GenAI-Powered Accessory Recommendations for Large-Scale E-commerce Catalog

target · E-commerce · 2025 · source Agentic: no · promotional bias: low · source-backed · extracted with prompt 0.5.6 (schema v2)

Generated from extraction.json — do not edit by hand. Every claim cites verbatim source evidence; stated means the quote directly asserts the claim, implied means it supports it indirectly.

Problem

Domain: e-commerce accessory recommendation

Target needed a scalable related-accessory recommendation algorithm for Electronics and Home that could identify which product attributes matter, capture aesthetic matches, and operate across a large catalog where manual expert scoring would be too slow.

Constraints:

• the catalog volume made manual attribute-priority evaluation impractical
• different products require different attribute priorities such as color, material, kid-friendliness, flavor, brand, or style
• some accessory matches depend on aesthetic coordination rather than exact attribute equality
• hundreds of thousands of Home-category items required time- and cost-efficient scaling
• recommendations needed to improve measurable business outcomes in add-to-cart placement

Architecture

Target built GRAM, a GenAI-based Related Accessory Model for Home. LLMs analyze product data to determine attribute importance and scoring rules for pairs of core and accessory item types. The scoring algorithm applies those rules to candidate accessory items, adds weighted relevance when core and accessory attributes match, uses sales rank as a tie-breaker, and parallelizes scoring/sorting across item-type pairs. Site merchants add human insight by curating commonly co-purchased accessory items for cross-category recommendations, and Target validates the capability with an A/B test before full production rollout.

Diagram confidence: reconstructed (reconstructed from prose).

flowchart LR
  product-catalog["Product catalog"] --> llm-attribute-analyzer["LLM attribute analyzer"]
  llm-attribute-analyzer --> scoring-rule-generator["Scoring rule generator"]
  scoring-rule-generator --> accessory-scoring-engine["Accessory scoring engine"]
  product-catalog --> accessory-scoring-engine
  sales-rank-tie-breaker["Sales rank tie-breaker"] --> accessory-scoring-engine
  human-merchant-input["Human merchant input"] --> accessory-scoring-engine
  accessory-scoring-engine --> recommendation-surface["Recommendation surface"]
  recommendation-surface --> a-b-testing["A/B testing"]

Components

Data flow

Product data for core and accessory item types is analyzed by LLMs to produce attribute weights and scoring rules. For each core/accessory type pair, the algorithm scores candidate accessory items by adding weights when relevant attributes match, sorts the candidates with sales rank as a tie-breaker, and augments the candidate space with merchant-curated co-purchase knowledge for cross-category recommendations. The resulting model is placed in the add-to-cart flyout and measured through A/B testing.

Why this architecture

Target used LLMs where product-pair semantics and aesthetic relationships were hard to encode manually, but constrained scaling by generating rules for item-type pairs instead of item pairs. Human merchant input complements the model where cross-category accessory knowledge improves diversity and basket expansion.

Models used: GRAM · Large Language Models · Home Accessory model · attribute-weight scoring rules · sales rank

Scale

• Target built the related accessory algorithm for both Electronics and Home categories.
  “asked to create a “related accessory” algorithm for both our Electronics and Home categories.”article, stated
• The Home category contains hundreds of thousands of items.
  “With hundreds of thousands of items in the Home category, it was crucial that our algorithm scale effectively.”article, stated
• The production A/B test ran in February 2025.
  “In February 2025, we conducted an A/B test where we added the Home Accessory model to the add-to-cart flyout.”article, stated
• The A/B test showed around an 11% increase in interaction rate.
  “Around an 11% increase in interaction rate indicating heightened engagement”article, stated
• The A/B test showed roughly a 12% increase in display-to-conversion rates.
  “A roughly 12% increase in display-to-conversion rates reflecting stronger content relevance and downstream impact”article, stated
• The A/B test showed more than 9% growth in attributable demand.
  “More than 9% growth in attributable demand”article, stated
• The model was fully rolled out to production in April 2025.
  “this model was fully rolled out to production in April 2025”article, stated

Alternatives considered (roads not taken)

Manual expert evaluation of attribute importance

Rejected because: The catalog was too large for human experts to manually prioritize relevant attributes for every accessory pairing.

Exact attribute matching only

Rejected because: Some accessory pairings require aesthetic coordination such as color harmony and stylistic coherence rather than literal attribute equality.

Evaluate every core-item/accessory-item pair directly

Rejected because: Target reduced time and cost complexity by scoring at the core/accessory item-type-pair level and then applying rules to item pairs.

Use model-only recommendations without merchant input

Rejected because: The model without HITL often suggested similar items, while merchant input produced a more diverse accessory set useful for basket expansion.

Patterns

human-in-the-loop-gate (recommendation quality)

Site merchants contributed commonly co-purchased accessory knowledge so the LLM-based recommender could support broader cross-category recommendations.

batch-offline-processing (scalability)

After LLM-derived scoring rules were established, Target scored and sorted all item pairs for relevant item types, parallelizing the operation for catalog scale.

latency-optimization (scalability)

The design avoids direct pairwise LLM reasoning over item pairs by constraining LLM work to item-type attributes and reusable scoring rules, reducing time complexity.

eval-driven-development (experimentation)

Target validated the Home Accessory model with an A/B test in the add-to-cart flyout and reported interaction, conversion, and attributable-demand lifts before production rollout.

Tradeoffs

How to generate accessory scoring logic

Chose LLM-derived attribute weights and scoring rules over manual expert scoring for each accessory pairing.

Rationale: LLMs made it practical to analyze product data and assign pair-specific attribute importance across a catalog too large for manual expert evaluation.

How to scale scoring

Chose core/accessory item-type pair rules over rules or LLM scoring for every individual item pair.

Rationale: Item-type rules reduce time and cost while still allowing the scoring engine to rank all items in the relevant types.

How to balance exploration and basket expansion

Chose support both model-only and HITL-enhanced recommendations over one universal recommendation mode.

Rationale: The model-only version helps guests compare similar accessories, while the HITL version exposes broader accessory types for basket expansion.

How to break ties between equally scored accessories

Chose sales rank as tie-breaker over arbitrary ordering among equal relevance scores.

Rationale: When attribute-weight scores are equal, sales rank provides an additional business-informed ranking signal.

Gotchas (what bit them)

Attribute importance varies by product pair.

Root cause: Guests may care about different features depending on the core and accessory items, so one global attribute priority would be wrong.

Resolution: Use LLMs to analyze product data and assign attribute weights for different core/accessory item-type pairs.

Aesthetic compatibility cannot be reduced to exact attribute equality.

Root cause: Some accessories need to coordinate as part of a complete set, so simple attribute matches miss visually coherent combinations.

Resolution: Let the LLM reason over color harmony, stylistic coherence, material, decor style, and related attributes.

Catalog scale makes naive pairwise scoring too expensive.

Root cause: Hundreds of thousands of Home items would make item-level pair reasoning costly and slow.

Resolution: Constrain rules to item-type pairs, score/sort all pairs in those types, and parallelize the operation.

Model-only recommendations can be too narrow.

Root cause: Without human merchant insight, the system frequently returned similar items rather than broader accessory types.

Resolution: Add HITL merchant lists of commonly co-purchased accessories to diversify recommendations and support basket expansion.

Evaluation

Target evaluated the recommendation system with an A/B test in the add-to-cart flyout and tracked interaction rate, display-to-conversion rate, and attributable demand.

Metrics: interaction rate · display-to-conversion rate · attributable demand · recommendation diversity · basket expansion

Outcomes

• The Home Accessory model increased interaction rate by around 11%.
• It increased display-to-conversion rates by roughly 12%.
• It drove more than 9% growth in attributable demand.
• The model was fully rolled out to production in April 2025.
• The approach is expected to adapt beyond Home because it scales through item-type pair rules.

The system is explicitly production-facing and catalog-scale: it was built for Target's Home category with hundreds of thousands of items, A/B tested in February 2025 in the add-to-cart flyout, and fully rolled out to production in April 2025 after double-digit engagement and conversion lifts.