AI-Ready Product Data: How to Prepare Your Catalogue

AI-ready product data gives an assistant complete, consistent facts about each product, including who it suits and how customers can buy it. Prepare your catalogue by filling missing attributes, keeping purchase information current and testing answers against the source records. A polished product page can still leave an assistant without the measurement a shopper needs.

What Makes Product Data AI-Ready

AI-ready product data is complete enough for shopper questions, structured into named fields and consistent across every place the product appears. It also explains suitability in plain language. A dining table record needs its dimensions and seating capacity, plus any limits on outdoor use.

Data written for browsing often relies on pictures or visual context. A size chart inside an image helps a person looking at the page. Copy the measurements into text fields too, so an assistant receiving only catalogue text can use them.

Judge data readiness for AI product by product, then compare categories. Clothing needs garment measurements and care instructions. Electronics needs model compatibility and power requirements. A complete dress record tells you little about the quality of your charger records.

How Large Language Models Read Product Data

Large language models use the product information supplied to them through search results, catalogue records or page content to compose an answer. A shopping system can retrieve text fields and named attributes, then give those records to the model. Structured markup supplies another source of labelled facts on a website.

An assistant’s access depends on its connection to the store. Some systems can analyse images, but a text-only feed won’t carry a measurement printed inside a photo. Write factual specifications as text even when images also show them.

Google Search accepts product data through page markup and Merchant Center feeds. Its guidance on catalogue data explains that using both helps it understand and verify the information. Accurate data must reach the system answering the question, as well as appear on your page.

Consider this hypothetical furniture record. A shopper asks whether a cabinet fits into a space 80 cm wide, but the record contains only its height. An assistant might guess from a photo or leave the cabinet out of its recommendations. Add a verified width of 76 cm and the relevant clearance instructions. The record now supports a width comparison without inventing the missing measurement.

Fields AI Shopping Agents Need

Supply shopping agents with fields that answer buying questions at the exact product and variant level. Record dimensions with units, material composition, compatibility and care instructions. Clothing also needs a sizing system and garment measurements. Keep stock and delivery details tied to the item the customer can actually order.

Explain use cases in ordinary sentences. A tag reading “outdoor” doesn’t explain whether a chair tolerates rain or needs covered storage. Write the supported use and its limits from the manufacturer’s instructions. For category-specific examples, use the guide to product attributes.

Google’s Merchant Center specification separates variant details from price and availability. It also recommends consistent units across product dimensions. Follow the same discipline in your own catalogue, while choosing additional fields around your shoppers’ questions.

Connect the same records to your on-site assistant. shoppa helps customers find products and compare their options using your catalogue feed and store documents. Customers also buy products in the chat. Its answers use your live product information and policies, so missing measurements or vague delivery terms can surface directly in a reply. Keep the source facts complete before using them to guide a purchase.

Product Data Checklist and Catalogue Completeness

Use this product data checklist to decide what each category must contain before an assistant can answer its common shopper questions. The priorities below are a suggested internal standard for AI answers, not a universal feed-submission specification. Required means the record needs that field for the questions your category must answer.

Field group	Field	Priority	What to record
Identity	Product identifier and page	Required	Stable identifier and the page for the exact item
Identity	Name and variant	Required	Product name, model and each purchasable size or colour
Purchase	Price and currency	Required	Current variant price and its currency
Purchase	Stock	Required	Availability for the exact variant, including preorder status
Purchase	Delivery rule	Required	Applicable shipping service, cost rule and delivery estimate
Purchase	Returns rule	Required	Applicable return conditions and product exceptions
Physical details	Dimensions and units	Required where fit depends on size	Product measurements, separate from packed dimensions
Physical details	Materials or ingredients	Required where they decide suitability	Verified composition and any relevant product warnings
Physical details	Size system and measurements	Required for sized products	Size label, sizing system and the measurements it corresponds to
Compatibility	Supported models or standards	Required for compatibility-dependent products	Exact supported models and connection type, with exclusions stated separately
Use	Suitability and use cases	Recommended	Who or what the product suits, with supported limits
Use	Care and maintenance	Recommended	Washing, cleaning or upkeep instructions
Use	Included items	Recommended	What comes in the box and what must be bought separately
Evidence	Warranty and instructions	Recommended	Applicable warranty terms and accessible instructions
Evidence	Reviews	Optional	Genuine product reviews with their source
Media	Additional images	Optional	Other views that help explain appearance or use

Promote a recommended field to required when shoppers need it to decide safely or correctly. Care instructions might be required for washable clothing. Mark genuinely irrelevant fields as not applicable, with a reason. An unknown compatibility value is missing data, not a reason to remove the field from the score.

Calculate each product’s data completeness as filled applicable required fields divided by all applicable required fields, multiplied by 100. Count a field as filled only when its value is specific, current and supported. Generic text such as “standard size” doesn’t complete a measurement field.

For a hypothetical product with ten required fields, eight complete fields produce an 80% score. If its width is missing, that score still cannot support a width question. Keep the missing-field list beside the percentage.

Report catalogue completeness by category using the average product score, the lowest score and the share of products with every required field complete. Fix the weakest category first, starting with fields that cause wrong answers. Comparing categories prevents a large, tidy clothing range from hiding incomplete appliance records. The guide to filling catalogue gaps explains how to fill repeated gaps in bulk.

Make Shipping and Returns Details Machine-Readable

Write shipping and returns rules as current text with explicit conditions, so an assistant can apply them to a product and destination. Separate dispatch time from carrier transit time. State whether an estimate uses business days and when oversized items or remote destinations follow a different rule.

A hypothetical policy saying “fast delivery Australia-wide” cannot answer whether a cabinet will arrive within five business days. A useful record states its dispatch range, destination-based transit estimate and any order cutoff. An estimate must remain an estimate in the assistant’s answer.

Record delivery costs or their calculation rule, including postcode restrictions and product-specific charges. Keep the source rule current when services change. Where the final charge depends on the basket and destination, the assistant needs those inputs before stating a total.

For returns, distinguish change-of-mind terms from faulty-product handling. State the return window, item-condition requirements, refund method and who pays return shipping. Identify any product exceptions. Keep policy wording consistent across your documents and product records, so retrieval doesn’t supply two conflicting answers.

A Readiness Test You Can Run This Week

Test readiness by asking ten real shopper questions about five products, then comparing all 50 answers with your verified records. Choose products from different categories, including a variant with limited stock and an item with special delivery terms. You need the current catalogue export, store policies and access to the assistant you want to assess.

1. Choose the five products and write the expected answers from their records. Use real questions from customer conversations where available. Add category-specific questions for measurements or compatibility.
2. Ask each question in a fresh conversation. For an on-site agent, use its private testing area if one is available. For ChatGPT or Gemini, give the exact product name and page address, then record which product source the answer uses.
3. Save the prompt, answer and source alongside the expected answer. Mark each answer correct, wrong, vague or unable to answer. Record the product variant and destination used in the question.
4. Assign each failure a specific fix and owner. A missing width goes to the product-data owner. An old shipping rule goes to the policy owner. Correct source data with a wrong reply goes to the assistant’s operator for retrieval or refresh checks.
5. Refresh the connected data and repeat the failed questions with the same product and destination. Confirm that each answer matches the corrected record before expanding the test to the rest of that category.

Use these ten question patterns, adapted to the product you’re testing. They are a test template, not reported assistant results.

• What is the current price of the exact variant I’ve selected?
• Is that variant in stock or on preorder?
• What are the product’s dimensions, and which units are used?
• What is it made from?
• Will it work with my specified model or existing equipment?
• Which size matches the measurements I’ve given?
• How do I clean or care for it?
• What comes with it, and what else do I need?
• What will delivery to my postcode cost, and how long is it estimated to take?
• Can I return it if I change my mind, and which conditions apply?

For categories where a question doesn’t apply, replace it with a real buying question of equal importance. Save enough context to repeat the question without changing the test. A public assistant’s old answer can reflect an outdated retrieved page even when your current catalogue is correct.

For a hypothetical failed answer, log “cabinet width guessed” against the product identifier. Assign the verified width field to your catalogue manager, then repeat the original space-fit question after refresh. The expected result is an answer using the recorded width and clearance requirements, or an explicit inability to confirm fit when a required fact remains absent.

Start with the lowest-scoring category and repeat this process until its buying questions have supported answers. For the broader steps around AI-led purchases, read the guide to preparing for agentic commerce.