When a machine misquotes a product twice, the interesting question is not whether the model behaved badly. It is why the wrong phrase was waiting there, close enough to be borrowed.
A founder sends me several screenshots. In one, an AI answer says her body cream is “fragrance-free,” which it is not. In another, a different tool calls it “dermatologist tested,” which the site does not claim. A third answer describes the product as “for eczema-prone skin,” a phrase the team has carefully avoided. The product page itself is tidy. The problem looks like it arrived from outside, wearing muddy boots.
The more typical picture, assembled from several beauty and wellness catalogues, is less tidy. The brand page says one thing, a stockist page says another, an old marketplace listing says something half-remembered, and reviews use buyer language that sounds close to a claim. The model then produces a sentence that is not quite invented and not quite sourced. It is a crooked echo. That is why repeated misquoting deserves more attention than a single alarming screenshot.
A misquote is often a clue, not a verdict
I do not like diagnosing a brand from one AI answer. A single response can be strange for reasons that have little to do with the catalogue: the prompt, the model version, the retrieved source mix, the user’s phrasing, or simple answer noise. One screenshot is an event. A repeated distortion is a pattern. The pattern is where the work begins.
AI misquoting of product descriptions is the repeated deformation of a product’s stated identity, because the available language around that product gives machines several plausible but conflicting ways to describe it.
That definition is important because it keeps the blame in the right place. Sometimes the model does make a clean error. More often in ecommerce, it is working from unstable material. The product has several names, several benefit claims, several category cues, or several versions across sales channels. The answer stitches them together and produces a line that feels made up because no human on the brand team would have written it. Yet pieces of it were lying around.
I use the term “borrowed wrongness” for this. Borrowed wrongness is a false or distorted product statement assembled from nearby language that was already public, semi-public, or strongly implied. It is not pure invention. It is a bad loan.
This distinction matters for founders. If the line is pure invention, the response may be hard to prevent beyond improving source clarity and monitoring repetition. If it is borrowed wrongness, the catalogue has repairable weak points. Old copy can be retired. Stockist descriptions can be corrected. Claims can be tightened. Review language can be supported or bounded. The machine’s error becomes a map.
The phrase usually comes from somewhere
In most cases, the first task is dull and useful: find the ancestor phrase. I lay out the product page, collection page, metadata, stockist listing, marketplace copy, old ads if available, review snippets, and common AI answer phrasings. Then I look for the family resemblance. The exact wrong sentence may not appear anywhere. A parent phrase often does.
A composite skincare example shows the shape. A body cream was described by an AI answer as “fragrance-free barrier repair for eczema-prone skin.” The live product page did not say this. It said “lightly scented,” “supports the skin barrier,” and “for dry, reactive-feeling skin.” A stockist page, however, had once used “suitable for eczema-prone customers” in a loose merchandising line. A review said, “I use it when my eczema flares,” which was a buyer’s own statement. The marketplace title used “barrier repair cream.” None of those fragments alone created the false claim. Together they made the wrong sentence easy.
The imperfect detail: the same answer also got the tube size wrong. That part looked like a plain retrieval error from an older listing. I would not build a whole diagnosis around the size error. The claim distortion was more consistent, and it appeared across several prompts. Patterns matter because they separate catalogue weakness from stray grit.
This is why founders sometimes feel gaslit by AI answers. The response is wrong, but not random. It sounds like the brand through a wall. The cadence is familiar. The claim is close. The category is almost right. That closeness is the danger, because a buyer may accept the sentence as a reasonable summary.
Product pages leave fingerprints
A strong product page has a narrow fingerprint. It can be paraphrased without losing its identity. A weak page has several fingerprints layered over each other. Machines can lift any of them. This is especially common where a product has changed position over time. A cream began as “natural body care,” became “clean clinical care,” then shifted towards “barrier support.” Old phrases remain in stockist copy, meta descriptions, review prompts, and collection pages. The current page may be right, but the public language field is still mixed.
I do not think every old phrase must be hunted down with panic. Brands evolve. Some residue is normal. The question is whether the residue contradicts the claim the brand now needs preserved. If an old line says “deeply nourishing” and the current line says “rich daily moisturiser,” there may be no real conflict. If the old line says “fragrance-free” and the product is lightly scented, the residue is dangerous. If buyer reviews repeatedly make medical-adjacent claims the brand cannot evidence, the page needs a boundary close to those reviews.
There is also a quieter fingerprint problem: style words that stand in for category words. “Kind,” “clean,” “restorative,” “ritual,” and “clinical” can all be useful in context. Used without hierarchy, they invite paraphrase. A model may preserve the emotional or tonal field while changing the functional claim. “Restorative” becomes “repairing.” “Clinical” becomes “dermatologist tested.” “Kind” becomes “suitable for sensitive skin.” Each step may look small. The final answer may cross a line the brand itself never crossed.
The machines are not reading tone with a lawyer’s caution. They are reading likely associations. A page that wants caution must build caution into the structure, not hope the adjective behaves.
Where repeated misquotes begin
Repeated misquotes tend to begin in one of five places. I call them the five loose seams. They are not a checklist for public display; they are a way to inspect the garment.
The first seam is naming variance. The product has different names across the product page, collection page, stockist, marketplace, and internal catalogue. A model may combine names or choose the one with the clearest category term. The second seam is claim expansion. A cautious brand phrase sits beside buyer or stockist language that states the claim more strongly. The third is category borrowing, where a product appears in a collection or marketplace category that is wider or more medical, technical, luxury, or budget than the brand’s intended category. The fourth is proof looseness, where evidence exists but does not sit close enough to the claim, so the answer fills the gap with a common category assumption. The fifth is old-copy residue, where previous positioning still appears in searchable places.
These seams explain why the same wrong line can return. It is not that the model loves the error. The public material keeps offering the error a route. A founder may correct the product page and still see the misquote because the answer has other sources, or because the corrected page still uses an ambiguous phrase that points back to the old idea.
A teaching example: imagine a cleanser once sold as “soap-free care for sensitive skin.” Later the brand changes it to “low-foam cream cleanser for dry skin.” Some stockists keep “sensitive skin.” Reviews mention rosacea. A collection page says “calming essentials.” An AI answer calls it “a rosacea-friendly cleanser for sensitive skin.” The product may indeed be gentle. The brand may not have evidence for that specific claim. The misquote begins where cautious product language and uncontrolled user language are allowed to shake hands without a chaperone.
That last phrase sounds fussy. It is also the work.
Correcting the answer starts before the answer
People often ask for prompt fixes when a product description is misquoted. I understand the instinct. The screenshot is visible, so the temptation is to fight the screenshot. But if the wrong phrase has public ancestors, a prompt is late in the chain. It may help test the pattern. It does not repair the material that keeps feeding the pattern.
The repair usually starts with a product-language trace. Take the misquoted phrase and split it into its claims, category terms, modifiers, and use cases. For “fragrance-free barrier repair for eczema-prone skin,” the trace would separate fragrance status, barrier claim, repair language, and eczema suitability. Then inspect where each part appears. Which parts are true and supported? Which are buyer language only? Which are old? Which belong to a stockist? Which are implied by a collection heading? The answer is often sitting there in pieces.
Next, the brand needs a canonical product description that is short enough to travel. I do not mean a stiff boilerplate paragraph. I mean a clear source description used consistently across product page, collection intro, marketplace listing, stockist notes, schema fields, and internal catalogue exports. If the product is a lightly scented barrier-support body cream for dry, reactive-feeling skin, then that structure needs to hold. It can be adapted for tone and space. It should not become “fragrance-free,” “eczema care,” or “clinical repair” elsewhere.
Finally, claims need boundaries near the places where they might be expanded. If reviews mention conditions, the page can acknowledge that reviews are personal experiences and then restate the evidenced use case. If a stockist wants stronger wording, the brand should give them approved alternatives. If the collection page uses “restorative,” it should say what kind of restoration is meant in product terms. Loose language is not charming when it makes a model cross a regulatory or trust line.
A stable description can still sound alive
Some teams hear “canonical description” and imagine dead catalogue prose. I would resist that. A stable description does not have to be sterile. It only needs to keep the same commercial bones wherever it travels. Tone can vary. The product identity should not.
A good description has a product type, a buyer situation, a supported claim, a bounded use case, and a few proof cues. It avoids dragging in claims the page cannot evidence. It also avoids overusing soft adjectives that require the model to guess the job of the product. “A lightly scented body cream for dry, reactive-feeling skin, built around barrier-supporting lipids and a cushioned finish” gives more shape than “a kind, restorative cream for skin that needs care.” The second sentence may be lovely in a founder note. It should not be the main source sentence.
The goal is not to make every AI answer identical. That is not possible, and it would be a strange ambition. The goal is to make wrong paraphrases less likely and less plausible. When several systems summarize the product, the same core should survive: what it is, who it is for, what claim it can carry, and what evidence supports that claim.
A repeated misquote is embarrassing, but it is also useful. It shows which part of the product language can be bent. Once you know that, the work becomes less mystical. Follow the phrase back through the catalogue. Find the loose seam. Stitch it where buyers and machines can see it.
The Shelf Note
Object: a body cream repeatedly described as fragrance-free eczema care. Distortion: the model may assemble a false claim from stockist residue, review language, and loose benefit words. Counterweight: a canonical product description, retired old copy, bounded claims, and proof placed close to the use case. Shelf line: A misquote is often the catalogue speaking with an old mouth.