AI agents
AI product descriptions and catalogue enrichment, step by step
This catalogue enrichment playbook helps ecommerce teams write AI product descriptions for hundreds of pages. An AI agent drafts titles, descriptions, attributes and alt text, and a person checks them. You export the catalogue, write the rules, review a sample, publish in batches and measure. AI for Marketing Ops can run it for you with our AI workflows and agents service.
Updated:
- Who it is for
- Ecommerce teams with hundreds of product pages that have short, copied or missing text.
- Time
- About 2 weeks for the first batch, from export to import. Search Console results need 4 to 5 more weeks.
- Outcome
- Product titles, descriptions, attributes and alt text that follow one set of rules, checked by a person and published in batches.
- Tools
- Your store’s CSV export and import · Claude Code · A spreadsheet · Google Search Console
Steps
- 01 Day 1
Catalogue export and first batch
Start with a full export of your products, so you work on a copy and keep the original. Every platform in this playbook has a documented export, so you need no plugin.
Shopify, WooCommerce and Magento export products to CSV from the admin. In PrestaShop, the SQL Manager runs a saved SELECT query and downloads the result as CSV. You choose the exact fields, such as descriptions.
What to do
- Export the products. Shopify: Products > Export. WooCommerce: Products > All Products > Export. Magento: System > Data Transfer > Export, entity type Products. PrestaShop: Advanced Parameters > Database > SQL Manager.
- Save one copy of the export that nobody edits.
- Choose about 100 products from one category for the first batch.
- Pick the category with the most empty fields or the most clicks in Search Console.
- Note the column your import will match on: handle, ID, SKU or product reference.
When it is done: A backup file and a batch file of about 100 products.
- 02 Day 2
Rules for the agent
Write one page of rules before the agent writes a word. The rules decide the quality of every draft, so this is the step to spend time on.
Save the rules as a CLAUDE.md file in the project folder. Claude Code reads that file at the start of every session, so each batch follows the same rules.
What to do
- List the fields to write: title, description, attributes, alt text, SEO title and meta description.
- Set a length for each field. Shopify allows SEO titles up to 70 characters and SEO descriptions up to 320.
- Allow only facts from the export or from supplier sheets. No invented sizes, materials or claims.
- Add your tone, the words you never use and 3 good product pages as examples.
- Write separate rules for each language you sell in.
When it is done: One page of rules, saved as CLAUDE.md next to the batch file.
- 03 Day 3
Facts for each product
Give the agent the facts it can use, and nothing else. AI models predict likely words, so they can invent a size or a material when a fact is missing. Google gives the same warning in its guidance on AI content.
Put the facts in columns, such as brand, material, size, colour, use and care. Leave a cell empty when you do not know the fact. The agent flags empty cells and never guesses.
What to do
- Copy the attributes you already have into fact columns.
- Add facts from supplier sheets and product labels.
- Mark products with no reliable facts and keep them out of the batch.
- Ask the product owner to confirm the facts the team is not sure about.
When it is done: A batch file where every product has a row of checked facts.
- 04 Day 4
Agent drafts of the new fields
Run the agent on the batch file. Claude Code reads and edits files in a folder on your computer. It writes each draft into a new column next to the original text.
Keep the original columns unchanged. Your reviewer compares old and new text side by side. You can return to the old text at any time.
What to do
- Ask the agent for a title, description, attributes and alt text for every row.
- Ask for a note on each product: the facts it used and the facts that were missing.
- Ask for alt text that says what the image shows, without a list of keywords.
- Run 10 products first and read them before you run the full batch.
When it is done: A draft file with a new column for every field and a note per product.
- 05 Day 5
Sample review and rule fixes
A person reads the drafts before anything goes live. Google asks site owners to check AI content for accuracy before they publish it. This includes titles, meta descriptions and alt text.
In the first batch, read every draft. In later batches, read 20 products and every product the agent flagged. When you find the same mistake twice, change the rules and run the agent again.
What to do
- Check every fact in the draft against the fact columns.
- Check each length against your rules and your platform limits.
- Mark each draft as approved, changed or rejected.
- Add each repeated mistake to the rules file as a new rule.
When it is done: A review sheet with a status for each product and an updated rules file.
- 06 Week 2
Publishing in batches
Import the approved drafts in batches with your platform’s own import. A batch with a name and a date is a batch you can measure and undo.
Test the import with 5 products first. Then open those 5 pages in the store and read them before you import the rest.
What to do
- Shopify: import the CSV with “Overwrite products with matching handles” selected.
- WooCommerce: import with “Update existing products”. It matches rows by ID or SKU.
- Magento: System > Data Transfer > Import with Add/Update, then Check Data. Never choose Replace, because it clears the existing product data.
- PrestaShop: Advanced Parameters > Import, one file per language. Never select the option that deletes all products before the import.
- Log the date and the product IDs of each batch.
When it is done: One batch live in the store, with a log of what changed and when.
- 07 Week 7
Measurement and the next batch
Compare each batch with its own past. The Search Console Performance report groups clicks and impressions by page. You can compare the edited pages before and after the import.
We judge a batch after 4 weeks or more, never in the first days. The newest data in Search Console is preliminary and can still change.
What to do
- Export clicks and impressions for the batch pages, 28 days before and 28 days after the import.
- Count the empty fields left in a new catalogue export.
- Count the hours the team spent on the batch.
- Choose the next category and use the same rules file.
When it is done: A before and after table for the batch, and the next batch in the plan.
Mistakes to avoid
- Letting the agent write without fact columns. It fills the gaps with likely words, and a wrong size or material reaches your customers.
- Publishing hundreds of unreviewed drafts in one import. One wrong rule then reaches every page, and you cannot measure what changed. Google says many AI pages that add no value for users can break its policy on scaled content abuse.
- Choosing Replace in Magento, or the delete option in the PrestaShop import. Both remove existing product data.
- Filling alt text with keywords. Google says this gives users a bad experience and can make the site look like spam.
- Sending AI titles and descriptions to Google Merchant Center without a label. Merchant Center asks you to mark product data made with AI.
What to measure
| Metric | How to measure | When |
|---|---|---|
| Drafts approved without changes | Count the statuses in the review sheet | After each batch |
| Wrong facts found in review | Count the drafts rejected for a wrong fact | After each batch |
| Empty product fields | Count the blank cells in a new catalogue export | Before the first batch, then once a month |
| Clicks and impressions on batch pages | Search Console Performance report by page, 28 days before and after | 4 weeks after each import |
| Team hours per batch | A time log kept by the reviewer | After each batch |
Rather have us run it
We run this playbook with you or for you
Sources
- Google Search Central: Using generative AI content
- Google Search Central: Image SEO best practices
- Google Search Console Help: Performance report
- Shopify Help Center: Using CSV files for products
- WooCommerce: Product CSV importer and exporter
- Adobe Commerce: Export data
- Adobe Commerce: Import data
- PrestaShop 8 user guide: Import
- PrestaShop 8 user guide: SQL Manager
- Claude Code overview (Anthropic)
Questions
Common questions
Can AI write product descriptions for my store?
Yes, when a person checks the drafts. Google accepts AI content that is accurate and useful, and asks you to check it before you publish. The main risk is invented facts. For that reason, the agent works only from the facts in your export and flags what is missing.
Does this work on PrestaShop, WooCommerce, Magento and Shopify?
Yes. All four platforms have a documented way to export products to CSV and import them again, with no plugin. The menu paths and the match column differ. Shopify uses the handle, WooCommerce the ID or SKU, and Magento the SKU. PrestaShop uses the product ID or the reference.
How many products should one batch have?
We start with about 100 products from one category. That is enough to find the mistakes that repeat in the rules. It is also small enough for one reviewer to read every draft. Later batches grow when the rules hold and the review finds few mistakes.
Can AI for Marketing Ops run this for us?
Yes. Our AI workflows and agents service builds the catalogue agent on your store data, as one project. Most projects go live in 10 working days. You approve every change before it goes live. On a monthly plan, we host the agent, monitor it and improve it every month.
Book a call about your catalogue
On a free 30-minute call, we plan the first batch with you. Before the call, our agents check the top 10 SEO problems on your product pages. A person reviews the check. You keep the short plan, even if we do not work together.
Book a discovery call