Amazon DSP Supports Creation of Lookalike Audiences From: The Three Seed Sources

What's actually being asked here

This phrasing shows up most often as a fragment of an Amazon Ads certification exam question, cut off before the answer choices. The full question asks which sources an advertiser can use as a seed to build a lookalike audience inside Amazon DSP — not what Amazon's own shopper data looks like, and not a general definition of lookalike modeling. The honest, complete answer is three seed types, all of them the advertiser's own existing audience data, not Amazon-sourced.

The three seed sources, one at a time

Hashed audiences are customer lists an advertiser already has — email addresses or other identifiers from a CRM system — uploaded to Amazon Ads and hashed before matching, so Amazon never receives the raw customer data itself. This is the most common seed for brands with an established customer database outside of Amazon.

Pixel or Amazon Ad Tag event audiences are built from behavior Amazon can observe directly through a tracking tag placed on the advertiser's own site — page views, add-to-carts, purchases — turned into an audience of people who took a specific action, which then becomes the seed for a similar-audience model.

DMP-transferred audiences are segments an advertiser has already built inside a connected data management platform and passes into Amazon DSP through that integration, rather than uploading a raw list or relying on Amazon's own tag.

Amazon's own advertiser-audiences documentation describes this as giving advertisers "a way of incorporating their existing audiences into their Amazon Ad campaigns," and describes building a lookalike "from their hashed audience" specifically — the seed is always something the advertiser already owns or has already defined, even when Amazon's modeling is what expands it into a larger similar audience.

How the modeling actually turns a seed into a bigger audience

Once a seed is in place, Amazon's similar-audiences capability uses machine learning to score the broader pool of Amazon shoppers against that seed's behavioral pattern, surfacing new people who resemble it without Amazon ever having to reveal who's in the original seed. This is where Amazon's retail signal becomes relevant — not as the seed itself, but as the pool the model searches to find new matches. The seed has to come from the advertiser; the matching happens against Amazon's own shopper data.

It's worth being precise about what this means for privacy: the advertiser never learns the identity of any individual matched into the resulting similar audience, and Amazon never receives the advertiser's raw customer data in the hashed-audience case — only the hashed values, matched against Amazon's own hashed records. Neither party sees the other's underlying identity data at any point in the process, which is also why Amazon can activate this feature without the advertiser needing its own separate consent or data-sharing infrastructure beyond the standard upload flow.

A worked example: what a thin seed actually costs you

Say a brand uploads a hashed CRM list of 400 past customers as its lookalike seed. That's below the volume most platforms need to find a stable pattern — a seed that small produces a similar audience the model is essentially guessing at, because there isn't enough shared behavior across 400 people to separate signal from noise. The fix isn't a bigger percentage or a different targeting setting; it's a bigger or more homogeneous seed. A pixel-based seed of everyone who purchased in the last 90 days, even if it's a few thousand people who all did the exact same thing, will consistently outperform a larger but mixed CRM list where browsers, one-time buyers, and repeat customers are all lumped in together.

The mistake, and what to check when a lookalike isn't working

The common mistake is treating any of the three seed types as interchangeable and picking whichever is easiest to export, rather than the one that best matches the campaign's actual goal. A DMP segment built for a different channel's definition of "high value" won't necessarily translate into a useful Amazon seed. If a lookalike audience isn't performing, check seed size first, seed homogeneity second — mixed intent inside one seed dilutes the pattern the model is trying to find — and seed freshness third, since a seed built from data six months old is matching against who the customer was, not who they are now.

In our own accounts, DSP prospecting audiences built this way run around two-thirds new-to-brand, and the visible effect tends to show up first in rising branded search on Amazon rather than in the DSP report itself — a brand watching only its DSP dashboard will systematically undervalue what a well-built lookalike is actually doing.

Where Dr. DSP fits

Dr. DSP is Amazon DSP run as a managed product by Full Circle, which has managed more than $500M in revenue across 100+ brands. Picking the right seed — and checking its size and freshness before blaming the audience — is a small, specific decision that quietly decides whether a lookalike campaign works. A reader who never buys anything from us should still leave this page knowing the exam-style answer to this question and, more usefully, which of the three seed types actually fits their own customer data.