What Data Does Product Catalog Automation Need?

The quality of product catalog automation begins with input data, not the writing model. If stable identifiers, price, or availability are missing, even fluent product descriptions cannot move safely into a sales channel. The right order is to organize product facts, generate content, and then validate the result against the destination channel.

Requirements differ by platform, country, category, and advertising program. Google Merchant Center and Shopify update their official documentation over time, so this article is a preparation framework rather than a replacement for the current specification of a destination channel.

Group 1: product identity

The product ID is the workflow’s primary key. Even when an external channel assigns another identifier, preserve a stable internal ID and mapping so that an update does not create a duplicate product.

Google’s product data specification changes the expectations for brand, GTIN, and MPN depending on whether a product has recognized unique identifiers. A workflow should not require the same identifier set for every product, and it should never manufacture a plausible-looking code when no source exists.

Group 2: customer-facing content and assets

Prepare a title, description, primary image URL or file, additional images, landing page, and a verifiable source for features and specifications. For consistent generated content, also provide:

• brand voice and prohibited wording;
• facts the workflow is allowed to use;
• target language and region;
• title, description, and bullet length limits;
• claims, effects, or comparisons that must not be made;
• image usage rights, purpose, and background requirements.

AI can adapt existing facts to channel formats. It should not invent material, certification, origin, performance, or warranty details. When a source does not confirm a fact, “missing information” is the correct result.

Group 3: price, inventory, and sales status

Common fields include currency, regular price, sale price and effective dates, availability, quantity, and product status. Price and inventory change frequently, so the record should include an update time and source system.

Automation can validate price formats, promotion windows, and conflicting states. Repricing and publishing have commercial consequences, however. A controlled workflow produces a diff and preview before an authorized person approves submission.

Group 4: variants and category attributes

Color, size, material, capacity, and bundle variants should share a parent relationship while maintaining their own SKU, price, availability, image, or GTIN when appropriate. Variant naming must remain stable or channel updates can create duplicate items.

Category attributes should come from the destination taxonomy and specification. Apparel and electronics require different data. Select a field template after identifying product type instead of forcing every attribute into one universal spreadsheet.

Group 5: logistics and compliance

Depending on the channel and region, a catalog may need weight, dimensions, shipping origin, service, handling time, return terms, tax category, age restrictions, dangerous-goods information, or energy-related data. These fields can affect cost, eligibility, and compliance and should never be guessed by a language model.

Local inventory programs may also require a store code, in-store price, and local availability. Google publishes a separate local product data specification for these cases.

A reviewable automation input package

A practical package can contain:

1. products.csv with one row per SKU or variant and factual fields;
2. brand-rules.md with voice, terminology, prohibited claims, and examples;
3. channel-mapping.csv mapping internal fields to destination fields;
4. an assets/ folder whose filenames correspond to product IDs;
5. approval-rules.md defining automatic transformations and required approvals;
6. test-sample.csv containing a small set of human-verified products.

Begin with 10 to 30 representative products, including normal rows, missing fields, and complicated variants. Compare import, generation, validation, and export results before increasing volume.

What the agent should validate

• Are required fields missing?
• Are identifiers, GTINs, and URLs syntactically valid?
• Are variant relationships consistent?
• Do price, currency, and availability conflict?
• Do titles and descriptions exceed channel limits?
• Does generated content use only approved facts?
• Which fields came directly from source data, and which were transformed?
• Which items require human confirmation?

The output should include both results and an issue list rather than only a “completed” message. That gives reviewers a direct route to the products requiring more information.

The boundary of DoPilot’s ecommerce example

The DoPilot Ecommerce AI Agent page presents a workflow design beginning with product sheets and brand rules and continuing through catalog content, multilingual variants, field checks, and operating reports. It is currently a product demo design. It does not claim completed connections to every ecommerce platform, and it does not present publishing, repricing, refunds, or bulk messaging as unattended actions.

Product catalog automation starts with a sourced, mappable, and reviewable product table. The clearer the field preparation, the more reliable generation, validation, and human approval can become.