Lookalike audiences

The version most people mean lives in Amazon Marketing Cloud: SQL resolves a seed of 500 to 500,000 user_id values, you POST it to the lookalike endpoint, and an AWS machine learning model — Amazon's own description — returns a larger set with similar characteristics. You cannot point the API at an existing rule-based audience; the SQL has to be resubmitted. Output size is not a percentage dial but one of five tiers on lookalikeAudienceExpectedReach, each an absolute, region-dependent band. The tiers nest: SIMILAR contains MOST_SIMILAR, BALANCED contains both, MOST_BROAD contains everything — moving up dilutes your best-matched population rather than swapping it out. The audience targets in DSP and Sponsored Display and modifies bids in Sponsored Products and Sponsored Brands. Amazon does not publish the model's features or a lift benchmark.

The failure mode is the seed count: the number you validate and the number the model receives are not the same. Amazon states only adUserId is targetable and calls 90% adUserId coverage an important factor limiting the full audience size; the final audience also omits users who have opted out of targeting and any users identified as potential children. Amazon warns that a seed above 500 user_id values can still come back too small when its ids aren't found in the matched population — a seed of 600 can come back FAILED hours later, reported in statusReason rather than at submit time. Size seeds in the low thousands and start at MOST_SIMILAR; MOST_BROAD in Europe and the Americas is 3M–10M. Sync takes 2 to 24 hours and refreshRateDays accepts only 7 to 21 — not a Prime Day week decision.