How the CYield Keyword Model Works?
Brands and advertisers can decide to target specific keywords in addition to those automatically generated by the platform. For Search placements, we enable two types of keyword models to work in unison:
Manual keyword management model, and
Automated keyword model
Delivering ads on the automated keyword model targets the most relevant and best-performing keywords for SKUs, which includes top search values and long-tail values that the engine identifies for the best conversion and return.
The automated keyword model shows great relevancy of more than 92% (versus only 39% for a manual keyword model), enabling campaigns to have a better win rate (66% of the campaigns using the automated keyword model have more than a 70% winning rate) and ultimately a more robust performance.
To further improve matching accuracy, the automated keyword model now also reads product images to understand visual attributes that catalog text may not fully capture — including color, product type, brand recognition, and other details visible in the image. This improvement is particularly valuable for retailers whose product catalog data does not include complete attribute information in text form. No changes to campaign setup or retailer configuration are required; the model applies image signals automatically and selectively.
This level of relevancy is correlated to the algorithm constantly linking new keywords to SKUs based on the first-party data that the Criteo engine ingests from Search and Product Pages.
To manage the search terms added by the automatic keyword model, leverage the negative keyword tool to ensure not targeting certain keywords.
Keyword Matching
Automated Keyword Model
Built on 12 years of Criteo machine-learning and first-party data ingestion, the Criteo Automated Keyword Model ensures optimal ad delivery and minimal bid filtration by automatically targeting the most relevant and performing keywords. It includes both top-performing and less common keywords and misspelled terms.
The model now uses both product text and product images to understand what products look like — not just what their descriptions say. This is especially valuable when a retailer's catalog doesn't include complete visual attribute information in text form, such as color, product type, or other details visible only in an image.
For example, if a product has no color listed in its catalog entry, the model reads the product image to determine its color and match it to color-specific shopper searches. Similarly, the model can use images to distinguish between men's and women's products, or to recognize brand identifiers that appear in product packaging but not in the product title.
A selective mechanism ensures image signals are applied only when they add genuine value. If a product image is missing or low in quality, the model relies on text signals only.
Note: This enhancement applies to Sponsored Products only.
Manual Keyword Targeting
All the promoted products selected in line item(s) already have keywords automatically associated with them by our advanced keyword model (see above). In addition, it allows negating terms or submitting new ones that would not be covered by our model, thanks to the manual keyword targeting tool at line item level.
The real-time negative keyword feature ensures capturing only the wanted traffic by allowing negating specific broad or exact match terms.
The ability to submit manual keywords allows reaching additional keywords if they are approved by the retailer.
Commerce Yield is designed to give its users as much flexibility and control as possible. Users can choose how strict and specific they want their matching to be. Commerce Yield's Automated Keyword Model automatically associates over 100 keywords to line items for targeting based on the products the line item features. The line items will then target shoppers who use search terms that match the terms assigned to the line item for targeting.
Keyword matching can be set to either Broad or Exact matching for both negative and positive targeting.
Match Types:
Broad Match and Exact Match have very different impacts on the delivery of campaigns.
Broad match has a broader scope, blocking words that are similar or variations of input terms.
Exact match allows for more precise blocking. It only blocks search terms that match the inputs.
The image on the left shows how negative keyword matching may impact ad delivery and targeting.