Target Uses AI and Language Models to Match Marketing Campaigns for Better Forecasts

Target has unveiled a machine learning system that assists marketing and analytics teams in planning campaigns by identifying and ranking similar historical efforts. Rather than predict outcomes directly, the platform surfaces relevant past campaigns that inform forecasting decisions, aiming to reduce manual effort and improve consistency across the marketing organization's growing number of channels and formats.

How the System Works#

The architecture relies on a multi-stage pipeline. Historical campaign data—including audience segments, product categories, channels, and intent—is converted into semantic embeddings that capture meaning from structured attributes. When a new campaign is created, the system generates an embedding from its metadata and retrieves candidate matches from an internal index. A large language model then ranks these candidates using structured constraints and contextual signals, returning an ordered list with explanations for each recommendation. Human marketing analysts review the model's suggestions and explanations before incorporating them into forecasting workflows.

Coverage and Performance Gains#

In evaluation using time-separated train-test splits across recent campaigns, the system achieved 75% coverage when considering only the top-ranked recommendation. When expanded to the top three suggestions, coverage rose to 100%—meaning every evaluated campaign had at least one viable historical match. This improvement significantly reduced the need for manual searching and corrections during campaign matching.

From Rules to Learned Similarity#

The new platform replaced an earlier rule-driven system that required constant maintenance and struggled as campaign formats evolved. Maintaining rule sets became increasingly burdensome, and older rules often failed to generalize to newer campaign types. The semantic approach sidesteps manual rule updates by learning similarity patterns directly from data, and it extends coverage to long-tail campaign types that previously had no reliable rules.

The system also incorporates a feedback loop: performance data from completed campaigns refines embeddings and improves retrieval quality, allowing the platform to adapt as campaign patterns shift over time. By separating embedding generation, retrieval, and LLM ranking into distinct stages, Target gained independent tuning and better visibility into each step—keeping human judgment central to the forecasting process.

For more details, see .