Amazon Marketing Cloud and Incrementality: Ad Sales Versus Ad Impact

Amazon Ads reports attributed sales: purchases that followed a click or view on your ad within a set window. Attribution tells you the ad was there. It does not tell you the ad was the reason. A shopper who types your brand name and clicks your sponsored result may have bought regardless. Amazon Marketing Cloud, usually shortened to AMC, is one place to examine that question, and a holdout test is the method that answers it. This article stays conceptual: what AMC is, why attributed and incremental sales differ, and how a test could separate them.

What Amazon Marketing Cloud is

AMC is a secure, privacy-safe analytics environment, often described as a clean room. Inside it, advertisers can query pseudonymized event data about ad exposure and shopper behavior, and in some setups combine it with their own first-party data. You work with queries or ready-made templates, and results come back aggregated, with privacy thresholds that stop individual shoppers from being identified. You get patterns, not names.

That makes it different from the standard campaign reports. The reports summarize each campaign on its own. AMC can look across ad types and across time, which lets you ask questions such as how many shoppers saw an ad before they bought, how often they saw it, and whether combining two ad types changes the outcome. It is not a tool for managing campaigns or a source of customer lists, and it is not an experiment engine on its own. It gives you the data environment, and the experimental design is your job.

Access rules and eligibility have changed over time, so check what is available to you through Amazon Ads or your agency. The concepts below hold regardless of the interface.

Attributed sales and incremental sales are different things

Attributed sales are the purchases Amazon credits to an ad because they followed a click or view inside the attribution window. Incremental sales are the purchases that would not have happened without the ad. The first is a bookkeeping rule applied to what happened. The second is a statement about what would have happened otherwise, which nobody can observe directly and which has to be estimated.

Take a hypothetical campaign credited with 2,400 orders for $8,000 of spend. That is about $3.33 per order, which looks excellent. The unanswered question is how many of those 2,400 customers would have bought without seeing the ad.

The gap is usually widest where ads reach shoppers who are already close to buying: searches for your own brand name, retargeting of people who recently viewed the product, repeat customers. It is usually narrower where ads reach people who have never heard of you, such as broad category terms. That is reasoning, not a law, and only a test can say. The gap can also run the other way, since ads can lift organic ranking or plant awareness that pays off after the attribution window closes.

Why a simple comparison misleads

The tempting shortcut is to compare shoppers who saw your ad with shoppers who did not, and call the difference the effect. It overstates the effect. People who see a Sponsored Products ad are searching for something related to your product, so they were interested before the ad appeared. People who never saw it are a different mix: some were never in the market at all.

Before-and-after comparisons have a similar flaw. If sales rise after you launch a campaign, the cause could be the campaign, a seasonal peak, a price change, a promotion or a competitor running out of stock. Without a control, you cannot separate these.

How a holdout test could separate the two

A holdout withholds the ad from a comparable group and measures the difference. The details depend on what each ad product lets you control.

Some ad types let you split an audience and exclude a share of it. Sponsored Products does not let you choose which shoppers see an ad, so a test there more often compares matched products, comparable regions or alternating periods. Each design has weaknesses. Matched products are never identical, and short on-off periods pick up noise. State the weakness before you read the result.

Suppose you split an audience of 200,000 shoppers evenly. The exposed group of 100,000 buys at 2.4%, which is 2,400 orders. The holdout group of 100,000 buys at 2.0%, which is 2,000 orders without any ad. The ads added about 400 orders, one in six of the 2,400 credited to them. In AMC terms, you would define the two groups, query their conversions and read the aggregated result. The lift is the difference in rates, 0.4 points, divided by the holdout rate of 2.0%, which is 20%.

A holdout has a price. Withholding branded ads from some shoppers can let a competitor's ad take the top spot for those searches, and any orders you lose are real. Keep the withheld share small, keep the duration short, and decide beforehand what loss you are willing to risk for the information.

Reading the result

The incremental cost per order is $8,000 divided by 400, which is $20. Compare that with what an order earns. Suppose an order leaves $14 of profit before ads. Then $20 per incremental order loses money on the first purchase, even though the $3.33 attributed figure looked great. A repeat purchase may justify some of the gap, but only if you can show that customers return.

Before acting, ask whether the difference is bigger than normal week-to-week wobble. A small gap in a small group can be chance. Repeat the test, or run it longer, before making a large cut. Do not extend one result to every campaign. A branded campaign and a category campaign can behave differently, and the same campaign can behave differently after a competitor changes its strategy.

The practical use of incrementality is to decide where the next dollar goes: toward campaigns that reach new shoppers, away from those that collect credit for sales that were coming anyway. Keep standard attributed reporting for daily management, since it is faster and still useful, and treat incrementality as the periodic check on whether that reporting is flattering you.

Where to start

• Find out whether AMC is available to you through Amazon Ads or a partner.
• Pick one campaign where you suspect attributed sales overstate impact, such as branded search.
• Choose a control design that fits the ad product, and write down its weaknesses.
• Freeze prices, promotions and stock during the test as far as you can.
• Compare cost per incremental order with the profit an order earns, then decide.