Where ChatGPT sends jewelry buyers.

A mention is not a path.

When ChatGPT recommends a jewelry brand, the recommendation is only worth money if the buyer can click through to the brand’s own store. Often the name lands but the link routes somewhere else. That is the difference this study measures.

And an owned path is itself one rung on a taller ladder. What a brand actually wants from ChatGPT is a climb: from being known at all to being consistently, accurately shoppable.

Birthstone and personalized jewelry is the standout: five brands named across three runs, and only Ana Luisa named in 3 of 3. Occasion gifts is the opposite kind of open: 19 brands and no clear owner. Meanwhile lab-grown engagement rings and solid gold returned the same regulars every run: Blue Nile, Brilliant Earth, Clean Origin, James Allen, and Ritani for engagement; Aurate, Baby Gold, Catbird, GLDN, Mejuri, Oradina, and Quince for solid gold.

For a jewelry brand, the move is not to fight the locked intents. It is to own a wide-open one before ChatGPT settles on its answer.

This is a correlation in a small sample, not a law. It is the kind of pattern we publish so it can be tested: if it holds at larger sample sizes, breadth of AI visibility and ownership of the buy path are separate problems that need separate fixes.

What sits underneath the answers.

Two layers below the visible answer shape these outcomes. The first is the catalog layer. Catbird is the strongest broad performer in this study: named in 20 of 30 answers, majority-owned paths. Yet in a separate product-level check we ran on July 1, its five tested products were missing from its own storefront catalog feed, and catalog lookups for its product names matched other sellers instead, including a “Katbird Bracelet” on a store that is not Catbird. A brand can be winning today’s answers while its catalog layer quietly belongs to someone else. That product-level problem is the subject of our next full research paper.

The second is the source layer: the pages ChatGPT leaned on while composing these answers. In the decoded retrieval and support payload of the 30 runs, the most frequent domains were not brand sites.

That is the trap inside “AI visibility”: a brand can be visible because a Reddit thread or a magazine roundup mentions it, not because of anything on its own pages. Visibility borrowed from third parties is visibility a brand does not control.

Limitations, stated up front.

This is a directional field study, not a market ranking. Each intent ran three times, so per-intent results are counts like “named in 3 of 3 runs,” not finely measured shares. Everything comes from one model, one mode, and one day: ChatGPT Instant, US sessions, June 29, 2026. These systems change without notice, so every figure carries its date, and brands named once or twice are anecdotes, not rates.

Path classification here is binary and brand-level: owned-store links divided by times named. This study does not split the non-owned 64.5% into retailer links versus no link at all, and it does not pin each owned link to a specific run. The source-domain counts come from the decoded retrieval and support payload, which includes search context, not only confirmed citations. And the brand-comparison prompt names four brands directly, which pads their mention counts; we footnote it wherever it matters.

Next in this series: a larger calibrated benchmark for the same category, with more runs per prompt and product-level path classification, so the binary split above can become the full owns-leaks-no-path picture.