stop thinking of ads and listing optimization as separate activities. They are two halves of the same system. Your listing quality determines your relevance score. Your relevance score determines your Ad Rank. Your Ad Rank determines your placement and your CPC. Advertising and listing optimization are the same discipline.
- Amazon's ad auction is not highest-bid-wins. Ad Rank equals Bid times Relevance Score, where relevance is predicted CTR and CVR computed per query in real time.
- Amazon uses deep learning models (PACC) that predict click and conversion rates simultaneously, with explicit position bias correction to ensure fair auction outcomes.
- A lower bid with a stronger listing systematically beats a higher bid with a weaker listing. Listing quality is an advertising lever that directly reduces your effective CPC.
- Dynamic bidding adjusts your bids in real time — up to 100% higher for top-of-search if Amazon predicts a conversion, and down to zero for low-value impressions.
- The 25% ROAS guardrail will automatically throttle underperforming campaigns after 30 days and 30 conversions. Build headroom into your ROAS targets.
- Real auctions operate under partial information and uncertainty. The auction mechanism itself is evolving through machine learning. Strategies built on listing fundamentals will remain robust; strategies built on exploiting system quirks will not.