Social Rocket
Amazon Ads wants to measure advertising impact beyond the immediate attribution window. Digiday reported on 1 October 2026 that Amazon, in an analysis of more than two million campaigns, found that around 20 percent of brand value only becomes visible after the usual attribution window.
Long-Term Sales keeps counting for up to twelve months
Amazon describes its Long-Term Sales metric as an ad-attributed measure that combines direct sales with the predicted value of customer interactions over the following twelve months.
To do this, Amazon looks at, among other things, new-to-brand customers and early signals such as brand searches or shopping carts. Historical purchase patterns are meant to show what additional revenue a campaign could generate over the long term.
For media teams this is appealing, because upper-funnel advertising moves closer to a monetary value. At the same time, Long-Term Sales remains a modelled forecast and not revenue that has already been realised.
Agencies want to check the forecast against their own data
According to Digiday, this is where the friction lies. Some agencies use the metric to better understand effects beyond classic last-click models. Others compare Amazon's forecasts with marketing mix models, D2C revenue or in-store sales and remain cautious when the figures don't match.
Advertisers know this basic problem from other automated media systems: the more a platform optimises and measures itself, the more important an independent control layer becomes. Our classification of AI in media buying fits this point.
Offsite impact is also becoming more important in the retail media market. Walmart Connect is expanding CTV, social and offsite, while with ChatGPT Ads measurement gaps are currently holding back larger budgets.
What marketers should take away
Long-Term Sales can be helpful when campaigns are not optimised solely for immediate purchases. For budget decisions, however, the metric should be considered alongside your own revenue, CRM and incrementality data.
The central question is not whether a platform can model long-term value, but how well that forecast can later be validated against actual sales.