Sample audit · our own site
The Gem Tray: who AI recommends instead
The Gem Tray is a jewelry guide site we run: moissanite and lab-grown gems, with Amazon picks. Before selling audits to anyone, we ran one on ourselves. This is a short edition (5 questions instead of 25), unedited: the numbers are not flattering.
What each assistant said
Names are the businesses or products recommended. Sources are the sites the assistant showed as citations.
| Question | ChatGPT | Perplexity | Copilot | Gem Tray |
|---|---|---|---|---|
| best moissanite earrings on amazon | Named TTR and SHASHI studs. Source: Shopbop listings. | Named IMOLOVE, BOYA, AINUOSHI, MDFUN. Sources: Amazon listings only. | Named GEMQUEEN, IMOLOVE, QUEENOU. Sources: Amazon, Reddit, BestProductsReviews.com, wellwhisk.com, Parade, elegantmoissanite.com. | Not named |
| is moissanite from amazon real | Sources: Gems Lore, Gemological Research Association, The Real Jewelry Company. | Sources: Amazon Seller Central, GIA. | No answer returned. | Not named |
| moissanite vs lab grown diamond which should i buy | Did not search. Answered from memory with no sources. | Sources: FTC, GIA, Grand Diamonds. | Not run. | Not named |
| is a GRA certificate for moissanite legit | Sources: a GRA report-lookup site, GIA, Gemmological Association of Australia, GRA. | Sources: gra-cert, Adamans, Amaral Jewelry. | Not run. | Not named |
| best affordable lab created sapphire or ruby jewelry | Named Macy's, Nordstrom Rack (Delmar), Kohl's (Gemminded), Mejuri, Zales, JTV. Sources: TheStreet, JTV. | Named JTV, Macy's, Ross-Simons, Kay, Etsy. | Not run. | Not named |
What this tells us
"On Amazon" questions go straight to the listings. Perplexity and ChatGPT cited store listings, not guides. A guide site cannot replace a listing there. Copilot was different: it cited three review and list sites (BestProductsReviews.com, wellwhisk.com, Parade) next to Amazon. That is a lane a guide can enter, but Copilot draws on Bing, so the page has to rank in Bing first.
Trust questions are where small sites win. For "is a GRA certificate legit", Perplexity cited two small jewelers and a certificate site, not big publishers. Both assistants also leaned on GIA and the FTC, named outside authorities, again and again.
Some questions have no sources to compete for. ChatGPT answered the moissanite vs lab diamond question from memory. No page change wins a search that never happens; the only lever there is being mentioned widely enough over time to be part of what the model already knows.
Broad retail questions belong to big retailers. The affordable lab sapphire question went to Macy's, JTV, Nordstrom Rack, Kay and Zales. A new site should not spend effort here yet.
The fix list, in order
1. Make the trust pages answer first, with a named outside source
The site already has pages for "is moissanite on Amazon real" and "is a GRA certificate real". The first paragraph under each heading should give the answer in two or three sentences and name GIA or the FTC next to the claim, because those are the sources both assistants reached for. Effort: an hour. Payoff: highest, since small sites already win these questions.
2. Get the buying guides into Bing
Copilot cites list-style guides for "best moissanite earrings on Amazon", but only ones Bing already ranks. Submit every guide to Bing Webmaster Tools and IndexNow, and keep the page title an exact match for the question. Effort: minutes. Payoff: opens the one lane on "on Amazon" questions that is not listings only.
3. Earn mentions where the assistants already look
Reddit and review roundups showed up as sources. A genuinely useful answer from a real owner of the jewelry, in the places buyers already ask, is how a name enters both search results and the model's memory. Effort: ongoing. Payoff: slow, but it is the only route for questions answered from memory.
4. Leave broad retail questions alone for now
Retailers own "affordable lab sapphire jewelry". Revisit after fixes 1 to 3 have had 30 days, when the site is indexed and has history.
What this sample leaves out compared with a full audit: 20 more questions, a per-rival count across all 75 answers, and a rerun to separate day-to-day noise from real patterns. Copilot answered one of the two questions we tried; the other returned nothing, so three rows are marked not run. We will rerun this audit in 30 days and publish what changed.