Writing product descriptions with AI: from spreadsheet to store
Few online stores have this problem when they start out. With a short catalog, someone on the team writes the descriptions by hand, reviews them, uploads them, and the product page looks fine. The bottleneck comes later, when the supplier sends the new collection in an Excel file with codes, measurements and a technical name, and each product’s description ends up as the line the manufacturer wrote, or blank.
This guide explains how an SME that sells online can automate product description writing with AI, from the product spreadsheet to the live product page on WooCommerce, Shopify or PrestaShop. It also covers when it isn’t worth it and how much work it takes to set up.
Why product descriptions stall as the catalog grows
Writing a good product description takes time. You have to look at the product or its spec sheet, decide what matters to the buyer, turn the specifications into something people understand, and make sure the text doesn’t read too much like the one on the next product. With a catalog that grows every season, the task lands on someone who also handles orders and returns, and it gets pushed back until there’s time.
You can see the result in any store that has been open for a few years. Old, carefully written pages sit alongside new ones that only have the manufacturer’s copy, the same text every reseller of that brand uses. Google finds the same paragraph on lots of sites and has no reason to pick yours, and the customer torn between two models doesn’t find the answer and leaves.
What AI changes is the cost of the first draft, which drops to almost zero, and that shifts the work to two places where you really do need a person: preparing the input data properly and reviewing what comes out.
What AI can write on a product page and what it shouldn’t touch
A typical product page has a title, a short description, a long description, a list of features, the attributes (size, color, material, dimensions) and, underneath, the title tag and meta description that Google sees. Not all of those pieces are handled the same way.
AI writes the long and short descriptions well, as long as you give it the data. It turns a list of specs into sentences that explain what each thing is for, adapts the tone to your store if you describe it, and produces variants when you have near-identical products that need different copy.
What it must not do is make up data. A language model that’s missing the material of a case or the capacity of a battery tends to fill the gap with something plausible, and that’s the dangerous part of the process, because made-up data on a product page ends in a return or a complaint. That’s why technical attributes, prices, sizes, certifications and any promise about delivery times or warranties come from your database or the supplier’s sheet. The AI receives them as input and simply puts them into sentences.
Don’t ask it to decide what to highlight about a product only you know, either. If an item sells because it’s dishwasher-safe, that fact has to be in the spreadsheet, because the model won’t guess it.
Step by step: from the product spreadsheet to the live product page
The workflow we describe works for any of the three platforms SMEs usually use, and it revolves around a spreadsheet. The data goes into the spreadsheet, the model writes in the spreadsheet, the person reviews in the spreadsheet, and the spreadsheet is what goes back to the store at the end.
1. Export the catalog and flag which products need copy
WooCommerce, Shopify and PrestaShop let you export products to CSV from the admin panel. That export is your starting point: open it in Google Sheets or Excel and add a few control columns at the end. One to flag whether the product needs new copy, another for the row status (pending, generated, reviewed, approved), and the ones the model will fill in, one for each piece of the product page you want to generate.
Before generating anything, check what data each row has. If the supplier gives you a spec sheet with more information than your store has, now is the time to bring it into the same spreadsheet, in separate columns. Whatever the model doesn’t see in the row it can’t use, and whatever it uses without having seen it will be made up.

You don’t need to start with the whole catalog, either. One category of similar products is enough to fine-tune the instructions and see how much needs correcting before you run it on the rest.
2. Generate the descriptions in bulk with n8n, Make or a script
This is where the actual automation comes in. A workflow in n8n or Make reads the spreadsheet row by row, builds an instruction with each product’s data, sends it to a model like ChatGPT, Claude or Gemini through its API, and writes the response into the reserved columns. If someone on your team codes, a Python script does the same.
Quality depends almost entirely on the instruction, the prompt, which you write once and reuse for every row. A template that works reasonably well has these parts:
- who is writing and for whom, with the store’s tone and the type of customer who buys
- the product data, taken from the row’s columns
- what it has to return, with the length and format of each piece
- an explicit ban on adding features, materials or figures that aren’t in the data
- what to do if information is missing, which is to leave a flag for the reviewer instead of filling the gap
The last one is what saves the most work later. Written out in the prompt, it might look like this: “Use only the data below. If you need a piece of information to write the description and it isn’t there, write MISSING DATA followed by the name of the field, and don’t infer it.”
It also helps to give the model two or three of your own product descriptions that you like, as a tone example. Without examples, the copy tends to all sound the same, with the same catalog adjectives on every product.
3. Review each description in the same spreadsheet before publishing
Nothing goes to the store without being read by someone who knows the product. The review happens in the spreadsheet itself, column by column, and the reviewer changes the row status once they’re happy with it.
What you check goes in order of severity. First, that every fact in the description is on the spec sheet, especially dimensions, materials, compatibility and anything related to health, safety or warranty. Next, the MISSING DATA flags, which mean completing the spreadsheet and regenerating that row. And finally tone and repetition, because when you read lots of descriptions in a row you notice right away if the model opens them all with the same sentence.

If the review finds the same mistake in many rows, the fix goes in the prompt, not by hand row by row. Correcting the template and regenerating the category is usually faster than editing each text.
4. Upload the approved descriptions to WooCommerce, Shopify or PrestaShop
Approved rows go back to the store the same way they came out, with an import CSV. All three platforms can update existing products if the file keeps each product’s identifier (the ID, SKU or handle, depending on the platform), so nothing gets duplicated. The safe move is to test first with a file of just a few rows and check in the store that the text has landed in the right field.
If you’d rather not deal with CSVs, the same n8n or Make workflow can write to the store through its API when a row is marked approved. It’s more convenient for a catalog that changes often, and it takes more work to set up.
Product page SEO: keyword, title tag and meta description
For a product page to rank, it needs a search query in mind. For products, people search by combining the product type with an attribute, whether that’s material, use, size or compatibility, and those combinations are usually already in your columns. The prompt can ask the model to use the way people search for the product in the title and first paragraph, if you give it to the model. Don’t ask the model for keywords: get them from Ahrefs or Search Console and put them in a column.
The title tag and meta description are separate from the description. Shopify and PrestaShop have dedicated fields for them on each product, and in WooCommerce they’re added by an SEO plugin like Yoast or Rank Math. The title has to fit on the Google result line, so ask for a short length and check that the product name doesn’t get cut off. The meta description doesn’t help rankings on its own, but it’s what people read when deciding whether to click, so that’s where your clearest selling point goes.
There’s a risk specific to bulk generation: every page in a category ending up with near-identical titles. Add to the prompt that the title must include the attribute that sets that product apart from its siblings, and review the column sorted alphabetically, which is where repetition jumps out.
When automating product descriptions isn’t worth it
With a small catalog that rarely changes, building the workflow costs more than writing the descriptions. If you add a handful of products a month, one person with ChatGPT or Claude open in another tab, pasting in the data and a good template, gets the same result without automating anything. Automation starts to make sense when the volume makes that manual task take up hours every week, or when entire collections arrive at once.
It isn’t worth it for artisan products or high-value items either, where the copy is part of what’s being sold. A handmade ceramic piece, a wine from a specific vineyard plot or a piece of jewelry with a story sell with copy that only someone who knows the piece can write. There, AI helps with editing, and the text is still a person’s.
And it isn’t worth it if your source data is poor. If your catalog has no materials, dimensions or uses in any column, the model will write generic copy or make things up, and the review will eat up all the savings. In that case, the first job is to clean up your product data; the writing comes after.
How much it costs to set up and how much time goes into review
There are three routes, and each one splits the effort differently.
The first is an app from your platform’s app store. Shopify has several that generate AI descriptions right inside the admin, and there are similar plugins and modules for WooCommerce and PrestaShop. They install in an afternoon and are usually paid by subscription. They work well one product at a time and less well in bulk with your own rules, because the app controls the prompt.
The second is the custom workflow we’ve described, with a spreadsheet, n8n or Make, and a model’s API. Building it and fine-tuning the prompt on a first category is a few days’ work for someone who already uses these tools, and quite a bit more if they have to learn them from scratch. Usage is paid by text volume to the model’s API, and for an SME’s catalog it usually weighs little next to the reviewer’s time. In return, you control the tone, the data that goes in and what gets published.
The third is to outsource it: have someone build the workflow on your spreadsheet and your store and leave it up and running. You pay for the setup and skip the learning curve.
With any of the three, review is the part that doesn’t go away, and you should plan for it from the start. Reading a well-generated description and checking it against the spec sheet takes a few minutes. What drags it out is made-up data, which is why the time you invest in step 1 and in the prompt pays off here. We think the app is enough as long as you write one product at a time, and that a custom workflow pays off as soon as you generate in bulk or your catalog has rules no app knows. To settle it, the sensible thing is to generate one category, time how long it takes you to get it ready to publish, and compare that with what it used to take you to write it by hand.
If you want to build this workflow on your catalog and aren’t sure which of the three routes to start with, get in touch and we’ll look at it with your data in front of us.