Why AI product descriptions get facts wrong, and what fixes it

AI product descriptions are accurate when the product data behind them is accurate, and wrong when it isn't. The model doesn't check anything. Give it a part with no confirmed fitment and it will still write a confident fitment line, because a plausible sentence is what it's built to produce.

You could have a description written for every product in your shop by Friday. I know, because I could do it for mine. It would read well and it would be the right length. Some of it would also be wrong, and nobody would notice until a customer bought the wrong part.

I've watched this play out in three businesses. An automotive parts retailer with over 64,000 products on Shopify. A heating and gas controls distributor. And my own brand, Drehwerks. Different sectors, same problem. Writing was never the hard part of a big catalogue. Knowing was.

This post covers why AI gets product facts wrong, what fixes it, what that looked like in each business, and a short checklist you can run against your own catalogue this week. It's long. Product data usually is.

Are AI product descriptions accurate?

For tone, length and structure, usually yes. For facts, they are only as accurate as what you feed them.

The errors cluster in predictable places:

• Fitment and compatibility. Which vehicles, engines, burners or models a part works with.
• Specifications. Dimensions, ratings, materials, thread sizes.
• Claims with legal weight. Road legality, certifications, approvals.
• Box contents. What's included, and what isn't.

These are also the fields customers make their decision on. A slightly flat sentence about build quality costs you nothing. A wrong thread size costs you a return, carriage both ways, and usually the customer.

Why AI invents product specs

A language model doesn't look things up. It predicts the most likely next words. Ask it to finish "Fits:" for a part with no fitment data, and the most likely words are a list of vehicles that sound right. Sometimes they are right. That's worse, because it teaches you to trust the next one.

Three situations cause most of the damage:

• The field is missing. The model fills the gap instead of leaving it.
• The sources disagree. Two supplier files list different fitment. The model blends them or picks one, and doesn't tell you which.
• A rewrite drifts. You ask it to improve a description that was correct. A 3/4" thread becomes 5/8" because that sounded better in the sentence.

Prompting helps less than you'd expect. "Only use the data supplied" reduces invention, but it doesn't remove it. The reliable control is structure. Give the model fields rather than freedom, then check what it wrote against those fields before anything is published.

What fixes it: product data provenance

Provenance means that every field on a product carries a record of where it came from and whether a person has confirmed it. It sounds like admin. It is admin. It's the difference between a catalogue you can trust and one you hope is right.

In practice, each field sits in one of four states:

• Empty. Nothing yet. Can't go live, and the section hides. Automation can fill it.
• Supplier. From a supplier or manufacturer file. Can go live, with care. Automation can overwrite it when the supplier updates.
• AI draft. Generated from the fields above. Not until reviewed. Automation can overwrite it.
• Confirmed. A person who knows has checked it. Can go live. Automation cannot overwrite it.

The last state is the one that matters. Once a person has confirmed a field, nothing automated should be able to change it. Without that rule, next month's supplier feed quietly undoes the work your team did this month.

Fitment data on a 64,000-product Shopify store

The first business was a UK automotive parts retailer on Shopify. It had over 64,000 products, and the same question sat behind every one of them: will it fit my car?

The answer arrives from suppliers in industry fitment files. ACES carries the vehicle fitment, and PIES carries the product attributes. Those files disagree with each other, with the manufacturer and with the customer more often than you'd hope. Shopify has no native home for any of it, which is a polite way of saying it's your problem.

So I built a system that takes the supplier files in, builds the fitment from them, drafts the copy with AI, and records where every field came from. Once a person has confirmed a field, nothing automated can overwrite it.

In its first twelve months the shop took £279k net across 578 orders, with no paid advertising behind it. I can't tell you how much of that was the data. I can tell you nothing else was holding it up, because there was no ad budget.

Spare parts lookup for a heating controls company

The second is a B2B heating and gas controls company. Nobody there is asking whether a part fits their car. An engineer is on site, sometimes with patchy signal, trying to find the right spare for the right burner.

The parts list I started from gave different parts identical names. Valves, mostly. Same name, different part. Which goes down well at seven in the morning in a plant room. No description, AI-written or otherwise, could fix that, because the information that told them apart wasn't there.

The fix was in the source data. I went back and added whatever separated one part from the other, so the lookup could tell them apart, and so could the engineer standing in a plant room.

Drehwerks: 374 products and an empty-field rule

The third is Drehwerks, my own brand. I'm in the middle of it, and on some days regretting it. It has 374 products across lamps, light bars, crate engines, engine parts, transmissions, tyres and wheels, and every type needs different information. A light bar needs a beam plot. A tyre needs a size table with the load index and speed rating. An engine part needs to say which engines it actually fits.

Each section of a product page reads one or two fields, so I gave myself one rule. If the field is empty, the section disappears. If the field is wrong, the page is wrong. If I'm not sure, I leave it empty.

About twenty engine parts are sitting in a queue marked "confirm". The supplier's catalogue suggests which engines they fit, and I'm not putting "fits" on a page until someone who knows has said so. Road legality works the same way. A product page never claims it unless the product carries the approval, and then it names the approval.

How to use AI for product descriptions safely

AI is properly good at volume. Reading a maker's spec sheet and drafting the specs, box contents and compatible parts for a few hundred products is exactly what I use it for. I'm not anti-AI. I'm anti-guessing. These seven steps keep that useful without letting it publish guesses:

1. Decide which fields matter. Fitment, key specs and anything with legal weight. Everything else is presentation.
2. Structure the data before writing anything. Fields first, copy second. If a fact isn't in a field, the copy can't rely on it.
3. Give the model fields, not freedom. Pass it the structured data and ask it to write from that alone.
4. Treat everything it writes as a draft. Nothing goes live until a person has read it and said yes.
5. Check the output against the fields. Every number in the copy should exist in a field. This is easy to automate and catches most drift. It's also the step everyone skips.
6. Lock confirmed fields. Once a person has confirmed a field, supplier updates and AI rewrites can't touch it.
7. Leave gaps empty, and hide them on the page. A missing section is honest. A confident guess is a returned part.

Check your own catalogue in ten minutes

Pick five products and one field that matters on each: fitment, a key spec, or a compatibility list. For each one, ask:

1. Where did this come from? A supplier file, a person, or a guess?
2. Who last confirmed it, and when?
3. What happens to it when the supplier changes theirs?

If you can answer all three for all five, your data is in better shape than most. If you can't, what you have on those pages is text. Nice text, perhaps. Still text.

Frequently asked questions

Can I use AI to write Shopify product descriptions?

Yes, and for large catalogues you probably should. Shopify has its own built-in tool, and plenty of apps do the same. The risk isn't the writing. It's publishing facts the model inferred rather than read from your data.

Do AI-written product descriptions hurt SEO?

Not because AI wrote them. Google's guidance is that it judges content on whether it's helpful, not on how it was produced. What does hurt is thousands of near-identical pages that add nothing, and wrong information that sends customers straight back to the search results.

What are ACES and PIES?

They are the automotive aftermarket's standard formats for product data. ACES describes which vehicles a part fits. PIES describes the part itself: dimensions, weights, descriptions and packaging. Most large parts suppliers publish both, not always in agreement with each other.

Do I need a PIM?

A product information manager earns its keep once you have more products than you can check by hand, or more than one supplier feeding the same catalogue. Below that, a well-structured spreadsheet with a column recording the source of each field gets you most of the way. Above it, the spreadsheet starts running you.

Send me ten products

None of this is exciting. One field at a time, one product at a time, written down once so it's right everywhere. I still like it.

If you've got a big catalogue and a feeling that your product pages aren't telling customers what they need, send me ten products. I'll tell you what I'd want to know as a buyer and couldn't find. I take the first five requests each month, and there's no pitch attached.

Get in touch