Open the outerwear category on almost any mid-sized retailer’s site in late October and read five parka descriptions in a row. “Stay warm and stylish all winter long.” “Crafted for comfort and designed for adventure.” “The ultimate layer for cold-weather days.” By the third one your eyes slide off the page, and by the fifth you can’t remember which jacket had the removable hood. Nobody wrote those sentences on purpose. A tool generated them in bulk, someone skimmed them for typos, and they went live alongside 300 others the same afternoon.
I’ve spent a fair amount of time cleaning up catalogue copy for retail teams, and the pattern is very consistent. The AI didn’t fail. It did exactly what it was asked, which was “write a product description for a women’s insulated parka.” The prompt had no facts in it, so the model filled the gap with the most statistically likely words about parkas. Those words are the same ones every other retailer’s model reached for.
So the fix sits mostly upstream of the writing. It lives in what you feed the tool, how you edit the output, and who on the team gets the final say.
Start with a fact sheet, not a prompt
The single habit that changes AI product copy the most is building a short fact sheet for each SKU before anyone opens a chatbot. It doesn’t need to be fancy. A spreadsheet row works. What matters is that it contains information only your business has.
For that parka, a useful fact sheet might look like this:
- Fill: 650 fill power responsibly sourced down
- Shell: matte nylon with DWR finish, sheds light snow and freezing drizzle
- Length: hits mid-thigh on a 5’6” model wearing size S
- Hood: removable, faux fur trim detaches separately
- Pockets: two hand-warmer pockets with fleece lining, one interior zip pocket that fits a large phone
- Fit note from returns data: runs slightly small in the shoulders, customers layering a thick sweater size up
- Store staff comment: “People like that it doesn’t look puffy from the side.”
- Care: machine wash cold, tumble dry low with dryer balls
Notice the last three lines. The fit note comes from your returns system. The staff comment comes from someone on the floor at your Yorkdale or Rideau location who has watched people try it on. Those details are the ones a generic model will never invent, and they are the ones that make a shopper trust the page.
Then the prompt becomes something closer to this:
Write a 90 to 120 word product description for the parka below. Audience: commuters in Toronto and Ottawa who take transit and walk 10 to 15 minutes in winter. Use only the facts provided. Do not use the words “ultimate”, “stylish”, “perfect”, or “must-have”. Mention the sizing note plainly. Plain, warm tone, short sentences, Canadian spelling. [paste fact sheet]
The “use only the facts provided” line does a lot of work. It stops the model from claiming the jacket is rated to minus 40 when nobody tested that, which is a legal problem as well as a tone problem.
Ban list, house style, and a few sample paragraphs
Every brand has words it overuses. Every model has words it overuses too, and the two lists overlap more than you’d expect. Pull the last 50 descriptions your team wrote by hand, the ones people liked, and read them out loud in a meeting. You’ll hear your own voice in maybe ten of them. Paste two or three of those into your prompt as examples of tone, and paste a ban list of the words that make you wince.
A typical ban list for apparel and home goods includes “ultimate,” “effortless,” “timeless,” “versatile,” “must-have,” “perfect for any occasion,” “take your X to the next level,” and “look no further.” Your list will differ. A hardware retailer probably has a separate set of tired phrases (“tackle any project”) that a fashion brand never sees.
House style also covers the small mechanical stuff: whether you write “colour” or “color” (for a Canadian audience, almost always “colour”), whether you use centimetres first or inches first, whether sizes are written “S/M/L” or “Small, Medium, Large.” Models get these wrong unless told, and inconsistent units across a category page look careless.
Where an AI humanizer fits in the workflow
Even with a good fact sheet, first drafts often come back with a recognisable rhythm: three short clauses, a benefit, a call to action, repeated across every SKU. When you’re producing a few hundred descriptions for a seasonal drop, rewriting each one from scratch defeats the purpose of using AI at all. This is where a rewriting pass helps. Running the draft through an ai humanizer varies the sentence structure and loosens the stiff phrasing so the copy reads less like a template, while keeping the facts you supplied. Quillbot’s version, for example, lets you paste the draft and get back a smoother rewrite you can then edit line by line.
Treat that step as a middle pass, not a finish line. The tool can change how the sentences sound. It can’t know that your Quebec customers hate the word “puffer” or that the zip pocket is the thing everyone mentions in reviews. A person still has to read every description before it ships, and they should be checking for three things specifically:
- Every factual claim matches the fact sheet. Rewriting tools occasionally soften “machine wash cold” into something vaguer, or drop a measurement.
- The sizing and care notes survived intact.
- The description sounds like it belongs on your site and not on a competitor’s.
A before and after
Here is a real-shaped example, built from the parka fact sheet above.
First AI draft (generic prompt, no fact sheet):
Stay warm and stylish all winter long with our Women’s Insulated Parka. Crafted with premium materials and designed for ultimate comfort, this versatile jacket is perfect for everything from city commutes to weekend adventures. Its timeless silhouette and thoughtful details make it a must-have for your cold-weather wardrobe.
Nothing in that paragraph is false, exactly. Nothing in it is useful either. A shopper learns zero about the jacket.
After fact sheet prompt, rewrite pass, and human edit:
Built for the walk from the subway to the office when the wind comes off the lake. 650 fill power down keeps the core warm without the marshmallow look from the side, and the matte shell shrugs off light snow and freezing drizzle. It falls to mid-thigh. The hood comes off, and so does the faux fur trim, separately. Two fleece-lined hand pockets, plus an inside zip pocket that fits a big phone. Sizing note: it runs a little narrow in the shoulders, so if you plan to wear a chunky sweater underneath, go up a size. Machine wash cold, tumble dry low with a couple of dryer balls.
The second version is longer, but every sentence carries something. And it reads like a person who has seen the jacket.
Bilingual copy needs its own pass
Retailers selling into Quebec already know the French requirements are not optional, and product information has to be available in French. AI translation from English has improved a lot, but direct translation of marketing copy still produces French that sounds translated. Quebec shoppers spot it immediately. The idioms come out European, or the sentence structure follows English word order a little too faithfully.
The better approach is to generate the French version from the same fact sheet rather than translating the finished English copy. Give the model the facts, tell it to write for a Montréal or Québec City shopper in Québécois French register, and have a francophone team member review it. Translating the English draft means you inherit every quirk of the English draft, plus a translation layer on top. Starting from facts means each language gets its own natural phrasing. ## Different products need different lengths
One mistake I see often is a single prompt template applied to the whole catalogue. A $14 pack of tea towels does not need 150 words. A $1,200 sectional sofa probably needs more than 150, plus dimensions, delivery notes, and fabric care.
A rough split that works for a lot of general retailers:
| Product type | Typical length | What the shopper actually needs |
|---|---|---|
| Low-cost consumables and basics | 30 to 60 words | Size, quantity, material, one use note |
| Apparel and footwear | 80 to 130 words | Fit, fabric feel, care, sizing advice from returns |
| Electronics and small appliances | 100 to 150 words plus spec table | Compatibility, what’s in the box, warranty |
| Furniture and big-ticket home | 150 to 250 words plus spec table | Dimensions, assembly, delivery, materials, care |
Set separate prompt templates for each tier, and you stop trimming 140-word descriptions of dish soap.
What to keep human, always
Some parts of the description should never be AI-generated without a person checking them against a source document:
- Safety and compliance claims (CSA certification, flammability, age grading on toys)
- Warranty terms
- Allergen or ingredient lists on food, cosmetics, and supplements
- Origin claims (“Made in Canada” has specific rules attached, and “Product of Canada” has stricter ones)
- Sustainability claims, which the Competition Bureau has been paying closer attention to
For these, paste the exact wording from the supplier spec sheet or your compliance team, and tell the model to reproduce it verbatim. Then check that it did.
Building it into a team process
A merchandising team of four can comfortably handle a few hundred descriptions a week this way, if the steps are clear. One workable setup:
- Buyers or product managers fill in the fact sheet when the SKU is set up, including the fit or use notes from samples.
- A copywriter or ecommerce coordinator runs the generation prompts in batches by category.
- The drafts go through a rewrite pass to break up repetitive phrasing.
- A second person reads every description against the fact sheet and the brand voice examples.
- Store staff feedback and early reviews get added to the fact sheet after launch, and the top-selling SKUs get their descriptions refreshed once real customer language is available.
Step five is the one most teams skip, and it’s the one that keeps the copy alive. Three weeks after launch, your reviews will tell you how customers describe the product in their own words. “Finally a jacket my arms can move in” is better copy than anything a model will produce, and it costs nothing to borrow the idea (paraphrased, with the reviewer’s words left in the review section where they belong).
Common mistakes
Generating from the product title alone. This produces the generic paragraph every time. No fact sheet, no useful copy.
Letting the model invent specs. Watch for temperature ratings, waterproof ratings, battery life figures, and thread counts that nobody gave it. These are the claims that end up in complaint emails.
Running everything through the same voice. Your kids’ department and your premium kitchen line probably shouldn’t sound identical, even if the brand voice is consistent overall.
Skipping the read-aloud. Read three descriptions in a row out loud. If you hear the same sentence shape three times, so will your customers. ## Tools and the people using them
AI does well at producing a first draft quickly and at rewriting the tenth description of a similar item so it doesn’t echo the ninth. It’s weaker at knowing which detail matters most to your customer, and it has no idea what your store staff hear at the fitting room. Feeding it those details is the job.
Using a rewriting tool on drafts that came out stiff is fine, and an ai humanizer is useful when you have a long run of near-identical products (twelve colourways of the same merino crewneck, say) and need each page to read a little differently without changing the facts. Just keep the fact sheet open beside it. The tool handles rhythm and phrasing; your team handles truth and judgment. ## FAQ
Will shoppers notice if product descriptions are AI-written? They notice generic writing, regardless of who wrote it. Descriptions built from specific facts, fit notes, and real use details rarely read as machine-made, even when a tool did the first draft.
Should we disclose that AI helped write product copy? There’s no general requirement in Canada to label AI-assisted product descriptions right now. What matters legally is that the claims are accurate. Some brands mention their process in an “about our content” page, which is a reasonable choice but not a standard one.
How often should descriptions be refreshed? Top sellers deserve a look after the first wave of reviews comes in, usually within a month or two of launch. Everything else can wait for the next seasonal update or a change in the product spec.
One small test to run this week
Pick your ten best-selling SKUs. Pull up the current descriptions and, for each one, write down a single fact a shopper would want that isn’t on the page right now: a sizing note, a care quirk, a staff observation, a common question from customer service. Then rebuild those ten descriptions from a fact sheet using the steps above and compare them side by side with the originals in a quick team review. You’ll know within an hour whether the process is worth rolling out to the rest of the catalogue, and you’ll probably find two or three of the old descriptions had a factual error nobody noticed.
