Skip to content
Artificial Intelligence · 10 min read

AI for Your Product Catalog: Attributes, Descriptions and Translations

The catalog is the most boring part of an e-commerce store and also the one that pays off the most. Descriptions copied from the supplier, half-filled attributes, rushed translations, images with no alt text: I see them in almost every store I work on. Generative AI really helps here, but only if you use it as a production tool inside a process, not as a “generate everything” button.

In this article I tell you where AI pays off on the catalog, what the risks are (technical, SEO and regulatory) and the workflow I recommend to my clients on Magento and WooCommerce.

Why the catalog matters more than before

Until recently, a good product page served Google and the customer. Today AI assistants (Google AI Mode, Gemini, ChatGPT, Copilot) read it too, answering precise questions like “which 30-liter backpack is waterproof and weighs less than a kilo?”. If the weight is missing or the waterproofing is written only in an image, that product doesn’t exist for the agent. I go into more detail in the article on agentic commerce and AI-driven checkout: the conclusion is that structured, complete data is the prerequisite for everything.

Where AI really pays off on the catalog

TaskHow much it helpsWatch out for
Extracting attributes from spec sheets, PDFs, supplier descriptionsA lotMade-up values when the data is missing
Normalizing values (colors, sizes, materials)A lotMappings to validate on a sample
Suggesting the category and the Google taxonomyA lotSimilar categories mixed up
Rewriting descriptions starting from verified dataWellIdentical texts, untrue claims
Translations for foreign store viewsWellTechnical terminology, measurements, units
Image alt textWellGeneric descriptions
Generating or editing product imagesWith cautionProduct shown differently from the real one, labeling obligations

As you can see, the highest value isn’t “writing nice descriptions” but turning scattered data into structured attributes. A material = aluminum attribute is used by the site’s filters, the Merchant Center feed, structured data and AI agents. A poetic sentence about aluminum is used only by whoever reads it.

The risks to know before you start

Errors and “hallucinations”

If you ask a model for a complete description, it completes it even when it doesn’t have the data: it invents a certification, a compatibility, a warranty. For the customer that’s information from the seller, and a misleading description remains your responsibility under the Italian Consumer Code (Codice del Consumo), regardless of who wrote it. The same goes for the safety information that the GPSR regulation requires in online sales (manufacturer details, warnings): it must be taken from the sources, never generated.

Duplicate content and “scaled content”

Google doesn’t penalize AI in itself. Its guidance on generative content does warn, though, that generating lots of pages without added value can violate the scaled content abuse policy, and it asks you to check content by hand before publishing it. A thousand descriptions with the same pattern and the same adjectives, perhaps similar to those of competitors using the same tool, won’t set you apart.

Labeling: Merchant Center and the AI Act

  • Google Merchant Center has precise rules. AI-generated titles and descriptions must be sent in the structured_title and structured_description attributes, with digital_source_type set to trained_algorithmic_media. Generated images must keep the IPTC DigitalSourceType metadata (for example TrainedAlgorithmicMedia), which is often stripped by compression and CDNs.
  • AI Act, Article 50: the transparency obligations apply from 2 August 2026 (the Digital Omnibus postponed the rules on high-risk systems, not these). For those who use AI, the obligation to disclose it concerns deep fakes and texts published to inform the public on matters of public interest. A normal product description is unlikely to fall into the second case. A generated image showing a product “as if it were a real photo”, on the other hand, comes close to the definition of a deep fake. The European code of practice on labeling is still being drafted: for doubtful cases, talk to a lawyer.

My practical advice: use AI to retouch real photos (backgrounds, lighting, settings), not to invent how the product looks, and keep track of what has been generated.

The workflow I recommend

1. Start from the data, not the text

Before generating any description, define the attribute set for each product family. In Magento you do it in Stores → Attributes → Product and in the attribute sets. In WooCommerce you do it in Products → Attributes. Then use AI to extract those values from the sources you have (price lists, supplier sheets, PDFs), explicitly asking it to leave the field empty when the data isn’t there.

2. Work outside the admin, in batches

Magento Open Source has no built-in text generator, and that’s fine with me: I prefer to export a CSV, process it via API with a versioned prompt and re-import it from System → Data Transfer → Import. In WooCommerce there are many AI plugins, but the principle doesn’t change: drafts, not direct publishing. Keep a column with the status (generated, reviewed, published) and one that flags generated content: you’ll need it for the feed.

3. A prompt with clear rules

Use ONLY the data provided as input.
If a piece of information is not present, do not write it.
Do not mention certifications, warranties or compatibilities that are not listed.
Tone: [brand tone]. Length: 80-120 words.
JSON output: { "short_description", "description", "missing_attributes" }

The missing_attributes field is the most useful of all: it tells you where the catalog is thin and gives the people managing the products a to-do list.

4. Human review on a sample, then full review on critical products

Manually review a sample for each category until the error rate is acceptable. For products with legal or safety risks (food, cosmetics, electrical, toys) the review must be complete. Pay particular attention to numbers, units of measure and absolute claims (“the best”, “100% natural”).

5. Translations with a glossary

For foreign store views, give the model a glossary of technical terms and names that must not be translated (product lines, patented materials). Have a native speaker proofread at least the main categories. Remember that in Magento the translation must be saved in the scope of the right store view, otherwise you overwrite the default.

6. Publish and propagate to the feed

New attributes are of little use if they don’t reach the feed. Map them to Merchant Center (color, material, size, GTIN, product_type) and, for generated texts, use the structured attributes mentioned above. If you use Magento, I cover it in the guide to Google Merchant Center for Magento 2. In 2026 Google also announced new “conversational” attributes in Merchant Center designed for AI surfaces: this area is still evolving, so keep an eye on it.

7. Measure

Compare before and after on a group of categories: errors and disapprovals in Merchant Center, organic CTR, product page conversion, internal searches with no results. If a category doesn’t improve, the problem is usually in the starting data, not in the prompt.

In summary

  • Use AI first for attributes, then for texts.
  • The model works only on verified data and flags what’s missing.
  • Drafts in batches, human review, then publishing.
  • Generated texts and images must be declared to Merchant Center with the required attributes and metadata.
  • Never invent safety data, certifications or how the product looks.
  • Glossary for translations and the correct store view scope.
  • Measure by category and act on the data, not just on the prompt.

Done this way, a job that would take months by hand is finished in a few weeks and the catalog becomes an asset that serves the site, Google Shopping and AI assistants. If you want to set up a workflow like this on your Magento or WooCommerce store, get in touch.

You might also like