AI Product Descriptions Detection: Fix Catalogue Copy

AI product descriptions detection usually flags catalogue copy for one reason: it says nothing only you could know. The fix is not rewriting in a different tone. It is adding measurements, materials, unit numbers and use cases that no model can invent, then varying the sentence templates across your range.

Why does generated catalogue copy read the same across a whole range?

A model producing a product description has no access to your stock, your supplier sheet or your returns data. So it reaches for the same safe claims every time: premium quality, perfect for everyday use, designed with care. Run that prompt across forty SKUs and you get forty variations of the same paragraph. The nouns change; the structure does not.

Our own measurement gives a useful reality check here. On 2026-09-13 we scored two corpora with the DeAIze detector at a 30% flag threshold. A set of 23 hand-written DeAIze guides came out at a mean AI score of 21%, with 5 of 23 flagged. A set of 12 unedited model outputs, generated from fixed prompts, came out at a mean of 33%. The interesting part is the spread: both corpora had passages in the same range. Detector output is a probability estimate, not proof of authorship, and a single score will not tell you which lines are weak.

That is why sentence-level detection matters for catalogue work. You do not need a verdict on the whole product page. You need to know which two sentences carry the signal, so you can replace them with something true.

What details make a product description sound human?

Specificity. Not adjective swaps. A buyer scanning a category page is comparing twenty tabs, and the only descriptions that survive that comparison contain facts the other nineteen cannot copy.

Four categories of detail do most of the work:

1. Measurements — internal dimensions, strap drop, cable length, pack volume, seat height, weight when assembled. Give the number and the unit.
2. Materials — not “premium leather” but the actual grade, the lining, the hardware finish, the recycled content if there is one.
3. Use cases — who buys it and for what. “Fits a 13-inch laptop and a water bottle” beats “spacious and practical”.
4. Unit numbers — pack size, quantity per case, servings per container, sheets per roll, hours of runtime. Unit numbers are the single most underused detail in ecommerce copy, and they are the hardest thing for a model to fabricate convincingly.

A test you can run on any draft: delete every sentence that would still be true of a competitor’s product. What is left is your description. If nothing is left, the copy is generic regardless of who wrote it.

If you want to see which sentences in a product page carry the AI signal before you rewrite, paste the text into DeAIze — the free tier covers 200 words on signup, no credit card, and you get a per-sentence breakdown rather than one opaque score.

How do you stop every product using the same sentence template?

Build a template bank, not a template. Most catalogue teams write one strong product template and reuse it, which is exactly the pattern readers and detectors pick up on. Instead, keep a small set of structural options and rotate them deliberately.

Opening move	Works well for	Trade-off
Lead with the measurement	Technical gear, furniture, apparel sizing	Reads dry if you never vary it
Lead with the use case	Gifts, accessories, seasonal items	Can sound vague without a number nearby
Lead with the material or origin	Premium goods, food, textiles	Weak for commodity items
Lead with the problem it solves	Tools, replacements, refills	Risks sounding like ad copy if overdone
Lead with the specification table	Components, parts, consumables	Poor for browsing-led categories

Rotate across the range and the catalogue stops reading like one voice on autopilot. Two practical rules: never open more than two consecutive products the same way, and never reuse the same closing sentence twice in a category.

If you are also running paid traffic to these pages, the same discipline applies to the ad copy that feeds them — see our guide to AI ad copy for the equivalent editing pass on headlines and hooks.

Where do SEO keywords fit without wrecking the phrasing?

Keyword placement in product copy has a narrow, honest job: make the page findable for the phrase a buyer types. It should not dictate the sentence.

• Put the primary term in the title, the H1 and the first line of the description.
• Use the natural variants a buyer would actually type — material, size, colour, use case — rather than repeating one phrase.
• Let the specification table carry the long-tail terms. That is what it is for.
• Never stack three keyword phrases into one sentence. It reads as spam to buyers and it is one of the clearest signs of machine-written catalogue copy.

Google’s guidance rewards helpful, original content and does not penalise AI-assisted writing as such. What it penalises is scaled, low-value pages — which is precisely what a range of near-identical descriptions becomes. Our piece on AI content SEO without penalty covers where that line sits for content teams.

A rewrite example: generic AI copy to seller-specific copy

Here is an illustration using invented copy, not a real product listing.

Before (generic, model-shaped):

This premium stainless steel water bottle is perfect for everyday use. Its durable construction and sleek design make it ideal for the gym, the office or travel. Stay hydrated in style wherever you go.

After (seller-specific):

750 ml stainless steel bottle with a powder-coated finish that survives a dishwasher cycle. The mouth is wide enough for ice cubes; the lid seals at 100 degrees so it can lie flat in a bag. Fits a standard car cup holder. We ship it with a spare gasket, because the gasket is the part that fails first.

What changed: a capacity and a unit, a material and finish, a real constraint (cup holder fit), a tested claim (dishwasher, sealing angle), and one detail that only the seller would know — the spare gasket. The second version is longer and far more useful. It is also much harder to mistake for a template.

Notice what did not change: no invented statistics, no fake awards, no “best in class”. Specificity carries the persuasion, not adjectives.

Should you humanise generated catalogue copy or rewrite it?

Honest answer: for a handful of hero products, rewrite by hand. You know the product, and a fifteen-minute edit beats any tool. For a 4,000-SKU long tail, hand-rewriting every listing is not realistic, and that is where a humanizer earns its place.

DeAIze runs a closed loop: it scores the text, rewrites only the flagged sentences, keeps a rewrite only when semantic similarity says the meaning survived, then re-scores and reports before and after. The site reports an average AI-score reduction of 60%+ after humanizing, best case 71%, measured on its own predicted scores rather than a third-party verdict. On the example paragraph used on the site, a 92% AI score becomes 4%.

The limits matter more than the numbers. A humanizer cannot invent the spare gasket. If your source copy contains no seller-only detail, the rewrite will be cleaner and still empty. Humanising is the second pass, not the first.

Here is the order that works for a range:

1. Pull the real specifications for each SKU into a spreadsheet.
2. Generate a first draft that includes those specifications, not just marketing language.
3. Run detection to find the sentences still carrying the generic signal.
4. Rewrite or humanise those sentences only.
5. Re-score and spot-check the meaning against the spec sheet.
6. Rotate the opening structure before publishing the batch.

The same triage applies to bulk work; our bulk humanizing workflow covers how to run it across a backlog without losing track of which listings were edited.

What should you check before publishing a catalogue batch?

A short pre-publish pass catches most of what makes a range read as automated:

• Does every description contain at least one fact from the spec sheet — a measurement, a unit number, a material?
• Do any two products in the same category open with the same sentence structure?
• Is there a claim you cannot verify from your own data? Cut it or source it.
• Does the closing line repeat across the range?
• Would a competitor be able to paste this description onto their product without changing a word?

That last question is the one that matters. If the answer is yes, the copy is generic, and generic copy is what AI product descriptions detection is picking up on — along with your customers.

The ideas, specifications and claims in a product description have to be yours. A tool can clean up the phrasing. It cannot supply the facts, and it should not be used to pass off someone else’s listing as your own.

---

Measurement note: figures in this article come from our own detector run on 2026-09-13 — n = 23 hand-written guides versus n = 12 unedited model outputs, median AI score 22% and 34% respectively. Method: /methodology/.