Writing Product Copy That Doesn't Invent Facts

Hallucination is a prompt design problem before it is a model problem#

When a language model writes "military-grade aluminium alloy" about a product you described only as "aluminium", it has not malfunctioned. It has done exactly what it was asked: produce text that resembles high-converting product copy. High-converting product copy in its training data contains phrases like "military-grade". The model has no way to distinguish the parts of that pattern that were grounded in fact from the parts that were decoration.

The fix is structural. You have to make accuracy the primary objective and persuasion the secondary one, in that order, explicitly.

Separate facts from instructions#

The single highest-impact change is to put the product facts in their own clearly delimited block, and to state that the block is the complete set of permissible factual content.

VERIFIED FACTS (the complete set — do not add to it):
- 1.7 litre capacity
- 2200W concealed heating element
- 304 stainless steel body
- Boil-dry protection, auto shut-off
- 1.2m cord, 2-year manufacturer warranty

Then instruct the model that any number, material, dimension, certification or origin claim must come from that block and nowhere else. This turns "be persuasive" into "be persuasive within this constraint set", which is a very different task for the model.

Give the model an escape hatch#

A model that is told "always state a benefit" will invent a benefit when none exists. A model that is told "if no benefit is verifiable, say so in a compliance note instead" will usually take the honest path, because it has been given a valid way to complete the task without fabricating.

This is why good copy tools return a notes field alongside the copy: "withheld the temperature rating — no evidence supplied" is more valuable to a seller than a fabricated number that gets the listing suppressed.

The taxonomy of an invented claim#

Invented content in product copy falls into five patterns. Knowing them makes them easier to catch in review:

• Invented specifications — a number that was never measured: "boils in 90 seconds" when nobody timed it.
• Inflated specifications — a real number rounded in your favour: 3.8L becomes "4L capacity".
• Borrowed certifications — "FDA approved", "CE certified", "OEKO-TEX certified" applied to a product that was never tested, often because a different product from the same factory was.
• Implied origins — "handmade in Portugal" when the product is factory-made and shipped from a warehouse elsewhere.
• Phantom comparisons — "2x stronger than the leading brand" with no test, no brand named, and no definition of "stronger".

Every one of these is a marketplace policy violation and, in most jurisdictions, an advertising-law problem as well — the FTC's truth-in-advertising standard requires claims to be substantiated before they are published, not after someone complains.

The red-flag vocabulary#

Some phrases are almost always decoration rather than fact, and experienced reviewers treat them as prompts to check the spec sheet:

• "military-grade", "medical-grade", "airline-grade", "professional-grade" — grades that name an authority which almost never certified this product.
• "lab-tested" — which lab, testing what, against what standard?
• "scientifically proven" — a study exists, or a marketer liked the phrase?
• "premium quality" — an honest opinion, but a flag that the sentence may be carrying weight it cannot bear.

None of these words is banned. Each of them is a check me signal — if you cannot attach evidence, replace the word with a fact.

What to do when the facts are thin#

Sometimes the spec sheet really is three lines long. The honest options are still good options:

• State what you know precisely, and stop. Three accurate sentences beat five padded ones.
• Use clearly subjective, non-comparative language: "designed for small kitchens" is an opinion; "fits any kitchen" is a claim.
• Say what the product is, not what it beats. Origin, material, use case, and dimensions are all copy-safe facts.
• Get the missing fact instead of writing around it — one email to the supplier is cheaper than one suppressed listing.

Test with adversarial inputs#

Try your own setup on a product you know inside out, then check:

• Did any number appear that you did not supply?
• Did any material, certification, or country of origin appear that you did not supply?
• Did the model repeat a claim you flagged as unverifiable?

Run this on ten products, not one. A single clean pass proves nothing; a model that hallucinates on 1 in 10 runs is still publishing 1 in 10 unsafe listings.

The tooling difference#

General chatbots let you paste guidelines every session and hope. A purpose-built tool bakes the constraint set into the prompt template, validates the output against it, and scans the result against a compliance rule set before you ever see it. That is the difference between a suggestion and a workflow — and for marketplace copy, a workflow is the only thing that scales.