A lot of us are drafting product descriptions, collection pages, and blog posts with ChatGPT now. It saves real time, but I keep running into the same problem: a chunk of that copy reads as obviously machine-written. Same flat rhythm, the “elevate your everyday” openers, filler like “in today’s fast-paced world.” Shoppers can feel it, and it makes a store sound like every other store.
So the question for the group: how are you handling this? Are you writing product copy from scratch, editing the AI draft heavily, or shipping it mostly as-is? And have you actually seen it affect trust or conversion, or is it a non-issue in practice?
Here is what has worked for me, in case it helps. Before I publish anything AI-drafted, I read it against a few tells and fix them by hand:
- The opening line. AI almost always starts with a cliche (“In a world where…” / “Elevate your…”). Rewrite the first sentence to say something specific about the product.
- Sentence rhythm. AI writes sentences that are all the same length. Break a couple up, let one run long and the next be three words.
- Filler transitions. “Moreover,” “furthermore,” “when it comes to.” Cut them; the sentence almost always reads better without.
- Buzzwords. “Leverage,” “seamless,” “unlock,” “game-changing.” Swap for the plain word you would actually say to a customer.
Full disclosure, I got tired of eyeballing this every time, so I built a free browser tool that flags these automatically. You paste the text (or give it a page URL) and it scores how machine-written it reads and highlights the specific lines to fix. It runs entirely in your browser, nothing gets uploaded, and there is no signup. It is not magic and detection is never 100 percent certain, but it is a fast gut-check before you hit publish.
Free here if it is useful: Free AI Detector — Check if Text Is AI-Written (Runs in Your Browser)
Genuinely curious how everyone else is approaching this though, especially anyone running a lot of SKUs where hand-writing every description is not realistic.
I agree with removing clichés and generic rhythm, but I would not use an AI-detection score as the main publishing threshold. A page can sound human and still fail to answer the buying question.
After nine years of Shopify CRO work, my stronger test is: could a competitor paste this paragraph onto their product page without changing anything? If yes, the copy is still too generic.
I would edit around four inputs: the customer’s own words, the exact use case, concrete product evidence, and the objection that prevents purchase. “Premium material” is weak; the material, construction, care requirement and why they matter are evidence. “Perfect for everyone” is weak; a specific situation and non-fit are more credible.
To judge conversion impact, compare product-view-to-add-to-cart rate and the questions customers ask before buying. If the rewrite reduces repeated uncertainty while add-to-cart improves for the same traffic source, that is more meaningful than whether a detector labels the prose as human.
The reason most AI-generated product descriptions sound generic (“Experience the ultimate…”, “Elevate your style…”) is that merchants feed LLMs raw supplier specifications without giving the AI structured brand rules or specific product context.
To make AI product copy sound authentic while maintaining bulk listing efficiency:
-
Provide “Negative Prompt” Style Rules: Explicitly instruct the AI in your prompt to ban fluff words (“ultimate”, “unrivaled”, “seamless”, “game-changer”) and force a specific tone (e.g. “Write in concise, conversational UK English focusing on material specs and dimensions”).
-
Structure Description Schema in Spreadsheets: Generate product descriptions into structured CSV columns (e.g. ‘Key Features’, ‘Materials’, ‘Care Instructions’) using a spreadsheet LLM macro, then concatenate them into clean HTML for Shopify’s ‘Body (HTML)’ column rather than asking the AI to write raw HTML paragraphs.
-
Don’t Let AI CSV Exports Break the Image Pipeline: When merchants use AI tools to generate 100s of product listings from Asian supplier feeds (like 1688/Alibaba), the AI CSVs often reference raw .webp supplier image URLs. Shopify’s CSV worker quietly skips WebP links, leaving your beautifully written AI product pages with blank variant images.
If you’re automating bulk product listings with AI copy and need to handle the image-pipeline side locally, consider using EasyCatch (a client-side Chrome extension). Its Local Canvas Transpiler converts supplier WebP images to static JPGs inside your browser sandbox and exports Matrixify-compliant ZIPs with pre-mapped variant rows in 1 click — ensuring your AI-written listings import with perfectly matched images. 100% Local-First so store catalog data stays completely private.
@Vlad_Gerasimchuk ,
I still use AI for the first draft because it saves a lot of time especially on larger catalogs but I never publish it without adding something the AI couldn’t know on its own.
The biggest improvement for me has been including real-world details why customers buy the product common questions they have asked what it pairs well with, or even who it isnt suitable for. Those are the things that make the copy feel authentic and genuinely helpful.
I dont worry too much about whether the text sounds AI-generated. I worry more about whether it gives a shopper enough confidence to make a buying decision. If it answers their questions better than the competing product page its doing its job.
HI @Vlad_Gerasimchuk
The biggest thing I’d avoid is trying to make the copy “pass” an AI detector. Those tools can produce false positives, and sounding human is better achieved through specificity and genuine brand voice.
For larger batalogs, I’ve found good workflow is:
AI draft → add product-specific facts → remove generic marketing language → add bran/customer context → human review.
Give the AI real source material-product specs, customer questions, reviews, differentiators, use case, brand vocabulary. The more specific the input, the less generic the output tends to be.
I’d also avoid publishing hundreds of descriptions with the same structure. Even if every description is technically unique, repetitive phrasing can make the whole store feel templated.
For SEO and conversion, I’d prioritize useful, accurate product information and a consistent brand voice over whether a detector thinks the text was written by AI.
Building an app that drafts product copy for Shopify catalogs, so this thread is basically my job description. Grain of salt accordingly.
The single change that mattered most for us was not letting the model see the product without its data. When you hand an LLM a bare title and ask for a description, filler is all it can write. When you hand it the materials, dimensions, care details and what the thing is actually for, most of the generic phrasing disappears on its own, because the model finally has something real to say.
Second: validate outputs against the inputs, not against a detector. We reject any draft that mentions a material or attribute that is not in the source product data. That kills the confident-sounding invented details, which are a worse tell than any cliche. Length limits and a short ban list of phrases catch most of the rest.
Icey.Lane’s competitor-paste test is the right bar. Specific beats human-sounding, and specific usually ends up sounding human anyway.