Ranking blocks

Default algorithmic signals

In addition to manual or variable-based blocks, AB Tasty Recommendations & Merchandising provides default algorithmic logic for several ranking blocks, for example Associated Products and Similar Products.

These are powered by three main signals:

• Co-occurrence

• TF-IDF

• Semantic

Co-occurrence

• Definition: Co-occurrence measures how often two products appear together in the same context (same basket, same purchase, or same browsing session).

• Logic: The more frequently two products are seen or bought together, the stronger their association score.

• Example use cases:

  • Associated products (“Frequently bought together”): Phone + Protective Case

  • Complementary products: Printer + Ink Cartridges

TF-IDF (Term Frequency – Inverse Document Frequency)

• Definition: A weighting method that adjusts the co-occurrence score to reduce the bias of overly generic products.

• Why it matters: Popular products (gift cards, batteries, etc.) tend to appear in many baskets and could dominate co-occurrence. TF-IDF down-weights such generic items and highlights more relevant, less obvious associations.

• Example use cases:

  • Associated products with relevance: Instead of recommending a gift card with every book, TF-IDF will surface the next tome in the same series.

  • Similar products: Boosts underexposed or niche products, giving visibility to new arrivals or long-tail items.

Semantic (on demand)

• Definition: Semantic similarity measures how closely products relate to each other based on their descriptions, attributes, and catalog metadata, rather than just co-viewing or co-purchasing patterns.

• Logic: By analyzing product text and attributes, semantic algorithms identify products that “feel alike” in style, function, or category.

• Example use cases:

  • Similar products on PDP: A customer viewing a red running shoe could be shown other running shoes of similar style or color.

  • Alternative discovery: Highlight stylistically close items when co-occurrence data is too sparse (e.g., for new products).

Semantic Algo – Multilingual The embedding model is trained on 109 languages, including those without spaces between words (e.g. Japanese, Chinese, Thai). This differs from the Content Interest Criterion (segment builder), which does not support such languages.

How AB Tasty uses them

• Both the “Associated Products” and “Similar Products” blocks sort results using:

  • Co-occurrence (frequency)

  • TF-IDF (relevance adjustment)

• Semantic similarity can be activated on demand to enrich Similar Products recommendations and improve discovery when behavioral data is limited.