To use AI for product descriptions without invented specs, give it only your source data, tell it what it must not claim, and make it mark every gap with a placeholder instead of guessing. Then check each claim against the source, start with a small batch, and review a sample before you scale up. The rule underneath it all is simple: if the source data doesn't say it, the description doesn't say it.
Why AI invents product facts
A language model is built to produce fluent, plausible text. Asked to describe a plant pot, it will happily tell you it's frost-proof, because plenty of plant pot descriptions say so. It isn't lying on purpose; it's filling a gap with something likely. In a catalogue, though, a likely-sounding spec that isn't true is a wrong spec, and it can turn into returns, complaints or worse.
The fix isn't a better model. It's a workflow that gives the model nothing to guess about, and catches it when it guesses anyway.
A workflow for catalogue teams
1. Give it only the source data
Paste in the supplier's spec sheet and your own product record, and tell the model to use nothing else. Dimensions, materials, capacity, care instructions, warranty: if it's in the source, the model can use it. If it isn't, the model can't.
2. Tell it what it must not say
Keep a short list of claims that need proof before they appear in copy, and give it to the model every time. Typical entries are claims such as "frost-proof", "waterproof", "food-safe" or "eco-friendly", and anything about warranty terms. Each one is either backed by the source data or left out.
3. Make it mark gaps instead of guessing
Ask the model to write a placeholder, such as [MATERIAL] or [DIMENSIONS], wherever the source doesn't give it a fact it needs. A placeholder is easy to search for before anything goes live. A guess looks exactly like a fact.
4. Check every claim against the source
Read each finished description next to its source data and tick off every factual claim. Anything you can't find in the source comes out. This is the slow step, and it's the one that makes the rest worth doing.
5. Start small, then scale
Run a small batch first, check all of it, and fix the prompt where the model went wrong. Only then run a larger batch, and keep reviewing a sample of each one. If the sample starts showing problems, stop and fix the prompt before carrying on.
Keep a brand-voice file it reads every time
Tone is where AI copy drifts most. A short voice file, read at the start of every run, keeps it steady. Keep one per brand, and say in each who you're writing for, which channels the copy appears on and what the tone is, with a few words you use and a few you avoid. Keep it short enough that someone on your team would actually read it too.
The same file helps with more than product copy. Clear, consistent product facts are also what AI assistants and search engines need to describe you correctly; see getting found in AI answers.
What good looks like
- Every claim in a description can be traced to the source data.
- No placeholders are left in anything that's published.
- Nothing on your list of claims that need proof appears without that proof.
- The tone matches your voice file across the whole batch.
For the wider set of AI tools that make this kind of work quicker, see building an AI stack to get things done fast.
Want more practical guides on putting AI to work in an ecommerce business? Browse the rest of the MCP-G blog.