Scaling Marketing Campaign Forecasting with Generative AI

8 May 2026

Problem

A retailer's marketing team must decide how much to spend on a campaign before any customer sees it. The estimate rests on how similar past campaigns performed, and the same matches train the offer-redemption model. Target's old system used rules and basic embeddings. As campaigns became more niche, it returned more false positives, and analysts had to fix matches by hand, which limited automation.

Example from the article: a new campaign that aims to drive trips among lower-engaged guests in apparel should retrieve past clothing promotions with the same objective.

Solution in brief

Treat each new campaign as a query. Retrieve the nearest past campaigns by embedding similarity, then let an LLM filter and rank them with a prompt that states the matching criteria in priority order. The LLM also returns a short rationale for each match.

Questions this case answers
  • How to select comparable historical cases for a new case described by structured and semi-structured attributes when rule-based matching misses unusual case types
  • How to make an automated analog-selection step auditable by non-technical stakeholders
  • How to find the most similar past campaigns for a new marketing campaign without manual search?
  • How to ground an offer-redemption model in relevant historical campaigns when campaign types get more niche?

Tools and methods

  • Embedding retrieval over a campaign index — gets the top 10 candidate past campaigns for a new campaign.
  • LLM filtering and ranking — picks the top 3 using an ordered matching-criteria hierarchy and explains each pick.
  • Unified offer pool — a cleaned, filtered table of past campaigns and offers with engineered features, used as the index.
  • Validation layer — extra check against LLM hallucinations (details not given).
Continue reading — it's free

Google or a one-time email code. No password.