Bulk product descriptions: drafted from your product data, published after editor approval
Bulk product description drafting is a workflow that uses AI to draft product descriptions and category copy in batches from the attribute data in a product list, checks each text against that data and makes publication depend on an editor's approval.
Short answer
Virelth's bulk product description solution drafts product descriptions and category copy from the attribute data in your product list. The model uses only the facts in each product's row, and every measurement and number is checked against that data. Drafts land in an approval queue; nothing goes live in your store until your team or editor approves it.
Key points
- Drafts rely only on your product data; a feature that is not in the data never makes it into the text.
- Every measurement, number and material in the text is compared with the product row automatically, and mismatches are flagged.
- A small sample batch is written and approved first; bulk drafting only starts once the rules are settled.
- Google asks for AI-generated content to be checked by hand before publishing, which is why the approval queue is mandatory.
- AI-written text sent to Google Merchant Center is labeled in the separate attributes Google requires.
What problem does it solve?
In a store with hundreds of products, the description is either the supplier's text or missing altogether. Supplier text is identical to what every other retailer of the same product uses, and it rarely answers the customer's real questions about size, material and use. Writing original copy for every product by hand takes the team weeks, and the job starts over when the new season arrives. Category pages are often just a product grid. This solution drafts from your product data, so the editor's job changes from writing to checking.
What does the pilot measure?
We do not promise rankings, traffic or sales. The pilot measures the quality of the drafts and the editor's workload.
Drafts approved unchanged
The share of drafts the editor approved without edits, reported per category.
Data mismatches caught
Drafts the automatic check flagged because a measurement, number or material contradicted the product data.
Editor time per product
The time the editor needs to check and approve a draft, compared with the time it took to write copy by hand before the pilot.
Missing data report
Products that could not be drafted and the fields they lack, showing which parts of your product data need completing.
How does the workflow run?
Product data comes in
At the start
The product list arrives as a spreadsheet (CSV) exported from your store or product feed: name, category, attributes, dimensions, material, use and care. Missing fields are reported.
Approval: Your team confirms which field is the source of truth for each fact.
Writing rules are set
Tone, length, structure (short intro, feature list, use and care) and banned phrases are written per category. Phrases such as 'best', guarantees or health claims go on the banned list.
Sample batch
A small sample from each main category is drafted. Your team reads and corrects it, and the rules are updated from those corrections.
Approval: The person who approves the sample batch decides whether bulk drafting starts.
Bulk drafting
Products are processed in batches. For each product the model sees only that product's row, never another product's data or anything from the web.
Automatic checks
Every number and measurement in the text is compared with the product row, banned phrases are searched for and near-identical texts are flagged. Problem drafts land in the queue marked 'needs a manual look'.
Editor approval
Drafts are read in an approval queue, product by product or per category. The editor approves, edits or sends them back with a note.
Approval: No text is exported to the store without approval.
Publishing
Approved texts are written to your store by bulk import or a connection. In a Google Merchant Center feed, AI-generated text is labeled in its own attributes.
Approval: Your team starts the import, trying a small group first.
Where a person approves
- Which field in the product data is the source of truth.
- The sample batch, and the decision to start bulk drafting.
- Every product description and category text, before publishing.
- Importing into the store and sending to the product feed.
- Every change to the writing rules and the banned-phrase list.
Why does a draft rely only on product data?
A language model can fluently invent something it does not know about a product: “waterproof”, “machine washable”, “lasts for years”. On a product page that is simply wrong information, and it comes back as returns, complaints and lost trust. So the rule is strict. For each product the model sees only that product’s row. Every number and measurement in the text is checked against the value in the row, and a mismatch flags the draft. A feature that is not in the data does not get into the text.
An example (representative data, not a real product):
| Product data | In the draft |
|---|---|
| Material: 100% cotton | “Woven entirely from cotton, …” |
| Size: 50 x 70 cm | “At 50 x 70 cm it fits standard pillows.” |
| Washing: 40 °C | “Machine washable at 40 °C.” |
| (not in the data) | No claims about stain resistance or colorfastness. |
What separates scaled content from useful content?
Google looks at whether content is helpful more than at how it was produced, but it can treat generating many pages without added value as a spam policy violation. A thousand different sentences for a thousand products are not value in themselves. The value lies in information that answers the customer’s question: size, material, compatibility, use and care. That is why writing rules are set per category. For a pillowcase, size and washing come first; for a kitchen appliance, power and cleaning. Drafts that are too similar to each other are flagged automatically.
Which intent your product and category pages should answer in search, how filter pages are handled and how structured product data is set up are planned together as part of e-commerce SEO. If you are thinking about building or rebuilding the store itself, see our e-commerce websites page.
What is it built with?
Workflow automation
Importing product data, drafting in batches, comparing with the data, the approval queue and exporting approved texts to the store.
E-commerce SEO
Which information should lead on product and category pages for each search intent, plus titles and structured product data.
Content marketing
Writing the tone rules and, if you like, having our editor handle the approval step.
Data and privacy
Data processed
Product names, categories and attribute data. This is usually not personal data; personal fields such as supplier contact details are removed from the sheet.
Commercial data
Purchase prices, stock levels and supplier terms are not sent to the model; only the fields needed for the published text are.
Supplier text
Supplier copy is read only as a source of product facts and is not copied into the draft. You know the scope of your rights to supplier content; if it is unclear, we leave that field out.
Logs
Each draft is stored with the rules version, the check result and the approver, so you can always trace why a text reads the way it does.
Provider and location
The AI provider, the country where data is processed and its retention terms are stated in writing before setup.
Who is it for, and who is it not for?
A good fit
- Stores with hundreds or thousands of products whose descriptions match the supplier's text
- Brands that add many new products every season
- Cross-border sellers who sell in several languages
- Catalogs with clean product data but no copy
Not a good fit
- Stores with a few dozen products; writing by hand gives better results.
- Products whose claims are regulated, such as supplements, cosmetics or medical devices, when no expert reviewer is involved
- Technical products where the manufacturer's text must be kept word for word
- Anyone who wants texts published without approval
Pilot plan
Data
The product list for one or two categories, as an exported sheet or a product feed.
Tool
Drafts are written into a shared approval sheet; nothing is exported to the store yet.
Owners
On your side, someone who knows the products and an editor; on ours, the person who runs the rules and checks.
Duration and metrics
A limited period stated in the quote; share approved unchanged, mismatches caught and editor time per product.
What determines the price?
- Number of products and categories
- Number of languages
- How clean the product data is and which fields need completing
- Whether categories need their own writing rules
- How texts get into your store and product feed
- Whether approval stays with your team or our editor
Price: From $890 per project + VAT (setup; automation platform licence and server not included)
The exact scope and price are set in a written quote after the free preliminary analysis.
Prices exclude VAT. The written quote states the exact amount; invoices are issued in Turkish lira (TRY) at the exchange rate of the invoice date.
Frequently asked questions
According to Google's guidance, using AI is not a problem in itself; content still has to meet Search Essentials and the spam policies. Generating many pages without adding value can count as scaled content abuse. Google also asks for AI content to be fact-checked by hand before publishing, which is exactly what the approval queue is for.
Yes. Merchant Center asks for product titles created with generative AI to be sent in the structured_title attribute and descriptions in structured_description, with the value trained_algorithmic_media. If your feed goes to Merchant Center, we label approved texts in those attributes.
The data gets tidied first. Because a draft relies only on what is in the product row, a missing measurement or material stays missing in the text; the model does not fill gaps with guesses. In the pilot we produce a missing-data report showing which fields need completing.
Most common platforms, such as WooCommerce or Shopify, support exporting and importing products as a spreadsheet, and the solution can work that way. A direct connection or export to a marketplace dashboard depends on the platform, and we check this in the free preliminary analysis.
Yes. Each language is drafted separately with its own writing rules rather than translated from one source text. For any language your team does not read, we require a check by an editor who does; we never treat a machine draft as approved copy.
Yes. Category copy is drafted from what the products in that category have in common and from the buying tips you provide, such as which size suits whom or what each material is good for. Category texts go through the same approval as product descriptions.
Your request has reached us
We will review your request and get back to you through the channel you chose.
We will review it and get back to you.
Your request could not be sent. Please check your details and try again, or write to us at info@virelth.com.