Shopify’s Commerce Trends 2025 reports that around 8 in 10 merchants use at least one AI feature in their store, and content generation, especially product descriptions, sits near the top of the list. The merchants who get hours back per week are not the ones using cleverer prompts. They are the ones who have stopped making the five mistakes that kill AI copy at scale. This post is the audit you can run on your own catalogue this afternoon.
Mistake 1: Asking for “a product description” with no anchor
The most common pattern: a merchant pastes a product name and a short feature list into ChatGPT and asks for a description. What comes back is competent and forgettable. It reads like every other product on the internet, because the model averaged the entire internet to write it. The failure mode here is subtle. The copy is not bad. It is just not yours, which means the customer who clicked into your product page from an ad has no reason to feel they have arrived somewhere specific.
The counter is to anchor the model with three things before asking for any copy: an example product description from your store that you wrote yourself and were happy with, your audience in one sentence, and the single comparison your product needs to win on the page.
“I’m writing product descriptions for [STORE NAME], a [PRODUCT CATEGORY] brand. Our customer is [ONE-SENTENCE DESCRIPTION, e.g. ‘a 32-year-old who buys quality basics every two years rather than fast fashion every month’]. The single most important comparison this product wins on is [COMPARISON, e.g. ‘softer than 100% merino at the same weight’]. Here is an example of one of our existing product descriptions that I am happy with: [PASTE]. Match its rhythm, sentence length, vocabulary, and opening structure. Now write a description for [PRODUCT NAME], using these specs: [PASTE FEATURE LIST]. Length: 90 to 110 words. Open with a sensory or scenario detail, not the product name.”
That single change (“match the rhythm of the example I just gave you”) moves the output from generic to recognisable.
Mistake 2: Writing 200 descriptions in one prompt
The other extreme is the bulk request. Merchants paste a CSV of 200 SKUs into a prompt and ask for descriptions for all of them. The model produces them. The descriptions are interchangeable, structurally identical, and visibly written-by-AI. Customers cannot articulate why, but they bounce.
The fix is to batch in groups of five to eight, by product category, and to vary the structural brief across batches. Five descriptions for fragrance candles in a “sensory opener” structure. Five descriptions for ceramic mugs in a “use-case opener” structure. Five descriptions for kitchen knives in a “comparison opener” structure. The model holds voice better in small batches and the structural variation prevents the rhythm-fatigue that makes catalogue browsing feel flat.
Shopify Magic bakes this into the product editor for stores already on Shopify, and it is fine for a first draft. The bulk-batch approach above will outperform it on copy that has to feel handwritten.
Mistake 3: Letting AI write the bullets
Product feature bullets are where AI is most likely to insert a polished untruth. Specifications need to be exactly correct. The model does not know the actual spec, only what is plausible for the category. Two failure modes here: the bullet that is technically wrong (a fabric weight, a battery capacity, a dimension), and the bullet that is technically true but sounds like marketing rather than fact.
Hard rule: never let the model write feature bullets from a product name alone. Always paste the spec sheet, and always ask for verification:
“Using only the specifications I am pasting below, write five feature bullets for this product. Each bullet must use a number or measurement that appears verbatim in my spec sheet. After each bullet, in brackets, quote the line from the spec sheet you used as the source. If a benefit you want to mention is not directly supported by my specs, omit it. SPEC SHEET: [PASTE].”
The bracketed source-quote forces the model to ground each claim. When you read the output, anything that does not have a clean source quote gets cut. The bullets that remain are accurate by construction.
Mistake 4: Forgetting search intent
Product copy is not just for the customer who already arrived on your page. It is for Google, and Google’s algorithms read your copy in the context of what someone typed into search. Most AI-generated product descriptions read beautifully and rank for nothing, because they don’t include the words actual customers use to find this kind of product.
Run a five-minute keyword check before writing. Type the product category into Perplexity or Google and look at the auto-suggestions, the “people also ask” box, and the language reviewers use. Then feed that back into the prompt:
“Customers searching for products like this commonly use the following phrases: [PASTE 5 to 8 PHRASES FROM AUTO-SUGGEST AND PAA]. Reviewers commonly mention: [PASTE 3 to 5 PHRASES FROM REVIEW SITES]. Rewrite the description below to naturally include three of those phrases without making the copy feel keyword-stuffed. Keep the same tone, length, and opening structure.”
The counterintuitive part: SEO-aware AI copy will often outperform AI copy that was carefully voice-tuned, on a revenue-per-page basis, because the discovery layer matters more than the conversion-per-visitor layer for catalogue pages.
Mistake 5: Skipping the AI-tell pass
Every output from a general-purpose model has tells. The em dash mid-sentence (for a brand that doesn’t use em dashes). The “elevate your” or “transform your” opener. The triplet rhythm (“soft, sustainable, and stylish”). The over-confident closer (“you’ll never go back”). Customers may not consciously notice any single one, but in aggregate, AI-tell is what makes a catalogue feel like nobody works there.
Run every batch through a tell-detector pass:
“Read these five product descriptions and flag any phrase that reads as AI-generated marketing copy. Look for: triplet adjective rhythms, generic verbs (elevate, transform, tap into), hedge words (truly, simply, just), opening with the product name, and closing with a vague positive claim. For each flagged phrase, suggest a more specific, brand-anchored replacement. Output as a table: original phrase, why it reads as AI, replacement.”
Three to five edits per description. Five minutes of work per batch. The output reads as if a person wrote it, because by the end you have written about half of it.
What you actually save in a week
A merchant with a 200-SKU catalogue who writes everything from scratch budgets roughly 25 to 30 hours per quarter for new product copy. The same merchant running the workflow above (anchored prompts, small batches, spec-grounded bullets, search-intent passes, tell-detector edits) typically lands at 8 to 10 hours for the same volume, with copy that reads more recognisably brand-specific, not less. The hours you save aren’t because the AI wrote the copy. They’re because the AI handled the structural draft work and you spent your editing time on the parts that matter.
The merchants quietly compounding revenue from catalogue pages over the next year aren’t the ones using AI hardest. They’re the ones using it for the right 70% of the work.