On a marketplace product page where several sellers offer the same item, one offer gets the buy button. On Amazon that slot is the Buy Box, formally the Featured Offer. Winning it is close to the whole game, and losing it is close to invisibility.
That much is well known to anyone who sells on a marketplace. What is much less appreciated is the measurement consequence, which is this: when an algorithm decides who gets the buy button, your sales data is primarily a record of that algorithm's decisions. It is only secondarily a record of what customers prefer. Most marketplace analysis treats it as the reverse.
The short answer
The European Commission's own decision text puts the market fact more bluntly than any vendor blog would dare. In its decision of 20 December 2022 in cases AT.40462 (Amazon Marketplace) and AT.40703 (Amazon Buy Box), the Commission wrote:
"Winning the Buy Box is crucial to Amazon marketplace sellers as the vast majority of consumers only view the Buy Box and only buy the offer displayed in the Buy Box."
And, separately in the same summary decision: sales on Amazon's European websites "are driven, to a very large extent, by the system of the so-called 'Buy Box'," with becoming the Featured Offer "essential for third-party sellers to be visible to consumers and to convert their offers into actual sales."
Read that as a research statement rather than a competition one. If the vast majority of consumers only view one offer, then the comparison you assume they made did not happen. They did not weigh you against three rivals and pick you. A system picked, and they accepted the pick. Your conversion rate is substantially a measure of how often the system picked you.
The feedback loop that makes the data circular
The mechanics of marketplace ranking create a loop that is easy to state and hard to escape.
- The algorithm awards the Buy Box on a weighting of price, fulfilment speed, seller performance, and availability.
- Winning it produces sales.
- Sales, sales velocity, and the seller metrics that come with them feed back into eligibility and ranking.
- Which makes winning it again more likely.
The output of step 4 is then read, in the Monday deck, as evidence of customer demand. But the sales came from placement, and the placement came from prior placement. The metric is downstream of itself.
This is the same structural problem we described for promotions in trade promotion effectiveness, where a volume bump gets read as a durable share gain, and for share loss in why did market share drop. The marketplace version is more severe, because the intervening variable is not a promotion you chose to run and can therefore control for. It is a continuous, opaque, third-party ranking decision applied to every impression.
The inversion: rank can move opposite to quality
Here is where marketplace data becomes actively misleading rather than merely noisy.
If the ranking system were a neutral, faithful aggregator of customer preference, then reading preference off rank would be a reasonable shortcut. The public record suggests it is not always a neutral aggregator.
In the monopolisation complaint filed on 26 September 2023 by the Federal Trade Commission and 17 state attorneys general, the FTC alleges Amazon is "biasing Amazon's search results to preference Amazon's own products over ones that Amazon knows are of better quality," and separately that Amazon "can bury discounting sellers so far down in Amazon's search results that they become effectively invisible." The complaint also alleges Amazon degrades the experience "by replacing relevant, organic search results with paid advertisements - and deliberately increasing junk ads."
Amazon disputes these allegations and the litigation is ongoing; none of this is a finding of fact. But note what the allegation describes, independent of whether it is ultimately proven: a mechanism by which the product a shopper is shown can be worse than one they are not shown, and the seller offering the lower price can be the one made invisible. If a mechanism like that operates at all, then in that range rank moves opposite to both quality and price competitiveness - the two things a marketplace share number is usually taken to prove.
The European Commission thought the selection criteria opaque enough to require remedies. Under the accepted commitments, Amazon agreed to apply non-discriminatory conditions and criteria for identifying the Featured Offer, and to display at least one competing offer next to it where available and materially different on price or delivery time - the "Second Displayed Offer" - with both offers showing "the same descriptive information and operating at complete parity with respect to purchasing possibilities." The commitments run for five years across the EEA, are overseen by an independent monitoring trustee, and excluded Italy in view of the Italian Competition Authority's separate decision of 30 November 2021 in case A528.
One detail in the final commitments is the most interesting sentence in the whole file for a researcher. After the market test, the parties agreed to improve the presentation of the Second Displayed Offer "by making it more prominent and to include a review mechanism in case the presentation is not attracting adequate consumer attention."
A regulator mandated a second option, then had to build in a mechanism for the case where shoppers still do not look at it. That is the cleanest available statement of how much interface design, rather than preference, determines marketplace outcomes.
What this breaks in your research
| What teams infer | What the data can actually support |
|---|---|
| "Customers prefer our listing" | Our offer won the buy button on N% of impressions |
| "Our share fell, so demand fell" | Our Featured Offer share fell; demand is unobserved |
| "The price cut worked" | The price cut changed our ranking odds, which changed placement |
| "This variant is our best seller" | This variant was the one shown; the others were a click away |
| "Reviews drove the win" | Reviews are an input to the ranking model that produced the win |
Each row on the left is a claim about people. Each row on the right is a claim about a system. On a marketplace you observe the system directly and the people only through it.
There is a second, quieter problem. Because the ranking system reacts to price, a competitive repricing spiral produces exactly the data pattern that looks like high price elasticity. You cut, you win the box, volume moves. That is not shoppers responding to price. It is an auction responding to price, and shoppers accepting whatever it returned. Applying a demand model to it and concluding that your customers are price-sensitive can lead you to cut price in channels where nobody is running that auction at all. Our documentation on correlation vs causation in research covers the general failure; this is its most expensive marketplace instance.
How to recover the customer signal
The counterfactual you need - what would this shopper have bought if the box had gone the other way - does not exist in any marketplace report. It has to be created.
1. Measure Featured Offer share as a covariate, not a result. Log your Buy Box win rate by ASIN and by day alongside units. Any sales analysis that does not condition on it is attributing placement effects to marketing.
2. Interview buyers off-platform about the choice they made on-platform. This is the only route to the rejected set. Recruit recent category buyers and reconstruct the decision: what they searched, what they saw, whether they scrolled, whether they noticed there were other sellers at all.
3. Ask the question the interface prevents. Most marketplace shoppers have never consciously registered that a Buy Box exists. Asking whether they compared sellers, and what they believed they were choosing between, routinely produces the finding that changes the strategy.
Using the six structured question types, a compact instrument covers it:
| What you need | Type | Example |
|---|---|---|
| The decision as they experienced it | open_ended | Walk me through buying it - what did you look at? |
| Whether comparison happened at all | yes_no | Did you notice other sellers offered the same item? |
| What actually drove the pick | ranking | Rank price, delivery speed, brand, reviews, and habit |
| The real consideration set | multiple_choice | Which of these would you have accepted instead? |
| Channel default | single_choice | Where do you usually buy this category? |
| Substitution strength | scale | If it had shown a two-week delivery, how likely were you to buy anyway? |
The yes_no item is deceptively powerful. When a large share of your buyers report that they did not know other offers existed, you have direct evidence that your marketplace share is an interface outcome, and you can stop running pricing strategy off it. That is a finding you can take to a commercial review, and it is unobtainable from any dataset either you or the retailer currently holds - the same evidence asymmetry we described in category management and shelf space.
Where Koji fits
Off-platform buyer research has always been the right answer here and has rarely been done, for the usual reason: a marketplace team asking for 25 interviews with recent category buyers is asking for six weeks and a budget line, and the repricing decision is due Thursday.
Koji collapses that. Describe the decision you are trying to understand, and Koji's AI consultant builds the discussion guide. AI-moderated voice interviews run with real buyers at any hour, in parallel, with identical probing on every call - so a difference between your Amazon buyers and your direct-site buyers is a real difference, not an artefact of a moderator who found one group more interesting. Thematic analysis and a shareable report come back in one click, in hours rather than weeks, with no research headcount and no moderator bias.
Against the legacy alternatives, the contrast is stark: a traditional agency study on this question is a six-week engagement; a survey platform like SurveyMonkey or Typeform will give you counts but not the reconstruction of a decision; a repository like Dovetail helps only once someone has already done the interviews. Koji does the interviews.
Your marketplace dashboard can tell you how often the algorithm picked you. It cannot tell you whether anyone would have picked you. Start a free study on Koji and find out which one your growth actually depends on.
Related reading
- AI Search Visibility (2026)
- Share of Search and the Digital Shelf (2026) - what a results-page scrape can and cannot prove.
- Agentic Commerce Research (2026) - what happens when an AI agent, not a shopper, accepts the pick.
- Retail Media Measurement - why network-reported ROAS is not evidence.
- Private Label vs National Brands - forced substitution that looks like a value switch.
- AI User Research for Marketplaces - a playbook for two-sided platforms.
- Stated vs Revealed Preferences - the underlying method problem.
- Product Detail Page Research (2026) - what happens on the listing once the offer has been chosen for them.
Frequently Asked Questions
What is the Amazon Buy Box, and how much of sales go through it?
The Buy Box, formally the Featured Offer, is the offer that gets the add-to-cart button when multiple sellers list the same product. The European Commission's December 2022 decision states that sales on Amazon's European sites are driven to a very large extent by the Buy Box, and that the vast majority of consumers only view and only buy the offer displayed in it.
Why is marketplace sales data a poor measure of customer preference?
Because an algorithm selects which offer shoppers see, and the sales that result feed back into the ranking that produced them. Your units are therefore a joint product of placement and preference, with placement doing most of the work, and the two are not separable from the sales file alone.
Does winning the Buy Box prove my price is competitive?
No. It proves your combination of price, fulfilment, availability, and seller metrics scored well against competing offers at that moment under a weighting you cannot see. Price is one input. Reading price elasticity off Buy Box-driven volume typically overstates it, because you are measuring an auction's response rather than a shopper's.
What did the EU require Amazon to change about the Buy Box?
Under commitments made binding on 20 December 2022, Amazon agreed to apply non-discriminatory criteria for selecting the Featured Offer and to display at least one competing offer alongside it where available and different on price or delivery time, at full parity of information and purchasing ability. The commitments run five years, are monitored by an independent trustee, and did not apply to Italy because of a separate national decision.
How do I research what shoppers would have bought instead?
You have to ask them, off-platform, soon after purchase. Reconstruct the decision in their words, then use closed questions to size it: whether they noticed other sellers, which alternatives they would have accepted, and how they would have responded to a slower delivery or a higher price. The resulting rejected set is the counterfactual your marketplace reports structurally cannot contain.
Is this only an Amazon problem?
No. Any marketplace or retailer that selects a default offer or ranks a results grid creates the same feedback loop, including Walmart, Instacart, and app-store style platforms. Amazon is simply the case with the most public documentation, because two competition authorities examined it in detail.
