Every conversation about AI visibility eventually arrives at the same place: what exactly is supposed to be in the data? This is our answer, built from running the problem on our own jewelry catalog before taking it to anyone else's.
It splits into three tiers. Tier one is what an assistant needs to surface your product at all. Tier two materially improves how well you match real shopping questions. Tier three covers differentiators and long-tail intent.
Before the fields: one spelling per concept
The single most common failure has nothing to do with which attributes you carry. It is that the values are inconsistent.
Yellow Gold. yellow gold. 14KY. YG. 14k Yellow. Your team knows those are one thing. A machine treats them as five, matches none of them reliably, and quietly downgrades the lot. Before adding a single new field, normalise the ones you have to a controlled list. That work alone moves the needle more than most people expect.
Tier one — required to appear at all
Identity
A stable product ID that does not change between feed refreshes, because a rotating ID resets every ranking signal you have accumulated. A group ID shared across variants, without which six ring sizes look like six unrelated rings. A structured title. A canonical URL that stays put.
And an honest identifier field. Custom and made-to-order pieces genuinely have no universal barcode, and the correct move is to declare that explicitly rather than fabricate a number or leave it blank. Mishandling this is the most common single cause of jewelry products being silently dropped.
Metal
Metal type from a fixed list. Karat as its own field. And critically, a field distinguishing solid from vermeil, gold-filled and plated — this is what answers the enormous "won't tarnish" and "won't turn green" question cluster, and it is also where disclosure obligations sit.
Stone
Centre stone type, shape, and carat weight as separate numeric fields. Total carat weight, which is the correct filter for tennis bracelets and pavé work. Stone origin — natural, lab-grown or simulant — which is both a disclosure requirement and one of the highest-volume consumer filters in the category. Treatment disclosure, for the same two reasons.
Commerce and trust
This is the tier-one group almost nobody structures, and it is where the easiest wins are. Resize policy. Return window. Warranty type. Production lead time. Whether the piece is made to order. Whether it is final sale.
Shoppers ask assistants about exactly these things. "Jewelry brands with free resizing and a lifetime warranty" is a real, frequent, high-intent question, and it is answered from fields — not from your policy page.
Tier two — match quality
Setting style from a controlled list, because solitaire, halo, hidden halo, three-stone, pavé, bezel and cathedral are the actual vocabulary of bridal search. Ring size as a proper variant axis. Band width in millimetres, which men's band shoppers ask for explicitly. Chain style and length. Earring back type. Certification lab and number. Colour, clarity and cut grades. Occasion tagging, which carries the entire gifting query cluster.
Also in tier two, and badly underserved: nickel-free and hypoallergenic as boolean fields. There is a large, motivated audience asking about this and almost no structured data to answer them with.
Tier three — differentiators
Metal finish. Metal weight. Recycled metal. Accent stone details. Clasp type. Water resistance. Appraisal inclusion. Engraving availability. Financing availability. None of these will make or break you, and each one wins a specific question your competitors cannot answer.
How to actually get there
Do not try to fill sixty fields across ten thousand SKUs at once. Take your top-selling hundred products, get tier one complete and normalised on those, and measure what changes. The pattern you establish becomes the template, and the template is the deliverable that matters — it is what your team applies to every collection you launch from then on.
The standard is not the point. The discipline of maintaining it is.