Ask ChatGPT to recommend independent engagement-ring designers. Then open a fresh conversation and ask exactly the same question. Will you get the same brands?
In our first AI jewelry discovery experiment, we examined which independent brands ChatGPT recommended. This time, we tested something different: whether those recommendations remained consistent when the question stayed exactly the same.
We tested eight independent fine jewelry shopping questions, repeating each in three fresh ChatGPT conversations. Across 24 answers, ChatGPT featured 47 different brands in 146 recommendation slots.
Not one of the eight questions produced identical featured-brand lists across all three runs. That raises a practical question: what does it mean to be discoverable when the shortlist itself keeps changing?
24
fresh ChatGPT answers
47
different featured brands
0
identical three-run shortlists
Three runs. Sixteen engagement-ring brands.
Engagement rings produced the least consistent answers in this experiment. One run named six designers, the next ten, and the third seven. Across those answers, 16 different designers appeared. Ten appeared in only one run.
Only Marrow Fine appeared in all three. In the first answer, ChatGPT opened with Frank Darling. In the next two, it opened with Marrow Fine. ARTËMER appeared only in the third run; Anna Sheffield appeared in the second and third.
The average overlap between pairs of engagement-ring lists was 22%, using Jaccard similarity. This does not mean the recommendations were bad—it means a relevant designer could be present in one answer and absent from another.
Explore the experiment · 24 first answers
Same question. Different shortlists.
Choose a shopping question to see exactly which brands had their own featured entry in each of three fresh ChatGPT answers.
Dark green = featured recommendation. Pale gold = not featured. R1–R3 = three separate Temporary Chat conversations. A brand listed elsewhere in an answer but not given its own entry is excluded.
A few brands kept returning
Catbird appeared in 17 of the 24 answers, WWAKE in 15, and Jennie Kwon Designs in 12. Together, those three brands occupied 44 of the 146 featured slots—about 30%. Yet 19 of the 47 featured brands appeared in only one run.
Catbird is both a jewelry brand and a multi-brand retailer, a distinction worth keeping in mind. These figures are not a ranking of the best jewelers or a measure of market share. They record appearances in this specific experiment.
The shopping question changed the brand universe
Colored gemstones produced the highest agreement, 57%, and a strikingly different shortlist. Irene Neuwirth, Harwell Godfrey and Retrouvaí appeared in all three colored-gemstone answers; so did Brent Neale. The colored-gemstone set shared almost no brands with the other seven shopping questions.
The lesson is not to make distinctive jewelry less distinctive. It is to make a brand's distinctive materials, design vocabulary and shopping occasions legible to people—and to systems answering their questions.
The difference between being seen once and being understood repeatedly
One appearance can be useful, especially when a shopper is looking for a highly specific style. But it does not establish that the brand will be recommended the next time someone asks. Across this test, 30 of the 47 featured brands appeared under just one of the eight shopping questions. A brand can be visible for a particular need without appearing across the broader category.
Consider the engagement-ring answers. Marrow Fine was present in all three, but ten other designers appeared just once. A single screenshot of one recommendation list would make some of those designers look consistently visible and others invisible. Neither conclusion would be supported by the three-run record.
For a jewelry business, the more useful exercise is to choose questions grounded in what customers actually buy: materials, occasions, price constraints, durability, customization and design language. Run each question more than once. Record not only whether your brand appears, but what the assistant says about it and which competitors share the answer.
What the repeated names do—and don't—tell us
Catbird, WWAKE and Jennie Kwon Designs returned often, but this experiment cannot establish why. It does not prove that a particular SEO technique, press mention, structured-data field or website design caused an appearance. It also cannot tell us whether a brand absent from these answers is absent from other shoppers' conversations.
The opportunity for smaller designers is not to copy the most frequently named brands. It is to describe their own work with enough specificity that the right shopping question has something clear to connect to: a gemstone specialty, a recognizable design approach, a documented custom process, or a concrete material and price range.
Being named is not the same as being verified
We selected 58 specific claims about prices, policies, sourcing and production timelines for checking. Twenty-three were verified, ten partly supported, thirteen were not found on the pages examined, five conflicted with available information, four had only secondary-source support and three could not be checked because the relevant sites were inaccessible.
None was established as definitively false. This was a selected set of claims, not a representative accuracy sample. “Not found” is not the same as “wrong.” For brands, the useful question is whether important product and policy details can be located and checked easily.
What a jewelry brand can actually do
Instead of chasing a supposed permanent AI ranking, test a range of realistic shopping questions. Examine whether your brand appears for its actual strengths, whether its own pages explain those strengths clearly, and whether an assistant's factual claims match what you publish.
One repeated answer is a signal, not a verdict. Repeating the question and inspecting the details reveals much more.
How we tested
On October 15, 2026, we ran eight fixed prompts three times each in fresh Temporary Chat conversations in ChatGPT's consumer interface, using the model labeled “Instant.” We recorded first answers without regenerating or following up. All prompts specified a shopper in the United States, but the signed-in Plus browser was outside the US; this was not a US-controlled test. Three captured answers ended mid-sentence and were retained as displayed.
The study describes these 24 answers, not what every shopper sees or why any particular brand was selected. Independence was not verified against company registries for the recommended brands. This is a separate experiment from our earlier five-prompt, logged-out study; their results should not be pooled. Read the research methodology and underlying data (draft research archive).
Earlier research: We Asked ChatGPT to Recommend Independent Fine Jewelry.