Jewellery Product Descriptions: What AI Gets Wrong on Metals

AI-written jewellery descriptions most often get five things wrong: they call plated pieces gold, confuse carat with fineness, drop the word "lab-grown", state a stone size that doesn't match its weight, and invent hallmark or care claims. Check every metal, stone and size statement against your spec sheet before publishing.

The errors are rarely wild. They are small, plausible softenings: "gold-plated brass" becomes "gold ring", "9ct" becomes "9 carat gold stones", a 5 mm stone becomes "a generous 5 carat centre stone". A customer who receives a plated ring sold as gold has a refund claim, and in some markets a regulator has an interest too. The checks below are grouped by the part of the listing where each slip tends to hide.

Follow me on Instagram@sagnikteaches

Why jewellery trips AI up more than other products

A language model writes the most likely next phrase. In jewellery copy the most likely phrase around a metal word is a flattering one. "Gold" appears next to "ring" millions of times; "gold-plated brass" appears far less. Left to fill gaps, the model reaches for the common, glossier wording.

Connect on LinkedInSagnik Bhattacharya

Three more things make it worse. Jewellery vocabulary has double meanings (carat measures gold purity in one phrase and stone weight in another), the trade uses abbreviations that vary by supplier (GP, GF, VM, 925, 14K), and the facts that matter most, such as plating thickness and stone origin, usually sit in a supplier PDF that never reaches the prompt. Whatever isn't in the prompt, the AI makes up or smooths over.

Subscribe on YouTube@sagnikteaches

So the process has two halves. First, feed the model a clean spec record so it has less to invent. Second, check what it wrote against that record, one claim at a time. A spec record, a claim-by-claim check and a simple log cover both halves. If you also need pricing logic behind the same data, the jewellery pricing calculator uses the same fields.

Build a one-row spec record before you prompt

For each piece, write down the facts the description is allowed to use. One row per product, in a spreadsheet, with these columns:

• Base metal: what the body of the piece is actually made of (sterling silver, brass, stainless steel, 9ct gold alloy).
• Surface: none, gold plate, rhodium plate, vermeil, gold filled, enamel, oxidised.
• Plating detail: thickness in microns if the supplier gives it, otherwise the word "unknown".
• Fineness stamp: 375, 585, 750, 925 and so on, exactly as stamped.
• Stone: the name, the origin (natural, lab-grown, simulant such as cubic zirconia), any treatment (heated, dyed, filled), and how you know.
• Stone size: millimetres and, where you have it, carat weight, as two separate cells.
• Piece size: ring size range, chain length, pendant drop, earring height, in mm or cm.
• Weight: grams of the finished piece.
• Hallmark: yes, no or not required, plus what is struck on it.
• Care: what the supplier says, nothing more.

Where a cell is unknown, write "unknown" rather than leaving it blank. A blank invites the model to fill the gap; "unknown" gives it permission to say nothing. The prompt further down uses these columns directly.

The main example: one ring, eight errors

Here is an illustrative case. A small online shop sells a ring whose spec row reads: base metal brass; surface 18ct-colour gold plate, thickness unknown; no fineness stamp; stone cubic zirconia, 6 mm round, no carat weight given; ring sizes K to R; weight 3.1 g; hallmark not applicable; care "avoid water and perfume".

The owner asks an AI tool for a short description and gets this (illustrative output):

Treat yourself to our radiant 18 carat gold solitaire ring, set with a dazzling 6 carat diamond-style stone. Crafted for lasting wear, this hypoallergenic ring is hallmarked for quality and will never tarnish. Available in sizes K to R, with a petite 3 mm band that sits comfortably on any finger.

Now take it apart claim by claim against the row.

1. "18 carat gold": wrong. The spec says brass with an 18ct-colour plate. The model turned a colour description into a purity claim.
2. "6 carat": wrong unit. The 6 is millimetres. A 6 carat stone would be enormous and very costly; the AI swapped a diameter for a weight.
3. "diamond-style": ambiguous. It is a cubic zirconia, so say that. "Diamond-style" lets a skimming buyer read "diamond".
4. "Solitaire": fine, a single stone setting, and the picture shows one.
5. "Hypoallergenic": unsupported. Plated brass may contain nickel, in the alloy or in a layer under the plate, and the spec says nothing about it. Remove it unless the supplier states nickel-free and you can show the paperwork.
6. "Hallmarked for quality": invented. The record says hallmark not applicable, and a plated brass ring carries none.
7. "Will never tarnish": false. Plating wears and brass reacts with skin and moisture. The supplier's own care note says avoid water and perfume.
8. "3 mm band": invented. Band width wasn't in the row.
9. "Crafted for lasting wear": soft, but it contradicts the care note when read next to "never tarnish", so cut it.

The corrected version, again illustrative:

A solitaire ring in brass with a gold-coloured plate, set with a 6 mm round cubic zirconia. Available in sizes K to R and weighing 3.1 g. The plating will wear with time, so keep it away from water, perfume and cleaning products and wipe it dry after wearing.

It is plainer and shorter, and every claim in it can be pointed to in the row. Of the nine claims in the first draft's 51 words, eight were wrong or unsupported: roughly one problem for every six words. Keep that ratio in mind as you read the checks below.

Metal checks: solid, plated, filled, vermeil

These four words get swapped by AI more than any others, so learn what each means in plain terms.

• Solid gold means the whole piece is a gold alloy of a stated fineness, not just the surface. If you can't state the fineness, you can't say solid.
• Gold plated means a thin layer of gold over a different metal. One widely used set of jewellery trade guidelines treats electroplate as a layer at least 0.175 microns thick and heavy electroplate as at least 2.5 microns, with the gold layer at 10 karat or higher.
• Gold filled means a layer of gold of at least 10 karat mechanically bonded over a base metal, making up at least 1/20 of the weight of the metal in the whole item under the same guidelines. It is much thicker than ordinary plate and not the same thing.
• Vermeil means sterling silver as the base, with gold of at least 10 karat fineness at least 2.5 microns thick. If a base-metal layer such as nickel sits between the silver and the gold, the same guidelines expect that to be disclosed.

Those numbers come from one market's guidelines, and your market may differ. The practical rule travels anyway: if a word has a threshold behind it, only use it when your supplier can show you meets the threshold. AI will happily write "gold vermeil" for any gold-coloured silver piece.

What to look for in the draft

• Any use of "gold" without "plated", "filled", "vermeil" or a fineness number beside it.
• "Gold-tone" or "gold-coloured" on a solid-gold piece (it undersells; fix it) and "gold" on a gold-tone piece (it oversells; much worse).
• Phrases such as "real gold", "pure gold" or "genuine gold" unless the piece is 24ct (999).
• "Solid silver" when the stamp is 925: sterling is 92.5% silver, which is the normal meaning, but write "sterling silver 925" and let the number do the work.
• "Stainless steel" that has become "silver" in the tags or the alt text.

Carat, karat and fineness: the unit mix-up

Gold purity and gem weight use the same word. "9ct" or "9 carat" on gold is purity; "0.5ct" on a diamond is weight (one carat is 0.2 grams). A model merges these constantly, and a buyer can be misled in either direction.

Two checks matter here. First, make sure each metal word is paired with the stamp in your record, not with what the AI thinks that metal "usually" is. Second, run the arithmetic: pure gold in a piece equals its weight times its fineness. A 2.4 g ring in 9ct gold (375) holds 0.9 g of fine gold. Descriptions that quote "contains 2.4 g of gold" are wrong by more than half, and the number is checkable by anyone with scales.

Stone checks: origin, name, treatment and size

Stones add four ways to go wrong.

Origin

Natural, lab-grown, and simulant are three different things. A cubic zirconia or moissanite is a simulant of diamond, not a diamond of any kind. A lab-grown diamond has essentially the same composition and structure as a mined one, which is exactly why it must carry its qualifier every time. AI often drops "lab-grown" because "diamond" is the more frequent word, and the omission can be a disclosure failure, not just a style slip. The separate lab-grown diamond wording checklist goes through the terms in detail.

Name and shape

Shape words get promoted into stone names. "Emerald-cut" describes a rectangular step cut, not an emerald; "sapphire blue" describes colour, not a stone. Look for any gem name in the draft that isn't also in the spec record's stone column.

Treatment

Many natural gems are heated, dyed or filled. One major market's jewellery guides expect sellers to disclose a treatment when it isn't permanent, needs special care or significantly affects value, and the safest habit is to state every treatment the supplier records. If the supplier row says "heat treated" and the draft says "natural ruby, untouched", the draft contradicts the record. If the record is silent, write "unknown" and ask the supplier before listing.

Size and weight

Millimetres and carats are not interchangeable without knowing the stone. As a rough guide, a well-cut round diamond of one carat measures roughly 6.3 to 6.6 mm across, depending on its proportions, but a gemstone with a different density (a sapphire, say) gives a different diameter for the same weight. A single wrong unit can make a stone sound five or ten times larger than it is, as in the ring above. Whenever you see a number next to a stone, ask: is this mm or ct, and does the record say so?

Hallmarks, marks and certificates

Hallmarking rules vary by market and are legally enforced in some. A draft will often claim "hallmarked" on any precious-metal piece, or "certified" when all you have is a supplier invoice. The hallmarking tutorial covers what copy must not imply. For this checklist, the test is simple: the word "hallmarked" appears only if the record says so and you can photograph the strike. "Certified" needs a named certificate and number you can show on request.

Dimensions and fit

Dimension errors cost refunds, not legal letters, but they hurt just as much. Check these every time:

• Ring size system: AI mixes letter sizes, number sizes and diameters. State the system you use, and add the inner diameter in millimetres if your supplier gives one.
• Chain length: 16 inches and 40 cm are close but not equal (40.6 cm). A necklace that quotes "45 cm (18 in)" should match what you measured.
• Drop and height: earring drop (from the top of the hook) and pendant drop (excluding the bail) are measured differently by different suppliers. Copy the supplier's definition.
• Band width and thickness: the AI loves to invent "delicate 2 mm band". Only use a width you have measured.
• Weight: grams for the finished piece, never for the stone.

Tools that put jewellery on AI-generated models have their own scale problems; the true-to-scale test covers that side.

Care and comfort claims AI makes up

These are the sentences that read well and cost you in returns: "won't tarnish", "waterproof", "hypoallergenic", "suitable for sensitive skin", "lasts a lifetime", "safe to wear in the shower", "ethically sourced". Strip each one unless the supplier gives a written statement you can show a customer. Replace them with the supplier's own care instruction and, if relevant, a plain line such as "Plated pieces wear over time".

A prompt that gives the model less room to invent

This prompt pastes your spec row and sets the rules. Copy it and replace the bracketed block.

You write jewellery product descriptions for our online shop.

Use ONLY the facts in the spec record below. Do not add any metal, stone, size, weight, hallmark, certificate, origin or care claim that is not in it. If a field says "unknown", leave it out and do not guess.

Rules:
- Name the base metal and the surface in the first sentence (for example "brass with gold plate").
- Never write "gold" on its own. Use the fineness number for solid gold or "plated" / "filled" / "vermeil" for the rest.
- Keep millimetres and carats apart. Say "6 mm", never "6 carat", unless the record gives carat weight.
- If the stone is a simulant or lab-grown, say so in the same sentence as the stone name.
- No claims of hypoallergenic, waterproof, tarnish-free or lifetime wear.
- 80 to 100 words. Spell words our way (colour, jewellery). No exclamation marks.

After the description, list every fact you used and the field it came from.

SPEC RECORD:
[paste the row]

The last instruction matters. Asking the model to list its sources turns checking into a short comparison of two lists, not a hunt through prose. If a source line points at a field that doesn't contain that fact, you have found the error in seconds.

Running the check on a batch

Do this in a sheet, not in the chat window. Give each product a row and these columns: SKU, claim, record field, matches (yes, no, unsure), fix. For a batch of 40 products with roughly six checkable claims each, you will have 240 lines; most will be yes. The no's cluster, and the cluster tells you what to change in the prompt next time.

An example log (illustrative) after one batch: 14 "gold" with no qualifier, 9 carat/mm swaps, 6 invented care claims, 5 stone origins dropped, 4 wrong chain lengths, 2 hallmark claims on unstamped pieces. That is 40 problems across 240 claims, about one in six, and it tells you the first prompt rule to add is the "never write gold on its own" line.

Two review passes cost less than one return. Let someone other than the writer do the second pass; a fresh reader sees "ring is gold" where the author remembers "ring is gold-plated". The general method for this is in setting up human review without slowing down.

A card-based shortcut for very small catalogues

If you list fewer than 30 pieces, skip the sheet and use a printed card: one page with the ten spec-row fields, filled in by hand and clipped to each sample piece. Writing the description straight from the card, with AI polish afterwards, gives you the record and the evidence together. When a customer disputes a claim a year later, the card is what you find.

When the mistake got through

An illustrative case, because they are common. A seller lists a silver-coloured chain described as "sterling silver" by the AI. The supplier sheet said "stainless steel, silver colour". Nobody checked, because the photo looked right and the price was low. The first complaint arrives a fortnight later: the buyer had a jeweller test it. The shop refunds, rewrites the listing, and then finds the same wording on eleven other products in the same batch.

What would have caught it: the "surface and base metal" cell in the record, a source list after the description, and a rule that nothing says "sterling" without a 925 in the record. The fix cost an afternoon. The refunds and the review cost more.

Tags, titles and alt text carry the same errors

AI also writes your title, your marketplace tags and your image alt text, and each can repeat the mistakes in shorter form. A title such as "18K Gold Diamond Ring" for a plated CZ piece is the same error as the body copy, and it appears in search results where there is no room to correct it. Run the same claim check on titles, tags, meta descriptions and alt text. A useful habit is to search your own catalogue for the words gold, silver, diamond, sapphire, hypoallergenic and certified, then open every hit and compare it with the record. It takes minutes and catches the things a batch review misses. The Shopify product descriptions tutorial covers the template side if your store uses one.

Where to stop trusting the draft entirely

Some claims should never be left to a model and a quick read: anything about investment value ("will hold its value"), health ("good for circulation"), origin ("conflict-free", "ethically mined") and legal status ("hallmarked", "certified"). Write those by hand from documents, or leave them out. If you want a second opinion on where AI belongs in your listing process and where it doesn't, that's the kind of job a 1:1 call is for, linked from the contact page.

One-page version to keep beside the screen

1. Spec row filled, with "unknown" where unknown.
2. Prompt includes the row and the six rules.
3. The draft's source list matches the row line by line.
4. The word "gold" never stands alone.
5. mm and ct kept apart.
6. Stone name carries origin: natural, lab-grown or simulant.
7. No hallmark, certificate, hypoallergenic, waterproof or tarnish-free claim without paper behind it.
8. Sizes use one stated system.
9. Title, tags and alt text checked as well.
10. A second person reads it before it goes live.

Questions jewellers ask about checking AI copy

Can I paste a supplier's spec sheet into the AI and trust the output?

Trust it more, not fully. A spec sheet in the prompt cuts invented details, but AI still rounds numbers, merges rows and softens words such as plated into gold. Treat every metal, stone and size claim in the draft as unchecked until you have matched it to the sheet, line by line.

Is 'gold tone' an acceptable way to describe plated pieces?

It is safer than 'gold' because it signals appearance, not content, but it can still mislead if the base metal and plating are not stated nearby. Say what the piece is made of, for example brass with gold plating, and let 'gold tone' describe only the colour.

Should I ask the AI to flag its own mistakes?

It can help as a second pass, because the model will sometimes spot a contradiction between two sentences. It cannot know your actual stock, so it will miss anything the sheet does not contain. Use it to find inconsistencies, then verify against your records yourself.

How often should I recheck old listings?

Recheck whenever the supplier, the plating, the stone source or the size range changes, and whenever you rewrite a batch with AI. A quarterly spot check of ten random listings against the spec sheet catches drift without a big project.

Further reads

• AI Tools for a Jewellery Business, Compared by the Job They Do — Which AI tool suits which jewellery job.
• How to Build a Product Glossary AI Must Use in Every Draft — Lock your metal and stone terms into every AI draft.
• Jewellery Photography With AI: Clean Backgrounds, True to Life — Keep photos honest about colour and finish.
• AI Alt Text: How to Generate and Check Image Descriptions — Alt text repeats the same metal and stone errors.
• ChatGPT Prompts for Jewellery Shops: 8 for Listings and Clients — Eight prompts for listings and client messages.
• Online Shop AI Mistakes That Hurt Trust and Conversions — Other AI slips that cost a shop its buyers' trust.
• How to Use AI for Etsy Titles, Tags and Descriptions — If you list handmade pieces on a marketplace.
• AI for Jewellery Design: Concept Sketches to Client-Approved Renders — A brief-to-render routine for bespoke jewellers: ten brief fields, three rounds of variants, a buildability check and a concept sheet the client approves.
• AI for Fabric Stores: Cut-Length Orders and Fabric Descriptions — A fabric shop's workflow: supplier sheet to checked description, then a metre-based order form and emailed requests turned into cut-length lines.
• Can You Put Jewellery on an AI Model? Disclosure Rules for Sellers — A scale check with real numbers, visual checks, platform rules for AI images and caption wording to copy for listings, ads and social posts.
• Bulk Product Description Generator: From CSV Rows to Unique Copy — How to turn a 500-row product CSV into unique descriptions with AI, check for duplicates and invented specs, and import the result into Shopify safely.
• WooCommerce Product Descriptions With AI: Bulk Updates and Safe SEO — A staged way to rewrite WooCommerce descriptions in bulk: export, tier your pages, draft from facts, test on staging, check samples and watch results.
• AI Tools and AI Development: The Complete 2026 Guide — the AI hub, including every tutorial in the AI-for-business series.

Sources: published jewellery trade guides (16 CFR Part 23) for gold plate, gold filled and vermeil thresholds, unqualified 'gold' and gemstone treatment disclosure; GIA's definition of the metric carat; millesimal fineness marks as listed by an assay office.