An online store can list thousands of products while providing surprisingly little useful information for recommendations. Generic titles, duplicated descriptions and important specifications buried in long paragraphs make it difficult to distinguish products that look similar but serve different needs.
If you plan to use artificial intelligence to help shoppers find suitable products, start by reviewing the catalog. Recommendation quality also depends on what the system can know about each item, which fields it receives and how those fields are kept current.
You do not need to fill every available field before getting started. First identify the information that helps shoppers choose and purchase products in each category. Then make that information consistent and verify that it reaches the integration correctly.
Start with a question a shopper would ask
Consider a shopper looking for a backpack to carry a 15.6 inch laptop to work. A listing that says only “Premium backpack” cannot establish the size of the laptop compartment, the bag's capacity or whether it suits that use.
A useful record includes facts you can verify: a 22 liter capacity, a compartment suitable for laptops up to 15.6 inches, external dimensions, material and a separate section for documents. Water resistance should match the actual specification. Do not describe a bag as waterproof simply because the phrase sounds persuasive.
Repeat this exercise for your main categories. What would a salesperson need to know to recommend a product responsibly? Which differences cause shoppers to accept or reject an option? Those answers point to the attributes that deserve attention first.
Make the title identify the item
A product title should help shoppers recognize the item without reading the entire listing. Depending on the category, it can include the brand, model, product type and a key distinguishing feature.
“22 L urban backpack with a 15.6 inch laptop compartment” conveys more information than “Amazing backpack special offer”. The description still matters for construction details, use and care, but it should not have to repair an ambiguous title on its own.
Keep frequently changing campaign language out of the product name. A stable title makes the item easier to check across screens and systems. Promotional messages and purchasing conditions can live in the appropriate offer fields.
Structure the attributes that determine suitability
When important information exists only in free text, each listing may express it differently. “22 liters”, “22L” and “twenty two liter capacity” describe the same value, but require normalization before reliable comparison.
Structured fields let you separate capacity, dimensions, material and compatibility. Define the units and attribute names for each category, then apply the same rules to new listings. Checking units prevents centimeters and millimeters from being compared as though they were equivalent.
Descriptions can add context around those fields. A paragraph explains how a compartment is useful; an attribute records its dimensions. Both representations can contribute, depending on how the recommendation solution processes catalog data.
Separate the product from its variants
Products offered in several colors or sizes require clear boundaries between shared information and variant specific information. A model description may apply to all variants, while price, availability, images and the selected option's identifier need to represent the corresponding variant whenever they differ.
A recommendation for a T shirt should not imply that medium is available when shoppers can purchase only small. Whether the solution recommends a parent product or an individual variant is an implementation choice. The destination page must support a purchase that matches what the recommendation presents.
Record the relationship between the parent product and its variants explicitly. Verify that the integration retains that relationship rather than combining distinct options into a single record that leaves the actual offer unclear.
Keep identity and offer information consistent
Identifiers connect catalog records with browsing events, inventory and orders. Use stable references and clear mappings between systems. A store SKU and a standardized identifier such as a GTIN, where one exists, serve different purposes. Do not invent one merely to fill the field intended for the other.
Price and availability must reflect the offer shown to the shopper. Yesterday's catalog snapshot may no longer be suitable for recommending an item that sold out today. Define how changes reach the recommendation system and what update delay your operation can accept.
When an update fails, use a defined response. Depending on the available capabilities, that could mean temporarily excluding records with stale data, querying the source again or alerting the team. An offer should not remain active indefinitely without a known last validation time.
Check which fields the solution actually uses
Recommendation systems do not all interpret descriptions, images and attributes in the same way. Some focus on browsing and purchasing behavior; others combine those signals with product information. An organized catalog does not mean every field is automatically used.
Before undertaking a large cleanup, verify the integration contract: accepted fields, variant handling, update frequency and eligibility rules. Establish which events link an impression, click and purchase to the correct product as well.
This prevents you from claiming that a system considers technical compatibility when it never receives compatibility data. Where compatibility determines whether an item is suitable, that rule needs to be represented and validated in the solution. Otherwise, additional filters or review may be necessary.
Run a small review you can evaluate
Choose an important category and review a sample of products. Include simple products, products with variants, closely related alternatives and items that are temporarily unavailable.
For each record, check:
- Does the title clearly identify the item?
- Are the attributes that influence selection present and expressed in consistent units?
- Do variants carry their own information where needed?
- Do the images, price and availability match the option offered for sale?
- Does the system receive the agreed fields while preserving stable identity?
- Does the recommendation lead to a page where the shopper can buy the item presented?
Then test realistic searches and browsing scenarios. Compare recommendations against criteria agreed with the commercial team and record unsuitable results. Improving catalog data can provide better inputs, but the effect must be observed in the actual system. It does not automatically guarantee additional sales.
Build the checks into everyday catalog work
Fixing a sample helps uncover the issue. To preserve those improvements, document title, attribute and variant rules, assign responsibility for validating new records and track incomplete listings.
Whenever an integration or recommendation strategy changes, check that the necessary information remains available. The catalog needs to reflect shoppers' questions and the rules your operation uses to present an offer.
AGTI can help assess your e-commerce catalog and integration flows, identifying the changes needed to use product data consistently in search and recommendations. Tell us which product decisions your shoppers find difficult to make.
Technical references
These references illustrate product attributes and structured representations. They do not define a universal AI integration contract or guarantee search visibility.
- Google Merchant API — ProductAttributes: https://developers.google.com/merchant/api/reference/rest/products_v1/ProductAttributes
- Google Search Central — Product and Offer structured data: https://developers.google.com/search/docs/appearance/structured-data/merchant-listing
- AGTI — services and contact: https://www.agti.eng.br/