Retail MediaBriefing (Crabstone)
A brass robot bearing the Amazon smile arrow marks its own exam paper, headed New Customers, with a red A+, while a crab in a waistcoat holds up a ledger labelled Our Customer List.

Amazon's Ad Agent Grades Itself on Amazon's Definition of New

Amazon's Full-Funnel Campaigns let an AI agent choose the channels and audiences, then report success in new-to-brand customers and Long-Term Sales. Both rest on a definition of new that looks only at Amazon, so a shopper from a label's own shop may count as a fresh acquisition.

Sir John Crabstone

Amazon’s ad agent decides where a brand’s money goes under Full-Funnel Campaigns, then grades the result in a metric that sees only Amazon. At unBoxed this week, Amazon pitched it at “new customers who haven’t discovered you yet”. The advertiser sets a goal, and “AI handles the channel mix, creative, audiences, and optimization.” The customers it then reports as new need only be new to the brand on Amazon.

The grades are flattering, as self-awarded grades tend to be. Amazon’s launch note put the product in open beta for every US advertiser that sells on Amazon, and credits it with 67% higher Long-Term ROAS and 29% lower cost per purchase for new-to-brand customers. The footnote cites early beta data from Amazon’s records for June to August, set against a blended average of the same advertisers’ other campaigns. Long-Term Sales counts immediate conversions plus “the estimated future value from new-to-brand customer actions over 12 months”, so the long term in the headline is partly a forecast.

Amazon’s help page defines new-to-brand purchases as those from “first-time customers of your brand on Amazon”, with a twelve-month lookback. It covers Sponsored Brands and Display; Amazon does not say whether Full-Funnel Campaigns use that definition or how they treat purchases made elsewhere. The wording implies that Amazon sees only what is bought on Amazon. A woman who bought a label’s jeans in its own shop in March, then buys the next pair on Amazon in October, would count as new.

If the agent is rewarded for new customers, the cheapest ones it can find are those who already know the brand.

Digiday reported that Full-Funnel advertisers approve the creative but “cannot select individual channels or specific audiences”. Amit Bhattacharyya, an Amazon Ads vice president, told the site the company is “figuring out how to give them more transparency on the audience side”. Digiday’s questions for buyers include how the agent works with “third-party data, identity and measurement providers”. What counts as “new” is not among them.

Paul Kotas, senior vice president of Amazon Ads, said the agent “understands your end-to-end business”. DVA+ reserves advanced settings for buyers who, in Kelly MacLean’s words, “want full control over every decision”. Neither addresses what counts as new.

Amazon offers eligible advertisers a free check. A brand can upload its hashed customer list to Amazon Marketing Cloud and join it with Amazon’s ad data. If Full-Funnel conversions can be queried there, that could show how many of the agent’s new customers the brand already had; Amazon’s documentation does not say they can.

Amazon’s 2023 example shows a check of this kind. Mars Wrigley, working through Treasure Data, found that a high share of its Amazon conversions came from shoppers who had not bought on mms.com. Amazon presented that as proof of how well it wins new customers. A brand keeps the one list that could contradict the agent; the room where it can be read against Amazon’s numbers is Amazon’s, though Amazon cannot see inside it.