Search matching and ranking
Merchandising Services with the composable catalog data model (CCDM) ranks results so shoppers see the most relevant products first. The service boosts products whose catalog text closely matches the shopper’s query, then favors matches where terms appear together, and finally includes broader matches.
How matches are prioritized
Relevance uses three layers of matching strength (in addition to other scoring factors described below):
-
Exact and near phrase match — The full search phrase matches catalog text, or a near match after normalization such as stemming (for example, singular and plural forms resolve to the same root). These matches receive the highest relevance boost.
-
All words in the same field — Every query word appears in one searchable attribute (for example,
redandpantsin the product name). This layer receives the next highest boost. -
Words across different fields — Query terms appear in different searchable attributes (for example,
redin color andpantsin name). This match layer is the broadest and receives the lowest relevance boost. It can also match partial queries used by autocomplete—for example, when a shopper typesred panbefore finishingpants. For German catalogs, see Decompounding (German).
Example
For a query such as red pants:
- Products with the exact phrase red pants (or a close variant) rank first.
- Products where red and pants appear in the same field (for example, name) rank next.
- Products where the terms appear in different fields (for example, color and name) follow.
Decompounding (German) decompounding-german
German catalogs use many compound words. For example, spülbecken and spül becken can decompose into tokens such as spul and beck (after stemming) so a shopper who searches spul becken can still find Spülbecken. At this layer, decompounded subwords from a compound term must appear in the same field. Other query terms can match in different fields.
This AND requirement filters irrelevant matches where only one subword is present. For example, a search for Brauseschlauch no longer returns Schlauchstück when only part of the compound matches. A search for spülbecken can still match spülbeckventil because the longer word contains all expected tokens.
Example
For a search phrase like Brauseschlauch chrom:
- Exact and near phrase match — Looks for the full phrase brauseschlauch chrom as typed, without decompounding (stemming still applies).
- All words in the same field — Looks for brauseschlauch and chrom in the same searchable attribute, still without decompounding (for example, both in name).
- Words across different fields — Decompounds Brauseschlauch into brause and schlauch. Those tokens must appear in the same field (not necessarily as an adjacent phrase). chrom can match in a different field (for example, brause and schlauch in name, chrom in color).
Set Language to German on the Language tab in Settings so decompounding rules apply. Validate high-value German queries on a staging storefront before you enable changes in production.
Decompounding is rule-based and can add edge cases at this layer. If a subword is missing from the dictionary, tokenization is incomplete and returns broader matches than expected—for example, gas missing from gaszähler emits only zahl, or stat missing from thermostat. The stemmer can also produce unexpected roots (for example, schrauber stemming to schraub, or schelle to schell). Adobe updates the dictionary and stemming overrides for known cases as issues are identified.
What else affects ranking
Phrase matching alone does not determine relevance. Several signals interact:
- Boost from exact / near phrase matching
- Boost when all query terms appear in the same field
- Intelligent ranking (when enabled), which blends textual relevance with behavioral signals — see How intelligent ranking scoring works
- Search weight on each attribute and other textual relevance factors (for example, how often terms occur in name or description length). For Adobe Commerce deployments, configure searchable attributes and their weights in attribute Storefront Properties. For stand-alone Adobe Commerce Optimizer deployments, use the Metadata REST API.
- Merchandising rules such as pin, boost, and bury
Because signals interact, a broad match sometimes ranks above a tighter phrase match when search weights or term frequency outweigh a weaker phrase match.
Example: If red pants appears in description with search weight = 1, but red and pants appear separately in name and color with search weight = 10, the split match outranks the phrase match.
Manual pin and bury rules remain strong; boost rules may require tuning to overcome new phrase and same-field boosts. Validate important queries after changing weights or rules.
Search weight 1 and combined indexing
Attributes with minimum search weight (weight 1) and no special match modes combine into a single internal field (defaultSearchField) to reduce mapping overhead. Treat that as one searchable surface for same-field matching: tokens that land only in those combined low-weight fields are evaluated together rather than as separate per-attribute fields. Adobe may refine this optimization over time as matching evolves.