Recognize which words are a brand, product type or hard feature.
Entity recognition names the meaningful parts of a search query. The system not only sees loose words, but recognizes, for example, a brand, model code, size, material and product category. This structure enables more precise search and filtering.
What is an entity in online store search?
An entity is a recognizable part with a specific role. In “Nike running shoe women size 39” Nike, running shoe, women and size 39 are different entities. They should not only appear in the product text; they each have a different meaning for filtering and ranking.
By recognizing those roles, the search engine can apply a brand filter, select the right product type and treat size 39 as a hard condition. Without entity recognition, a result can happen to contain all the words, but still show a wrong brand, target group or size.
What entities are important?
- Brands and product lines, including writing variants and abbreviations.
- SKUs, EANs and model codes which should stay exactly.
- Product types and categories by determining the relevant catalogue space.
- Attributes such as color, material, size, power, content and connection.
- Applications and compatibility, for example, “for outdoor” or “suitable for model X”.
- Quantities and units, where 10 mm means something other than size 10.
The right set varies per assortment. Fashion has different core entities than electronics, building materials or industrial parts. The product taxonomy and real search questions determine which types are given priority.
Recognize and link
Recognizing a term is only the first step. After that, the term should be linked to the value in the product data. For example, “Dark Blue”, “navy” and “navy” may refer to the same normalized color value, while the original customer language is retained for explanation and analysis.
Brands and model codes require caution. An automatically corrected code may refer to another product. Exact catalogue values, well-known aliases and controlled synonyms are therefore more important than broad language associations.
Context prevents wrong labels
The same term can have multiple roles. “Orange” can be a color, brand or product name. “Pro” can be a serial designation, but also a general word. Entity recognition uses surrounding words, category context and catalog knowledge to determine the most likely role.
When in doubt, the system can take multiple interpretations with it and let the result decide later by available products, filters and confidence. This way, uncertain recognition does not immediately become a hard exclusion.
Product data as a foundation
Entity recognition becomes stronger when brands, categories, and attributes are in separate fields. If all information is only in one long description, it is more difficult to reliably link a recognized value and use it as a filter.
Therefore, check fixed units, consistent value lists, complete identifiers and a clear category tree. Deviating spellings can be mapped, but conflicting source data must be resolved at the source. The search layer should not invent an uncertain product fact.
Practical example
At “red leather women’s boot size 40” the search engine can recognize boot as a product type, red as color, leather as material, ladies as a target group and 40 as a size. Products that look semantic to boots but are not available in size 40 will fall off when stock per variant is reliably available.
If material data is missing from part of the assortment, the search engine can arrange explicitly marked leather products higher and treat unknown cases separately. That is more transparent than assuming that every appropriate description means learning.
Quality control
Measure recognition by entity type. Look at missed values, mislabeled terms, and erroneous links to catalog fields. Use real search questions, especially combinations with multiple attributes and terms that mean something different in different categories.
Findoviq uses this structure to control query understanding, constraints and hybrid retrieval more specifically. This protects model codes and hard properties, while making natural language usable for wider product discovery.
Discuss your search questions
Do you want to know how this approach fits your assortment, product data and customer behavior? Together, we look at which query types have priority and where exact, semantic and business signals need to complement each other.