AI Can Write Your Listings: Who Checks Them?

It’s a six-inch stoneware bowl with a matte glaze, the kind you’d eat cereal out of, and in the supplier’s spreadsheet the entry for it reads, in full, “Stoneware bowl, 6 in, matte glaze, hand wash only.” Dull as dishwater, no pun intended. Then a seller I’ll call Dana fed that row into an AI writing tool along with fifty-nine others, and what came back for the bowl was a lovely little paragraph about how it’s dishwasher safe and microwave friendly and perfect for everyday use.

Dishwasher safe. Where did that come from?

Nowhere, really. The spreadsheet said hand wash only, but somewhere between the row and the paragraph that line got lost, or the tool just never weighed it properly, and when a language model is missing a fact it doesn’t leave a hole. It puts in whatever is usually true of bowls. And usually, bowls are dishwasher safe, so that’s what it wrote, in the same cheerful tone as everything else on the page. Three people looked at that listing before it went live. One of them even said “nice.” Ten days later a customer sent a photo of the bowl in two pieces, very politely, and Dana spent her evening on a refund, a replacement and an apology she’d have happily skipped.

I like that story because it isn’t really about AI being bad at writing. The writing was fine. It was better than fine, plainly better than what Dana would’ve written herself at nine at night. The problem is that she’d traded a slow job she understood for a fast job nobody was responsible for.

Fast drafts, slow everything else#

When something gets nearly free, people make way more of it, and product copy is no exception. Back when each description took half an hour, you wrote one per product and went to bed. Now it costs nothing to ask for the long version for your own site, a shorter one a marketplace will accept, something punchy for the social shop, a slightly cozier one for Etsy, and why not a Spanish version while we’re at it. Three hundred products across four places is twelve hundred pieces of text, and every one of them has your shop’s name on it.

Nobody budgets for the reading. That’s what bit Dana. She figured the AI would save her a week, and it did, right up until the point where she realized someone still had to read the output, check the facts, answer the customer whose package arrived with the wrong care label, and fix the listing so it wouldn’t happen again tomorrow.

The ways it goes wrong#

The mistakes sellers describe tend to cluster, so once you know the kinds you can almost hear them coming.

There’s promotion, where “water-resistant” gets bumped up to “waterproof” because the tool leans toward optimistic prose. There’s cross-contamination, where details from the previous row on the sheet bleed into the next one, so a wool sweater ends up with a washing instruction meant for a polyester fleece. There’s hallucinated compliance, where the tool adds “BPA-free” or “certified organic” because those words sound like things people want to hear about containers and cotton. And there’s silent drift across channels, where the Shopify description promises one thing, the Amazon bullet points promise another, and the customer who read both expects both.

The sneaky bit is that none of it looks like a mistake. Bad writing looks bad and you fix it without thinking. A factual error dressed up in fluent, confident English looks like great copy. You read right past it unless you’re holding the supplier sheet in your other hand.

What Dana did next#

She didn’t ditch the tool, which would’ve been silly. She just stopped trusting it to be alone with her store.

The first thing she did was build a plain sheet for every product, nothing but the details she couldn’t afford to have wrong. Material, dimensions, weight, care instructions, warranty, what comes in the box. No marketing language, just the hard facts. If the AI output didn’t agree with the sheet, the sheet won.

The second thing was giving up on reading every listing. By about number thirty she was nodding along, skimming, missing things anyway. So she switched to sampling: check ten percent of every batch against the sheet, and if more than two had an error, reject the batch and run it again with tighter instructions.

When she checks she’s not proofreading, she’s hunting for promises. Waterproof, organic, BPA-free, handmade, lifetime warranty, dishwasher safe. Any word that makes a claim a buyer could hold her to goes straight to the top of the pile.

Every so often she opens one product on every channel, just to see whether the price and the stock and the title match. You’d be surprised how often they don’t, especially after a midnight edit on one marketplace that never got copied over to the others.

Sam asked once whether a better prompt could make all of this unnecessary. Fair question. Better prompts help, of course, and you should use them. But a prompt cannot fix a spreadsheet row that was ambiguous to begin with, and it cannot keep a third-party marketplace from silently truncating your third bullet point because its character counter counts spaces differently than yours does.

The part nobody planned for#

Then came the other pile, which wasn’t listings at all. Where’s-my-order emails. A return. A review that said the color was off. A question on Instagram about whether the bowl comes in blue. None of it was complicated, but all of it was waiting, and the faster the catalog grew, the bigger the pile got.

Dana’s least favorite job was reviews, so she did what a lot of owners do and ignored them. She shouldn’t have. A calm, specific reply to a mediocre review is often the thing that convinces the next shopper to buy anyway. People aren’t looking for perfection, they’re looking for evidence that somebody is home.

Stock is quieter and nastier. A supplier delays a shipment by two weeks, and somebody has to notice, pause the listing or push back the ship date before thirty people buy something you don’t have. An AI tool will happily write thirty gorgeous descriptions of an out-of-stock bowl. It does not care.

Bringing in some help#

Dana hired a part-time helper, someone I’ll call Marisol, and she did it slowly, which I’d recommend to anyone.

In week one, Marisol got the support inbox and the order system and nothing else. Dana wrote her a short cheat sheet: here are our shipping times, here’s how returns work, here’s how we handle a broken item, here are the three questions we get five times a day. Dana read every draft reply before it went out. By day four she was changing almost nothing.

Around week three, once support felt smooth, listing checks and order work joined the list. By then Marisol already knew the products because she’d been talking to the customers who bought them. That turns out to matter. A person who has answered three emails about bowl care is much faster at spotting an unverified “dishwasher safe” in a product description than someone reading it cold.

Is that overkill for a shop this small? Maybe, depending. Dana did the math on her evenings, though. Ten hours a week at ten or twelve dollars an hour was about the cost of a nice dinner out, and in exchange she stopped waking up at two in the morning wondering whether she’d answered that refund email from Tuesday.

How did she know it was working? Replies went out faster, tickets closed sooner, returns stayed flat even as the catalog grew, and when a supplier quietly changed a spec on a ceramic pitcher, Marisol caught it before it went live. That last one paid for a month of her time on its own.

If you want to do the same#

There are two ways that I’d look at. If you want a dedicated long-term hire on your own team, Globaltize can find, vet and place an ecommerce assistant who already understands order flows, customer tickets and store admin, and they handle the payroll and paperwork so you don’t have to become an international employer on day one. If you’d rather have someone run the search while you stay in charge of the vetting, Hire A Virtual Assistant gives you pre-screened candidates across administrative, sales and store tasks, and you interview and choose who joins.

Where in the world matters a little. The Philippines has a lot of ecommerce and support experience, strong English and great familiarity with US platforms. Latin America gives you real-time timezone overlap with North America and native Spanish if you sell into those markets.

Whoever you choose, start with a small paid test. Ten real support tickets and five listings to check. You’ll know within two days whether you communicate well, and two days of testing beats two months of polite mismatch every single time.

Back to the bowl#

Dana still has one of those bowls on her desk, the replacement, holding paper clips. She says it reminds her of the difference between writing copy and running a store.

So let it draft. Have somebody trained look before anything goes out, and keep the real decisions for yourself. The math works, your listings stay honest, and your evenings get your own again.