Operator glossary
Incrementality
Whether an ad actually caused a purchase — the gap between an exposed group and an equivalent unexposed control — expressed as incremental sales, lift percentage, or incremental ROAS rather than console-attributed revenue.
Amazon's console reports attributed sales, not caused ones. The standard model is last touch — a conversion credited to one ad click or view, usually the last one encountered — so every attributed sale lands on some ad whether or not the ad changed anything. The US multi-touch attribution beta splits that credit across touchpoints instead, and Amazon says it draws on experiments and machine learning, but the total still sums to attributed conversions, so it redistributes credit rather than measuring what the advertising added. Causal measurement needs a control group. Amazon Brand Lift is the packaged one — surveys through the Amazon Shopper Panel to an exposed audience and an unexposed control, reporting absolute lift as the difference in response rates — but on perception, not sales, and it excludes Sponsored Products.
For sales you go to Amazon Marketing Cloud and compare exposed against unexposed in SQL. That comparison is not an experiment: the unexposed group is whoever the bidder declined to serve, selected on the exact purchase propensity you are measuring, so lift reads high — which is how retargeting and defensive branded terms keep proving themselves. Last touch seats them nearest the purchase by construction, so the best-ACoS campaign is often the least incremental. A holdout randomizes before exposure and suppresses delivery — a DSP capability, where an audience activates once 2,000 members match, while Sponsored Products has no audience exclusion, leaving geo splits, staggered pauses, or control ASINs read on total units. Size the test accordingly: two weeks on one ASIN measures nothing, and a big clean number usually measures selection.
The numbers
- Amazon states its standard attribution is last touch — one conversion credited to a single ad click or view — and its multi-touch alternative, which divides credit among earlier ads using experiments and machine learning, is a US beta shown alongside the last-touch metrics.
- Amazon Brand Lift is survey-based — it surveys an ad-exposed audience and an unexposed control through the Amazon Shopper Panel and reports absolute lift as the difference in response rates — and it measures brand perception, not sales, with Sponsored Products ineligible.
- Amazon DSP requires at least 2,000 matched audience members before an audience becomes active, and AMC privacy rules never return user_id in workflow output and require at least 100 distinct users per output row for medium- and high-sensitivity columns.
Primary sources: Amazon Ads - Multi-touch attribution (beta): last-touch definition and beta scope · Amazon Ads - Brand Lift (exposed vs. unexposed control methodology) · Amazon Ads API - AMC data aggregation thresholds