Who the Machine Names When a Shopper Looks for a Jeweler: Independent Visibility Across Five US Metros

Introduction

A shopper buying jewelry is rarely buying a specification. They are buying a claim they cannot check. Is this diamond really the clarity grade on the tag? Is the gold eighteen karat all the way through? Is the setting the work of a skilled hand or a cheap casting that will loosen inside a year? For most buyers these questions stay open even after the ring is on the finger and the receipt is filed. In the language of economics, jewelry is a credence good: its quality is hard to judge before the purchase and hard to verify after it. That single property shapes how the whole category is searched, shortlisted, and chosen.

Unable to verify craft, the buyer substitutes proxies for it. They read reputation, the depth and tone of other people’s reviews, how long a shop has held its corner, whether a friend was treated well. The purchase is also occasion-driven and infrequent, among the largest discretionary outlays a household makes, which raises the stakes and stretches the search. Industry surveys put the average US engagement-ring spend at roughly 5,200 dollars in 2024, with more than half of couples spending one to four months choosing (The Knot and industry press, 2024). A category that is high-ticket, rare, emotionally loaded, and impossible to inspect is a category where trust is the product.

The competitive terrain follows from that. The independent local jeweler competes with more than the shop across town: national mall chains, luxury houses, and increasingly online-first sellers such as Blue Nile and James Allen that turned the diamond into a filterable database. The question this study fixes on is narrow and answerable: when a shopper looks for a jeweler through the machine, whom does the machine name, and where does the independent local shop actually stand?

Data and method. We treated the machine’s local answer as an observable object. In the capture on 22 July 2026, we ran a grid of five real shopper prompts across five US metros and captured, for each, the Google local pack, the organic top ten, and the Google AI Overview where one appeared. We read US search demand for the category’s buyer language. We then asked, in a small and deliberately modest sample, whether the independents the machine names are even structurally ready to be read. Every figure below is dated to its capture and tagged with an evidence tier. Where an engine was not captured, it is named as uncaptured, never estimated.

Background and literature

The economics of this category were mapped before the internet existed. Nelson (1970) split goods into search goods, whose quality a buyer can assess before purchase, and experience goods, whose quality is revealed by use. Darby and Karni (1973) added the harder third class, the credence good, whose quality the buyer struggles to judge even after consuming it, and showed that this information gap creates a persistent temptation for sellers to over-supply or misrepresent. Zeithaml (1981) carried the same taxonomy into marketing, arguing that goods and services high in credence properties change how consumers evaluate them: buyers lean more on word of mouth than on advertising, weigh price and physical surroundings as quality cues, and do most of their judging after the purchase rather than before it. Dulleck and Kerschbamer (2006) then formalized the credence-goods market and named its defining ingredients: an expert who knows what the customer needs, a customer who does not, and the resulting scope for both under-treatment and over-charging. Jewelry sits squarely in this class, alongside the doctors, mechanics, and specialists of their title.

The information gap is not a neutral inconvenience; it invites fraud. Emons (1997) modeled a market where sellers act as the experts who diagnose what the buyer needs, and showed that this dual role gives them a standing incentive to over-treat and over-charge unless something disciplines them. Akerlof (1970) had already shown the darker endpoint: unchecked information asymmetry can unravel a market entirely, as buyers who cannot tell good from bad refuse to pay for quality they cannot confirm. The counterweight, in Spence’s (1973) signaling model, is a costly signal the honest seller can send and the dishonest one cannot cheaply imitate. In modern retail the dominant such signal is the accumulated review. Luca (2016), studying Yelp, found that a one-star increase in rating raised revenue by five to nine percent, and, tellingly for this study, that the effect was driven by independent businesses rather than chains, whose brand already supplies the trust that reviews otherwise provide. For a credence good sold by an independent, the review is not decoration. It is the eligibility signal.

How buyers move through that information is also well studied. Pirolli and Card (1999) described information foraging: people follow the strongest scent of value for the least effort and abandon a trail that does not reward them quickly. Iyengar and Lepper (2000) showed that too much choice can demotivate a buyer into inaction, which matters in a category where an online seller can present tens of thousands of stones. Verhoef, Neslin and Vroomen (2007) documented the research-shopper phenomenon, the now-standard pattern of researching in one channel and buying in another, which in jewelry usually means researching online and completing in a showroom for the high-ticket piece. Offline still accounted for roughly 83.9 percent of global jewelry revenue in 2025 (industry press, 2025), even as online grew fastest. The through-line from Nelson to Verhoef is consistent: for this good, the search happens on a screen and the trust that closes it is borrowed from other people.

The newest layer is the generative answer, and it changes what a signal has to reach. Aggarwal and colleagues (2024), in the GEO paper presented at KDD, showed that visibility inside an engine-generated answer responds to specific, largely content-side levers, and that being cited by the answer is a different contest from ranking beneath it. That distinction is the hinge of our findings: a review a human reads and a review an engine can parse are not the same asset. The US jewelry-store industry where all of this plays out is large and fragmented. IBISWorld counted roughly 74,000 US jewelry-store establishments in 2025 and again in 2026, under NAICS 44831, with no single chain dominating and the great majority small, owner-run shops. Fragmentation is the independent’s context and, as the data will show, partly its advantage.

Findings

The clearest result is that the machine’s local answer for jewelry is neither a chain answer nor a marketplace answer. It is a local answer. Across all 25 prompt-and-metro cells, Google returned a map three-pack every time, and those packs were overwhelmingly filled by independent local shops. Of 75 three-pack slots captured, 70 (93.3 percent) went to independent local jewelers, 5 (6.7 percent) to national chains, and none to online marketplaces or directories. The shops named were real and deeply reviewed: review depth across the 75 pack entries ran to a median of 323 ratings, a mean of 614, and a maximum of 3,200, with visible star scores clustered between 4.7 and 5.0. Names recurred across prompts within a metro, which is what a stable, review-earned local presence looks like.

The organic top ten told a quieter version of the same story. Of 244 organic slots, independent jewelers held about 53.7 percent, national chains about 20.5 percent, directories and editorial listicles about 25.0 percent (Yelp alone appearing 27 times), and online marketplaces under one percent. We label the organic composition directional because classifying a domain as independent, chain, or directory at scale is imperfect, and several online-only brands and local-news pages sit near the boundary. Even so, the shape holds: below the pack the independent still leads, but the directory and the listicle claim a full quarter of the page that the pack denied them.

The AI Overview is where the terrain shifts. Google generated an AI Overview for only 2 of the 25 queries we ran, and both were the ambiguous, non-local prompt "best jewelry store". The four "near me" and category prompts, among them "engagement rings near me" and "jewelry repair near me", triggered no AI Overview in any of the five metros in this capture. For jewelry, that is to say, Google still answers the local question with the map pack, not with a generated paragraph. But when it did generate one, the citations parted ways with the pack: of 26 source citations across the two firing answers, 80.8 percent pointed to third-party directories and best-of listicles, 15.4 percent to independent jewelers, 3.8 percent to a chain, and none to a marketplace. The generated answer, when it appears, is assembled from lists about jewelers rather than from the jewelers.

Demand sits underneath all of this and points the same way. The category’s buyer language is generic and location-anchored, not identity-anchored. "Jewelry store near me" draws about 673,000 US searches a month; "jewelers near me" about 74,000; "jewelry repair near me" about 110,000; "engagement rings near me" about 40,500. The word buyers almost never type is "independent": "independent jewelry store" draws about 210 searches a month and "independent jeweler near me" about 110. The framing a local shop might most want to claim carries effectively no demand behind it. The shopper types the generic near-me query, and the map pack answers it.

The AI answer engines: who ChatGPT, Perplexity, and Gemini name

The Google surfaces above are only part of the machine. To read the standalone answer engines the same way, we captured three of them, ChatGPT, Perplexity, and Gemini, with web search on 23 July 2026, running two shopper prompts, one reputational ("the best independent jewelry stores in [metro]") and one purchase-intent ("where should I buy an engagement ring in [metro]"), across New York, Chicago, and Dallas, twice each, for 36 successful captures. The headline runs the AI Overview result in reverse. All three engines named real independent local jewelers, in all three metros, and named a great many of them. ChatGPT was the most concrete: it returned ranked shortlists of named shops with addresses, star scores, and review counts, linked to the jewelers’ own websites, and named on the order of forty distinct independents across the runs. Its lists admitted a single national name, almost always Tiffany and Co., on the engagement-ring prompt, and were otherwise independent throughout.

Perplexity named independents just as freely in its prose, Adornment and Theory and Steve Quick in Chicago, Eiseman Jewels and Matthew Trent in Dallas, Greenwich St. Jewelers and Catbird in New York, but its citations retold the AI Overview story one layer down. Of its 169 source citations, the largest share pointed to community threads and editorial lists rather than to the shops themselves: Reddit alone accounted for 42, with The Knot, D Magazine, CultureMap, Time Out, and CBS local guides close behind. It anchored the purchase-intent answer on shopping districts, Jewelers Row and the Diamond District, and folded in chains such as Diamonds Direct, Robbins Brothers, and Zales more readily than the other two. Gemini named independents almost exclusively, with owner names, neighborhoods, and character, and framed them outright as a way to "bypass" Tiffany and Cartier. Yet Gemini masked all 347 of its citations behind vertexaisearch.cloud.google.com redirect URLs, a single opaque domain, so its sources cannot be classified at all. Its named businesses had to be read from its answer text, and that opacity is itself a finding.

Across the two runs the named set held rather than reshuffling for every engine, the more reassuring end of the non-determinism scale. ChatGPT’s purchase-intent lists were near-identical run to run; the reputational lists and the other engines kept a stable core and rotated the tail. Gemini repeated its top recommendations but shifted some founder and owner attributions between runs, and at least one reading appeared simply wrong, a reminder that a fluent synthetic answer is not a verified one. The three engines converged on a handful of anchor independents in each metro, Greenwich St. Jewelers, Catbird, and Mociun in New York, Eiseman Jewels and Shapiro Diamonds in Dallas, Ethan Lord and Adornment and Theory in Chicago, while each also named a long tail the others did not.

Read against the study’s thesis, the split is instructive. In the map pack, and now in these three chat engines, the independent is named; in Google’s own generated Overview it mostly is not, and the difference is which sources each surface reads. The forward risk is the one the Overview already exposed. Where an engine builds its answer from directories and lists, as Perplexity visibly did, a shop missing from those lists is missing from the answer; where an engine hides its sources, as Gemini did, a shop cannot even see why it was or was not named. Being named by the chatbot today, like winning the pack, is earned; staying named as the engines change their sourcing is the work.

Discussion

Put the surfaces together and the independent jeweler’s position is more complicated than the usual story of the local shop crushed by national brands. In the surface that still answers most local jewelry queries, the map pack, the independent is dominant, and the mechanism is exactly the one the literature predicts. Because the buyer cannot verify a credence good, buyer and machine both fall back on the strongest available trust proxy, the depth and quality of reviews. Luca’s finding that review effects run through independents rather than chains shows up here as structure: the packs are full of shops with hundreds or thousands of ratings, and that accumulated trust is what qualifies them into the answer. The wider review economics point the same way. Chevalier and Mayzlin (2006) found that a better review profile lifts a product’s relative sales and that negative reviews weigh more heavily than positive ones, while Anderson and Magruder (2012), using a regression-discontinuity design on Yelp, showed that an extra half-star made a restaurant sell out markedly more often, with the effect largest where other information was scarce. Jewelry is precisely such an information-scarce purchase, and the review is its eligibility signal. The independents that have earned it are winning the surface that matters most today.

The exposure sits one layer up and one layer ahead. Where Google generates an answer rather than a pack, it currently assembles that answer from directories and best-of lists, not from the jewelers, so a shop’s presence in the generated answer depends less on its own site than on whether third parties have listed it. The GEO literature frames this exactly: being cited by an engine is a distinct contest from ranking beneath it, responsive to different levers. For now the contest is small in jewelry, since Overviews fired for only 2 of the 25 local queries we ran, but the direction of travel across categories is toward more generated answers, not fewer. There is a second-order trap in the mechanism. A signal the shop cannot directly control, its place on someone else’s list, now gates part of its visibility, and the same opacity that let Gemini mask 347 citations denies the shop any read on why it was named or passed over. The independent invisible to the directory layer today is invisible to the generated answer tomorrow.

This is why we measure visibility as one thing across every surface rather than as a rank on one of them. The RavenEye model reads the terrain as a Corpus, the map of where a market’s attention actually sits, then works the surfaces through Search Surface Optimization, the single method that engineers the map pack, the organic page, and the generated answer together, and reports the result as one number, the Machine-Readiness Score. For jewelry the Corpus reading is unusually clear. Attention concentrates on generic near-me demand answered by a review-driven map pack, with a small but structurally different generated layer fed by directories. A shop that pours effort into the word "independent" or into ranking a blog post is polishing a surface buyers do not use. The advantage lies in three moves the data points to directly.

First, win the generic near-me query on the map surface, because that is where the demand and the eligibility both sit, which means an accurate and complete Google Business Profile and a steady flow of legitimate reviews from real customers. Second, earn a place on the third-party lists that feed the generated answer, since that is where citation currently comes from, a digital-PR and directory-hygiene task rather than an on-site one. Third, close the machine-readability gap our small audit surfaced, so that when the generated layer does read a jeweler directly, it can resolve who they are and where they are. None of these is a promise of a ranking. They are the levers the terrain rewards.

Implications for retailers

For an independent jeweler the practical reading of this study is encouraging and specific. In the surface that answers most local jewelry searches today, the map pack, independents like you took more than nine in ten of the slots we measured across five major metros, and they earned them on review depth rather than ad budget. The near-me demand is enormous and generic, which makes the contest winnable by any shop that is accurately represented and genuinely well reviewed. The work is not glamorous, but it is legible: a Google Business Profile that is complete and correct, business facts that match across every directory, and a compliant system for earning and answering reviews from real customers only.

The forward-looking work is different in kind and worth starting before it turns urgent. The generated-answer layer is small in jewelry today, but it draws on directories and lists rather than on your own site, so the task is twofold: be present and consistent on the third-party surfaces that feed it, and fix the machine-readable signals on your own site so that when the layer reads you directly it can resolve who and where you are. Rankings and citations are decided by the engines, not by any firm. What can be done is to measure where you stand across all of these surfaces, fix the gaps in priority order, and hold the position as reviews decay and engines change.