Rings.com Research & Data

Rings.com Research  ·  Edition 5

Rare Carat Doubled Its AI Visibility Score in One Month, and Brilliant Earth's Lead Held

---

September 30, 2026  ·  OpenAI, Gemini, Perplexity, Claude + Google AI Overviews  ·  30 queries  ·  36 brands scored  ·  2,400 responses + 600 AI Overviews

Five months in, this study recorded the largest single-month move it has ever measured, in points and in standard deviations. Rare Carat's weighted visibility score, the study's position-weighted measure (10 points for a first mention, down to 1), rose from 3,494 in August to 7,051 in September, an increase of 3,557 points (+102%) that roughly doubled it, lifted it from #8 to #4, and sits 31.4 standard deviations from its August level on the study's noise test (the next-largest move this month is 9.5; in percent terms Keyzar's rise from a small base, +638%, is larger). The score is not a mention count: raw mentions rose 63% (from 530 to 862), and the number of responses listing Rare Carat first rose from 54 to 336, mostly in Claude (18 to 172) and Gemini (26 to 93).

Excluding Perplexity, Rare Carat's score rose 2,611 to 4,421, +69%, not +102%. Perplexity's answers also became far more uniform across runs (Rare Carat appeared in all 20 runs for 13 queries in September versus 2 in August), and because the noise test measures same-session scatter, its standard deviation understates variation for that surface. The score rose in three models (Perplexity +1,747 points, 36.4 SD; Gemini +941, 10.8 SD; Claude +810, 17.1 SD) and marginally in OpenAI (10 mentions, 59 points, 3.3 SD, supporting evidence only); these per-model tests are post hoc. This study measures what the models say, not why, and offers no explanation.

Brilliant Earth kept the top weighted score over Blue Nile, 15,860 to 15,448; the 412-point gap is 3.73 standard deviations on the study's committed unpaired formula (August 4.17), marginally beyond the strict bar. A secondary paired per-run test gives 4.19 standard errors of the mean per-run gap, the basis on which Edition 4 reported 4.7 for August; the change between the two months' gaps is within noise. The mall chains rearranged again, and Google AI Overviews fired on 564 of 600 query executions, up from 520; most of that rise (34 of 44) is in overviews returned without source lists.

Rare Carat's One-Month Doubling

Rare Carat's weighted score was 3,494 in August, eighth on the leaderboard. In September it was 7,051, fourth, ahead of VRAI (4,834), Ritani (4,352), Whiteflash (3,984) and Tiffany & Co. (3,560). Both months collected 20 runs, so the raw change is directly comparable: +3,557 points, or +1,778.5 in 10-run-equivalent terms. Its mentions rose from 530 to 862, and its share of all brand mentions from 4.3% to 6.7%.

On the study's noise test, which scales each brand's change by its own run-to-run scatter, the month-over-month move is +31.4 standard deviations. The strict bar is 3.5. The next-largest month-over-month move this edition is Keyzar's +9.5, and the largest in any earlier edition's audit was 9.4 (Shane Co., July). Measured against May, Rare Carat is +2,003.5 equivalent points at 28.9 SD, also beyond the strict bar. This is a one-month observation that is beyond measured noise; whether it holds in later months is something only later editions can show.

The increase appears in three models and marginally in OpenAI, and is concentrated: Perplexity accounts for 206 of the 332 added mentions (62%), Gemini for 85, Claude for 31, and OpenAI, which mentioned Rare Carat zero times in August, for 10 (a change of 59 weighted points at 3.3 SD, supporting evidence only). Rare Carat was named in responses to 24 of the 30 queries (21 in August). Mentions rose for 17 queries, were unchanged for 1, fell for 6, and were zero in both months for 6; mentions in the ethical-sourcing cluster fell from 30 to 17. The largest single query accounts for 7.5% of its mentions (at most 7.5% for any query). In weighted points the gain is Perplexity +1,747 (49% of it; 36.4 SD), Gemini +941 (10.8 SD), Claude +810 (17.1 SD) and OpenAI +59 (3.3 SD, supporting only); these per-model tests are post hoc; excluding Perplexity the score rose from 2,611 to 4,421, +69%. Perplexity's Rare Carat answers also became near-deterministic: Rare Carat appeared in all 20 runs for 13 queries (versus 2 in August), in none for 12 (19 in August) and in some runs for 5 (9); its per-run Perplexity count ran 15 to 18 in every run (August 4 to 8), and 275 of its 332 Perplexity responses naming Rare Carat cite rarecarat.com (August 74 of 126). The noise test's standard deviation measures same-session scatter, so it understates variation for a surface that behaves this uniformly within a session. Because the weighted score rewards earlier placement, it rose faster than the mention count: Rare Carat was the first brand listed in 336 responses (versus 54 in August), and its average position when mentioned moved from 4.4 to 2.8; the first-listed increase is mostly in Claude (18 to 172) and Gemini (26 to 93), with Perplexity 10 to 71. The per-query counts are in the expandable table below the summary.

Rare Carat	August	September
Weighted score	3,494	7,051
Weighted rank	8	4
Total mentions	530	862
Claude mentions	241	272
Gemini mentions	163	248
OpenAI mentions	0	10
Perplexity mentions	126	332
AI Overview mentions (in overviews with a source list)	99 (99)	132 (125)
General purchase queries (mentions)	227	407
Lab-grown diamond queries (mentions)	105	193
Ethical sourcing queries (mentions)	30	17
Custom and gifting queries (mentions)	3	64
Price and value queries (mentions)	165	181
Queries with at least one mention (of 30)	21	24
Largest single query, share of its mentions	53 of 530 (10.0%)	65 of 862 (7.5%)
Responses listing Rare Carat first	54	336
Average position when mentioned	4.4	2.8

Mentions are counts of responses (out of 600 per model per month) that name Rare Carat; 20 runs in each month. Cluster rows use the study's five query clusters (see Methodology). Mentions by cluster sum to the monthly total.

Show Rare Carat mentions by query ↓

Query	Aug	Sep
which online jeweler has the best diamond quality	35	65
best lab-grown diamond retailers	28	60
which online diamond retailer has the best reputation	30	60
best place to buy an engagement ring online	27	56
best place to buy lab-grown diamond rings online	19	56
where to get the best value diamond ring	45	56
cheapest place to buy a real diamond ring	41	53
affordable diamond engagement rings online	52	52
best place to buy an engagement ring	21	50
where should I buy a diamond ring online	37	50
where to buy a wedding band online	20	41
most reputable online diamond retailers	53	40
where should I buy a lab-grown diamond	53	39
lab-grown diamond vs natural diamond where to buy	5	38
best place to customize an engagement ring	0	22
where to buy a custom engagement ring	3	22
best jewelry brands for engagement rings	3	20
where to buy a wedding ring set	0	20
best alternative to Tiffany for engagement rings	0	17
best engagement ring for the money	22	16
where to buy ethically sourced diamonds	21	12
best online jewelry stores for rings	1	8
best conflict-free diamond retailers	9	5
best engagement rings under $3000	5	4
best websites to buy fine jewelry	0	0
is it worth buying a lab-grown diamond engagement ring	0	0
most ethical jewelry brands	0	0
sustainable jewelry brands for engagement rings	0	0
ethical engagement ring brands	0	0
best jewelry brands for anniversary gifts	0	0

Number of responses (across 4 models and 20 runs, 80 responses per query) that name Rare Carat. Totals: 530 in August, 862 in September.

What this study can and cannot say. It can say that the increase is large and beyond measured noise, that it is most concentrated in Perplexity, and that it moved Rare Carat's placement earlier as well as making it more frequent. It cannot say why. Several of this month's largest anomalies are concentrated in Perplexity: its Rare Carat mentions (126 to 332), its Blue Nile lab-grown co-occurrence (182 to 325 responses), and its Reddit citations (596 in July, 1,206 in August, 20 in September); its total captured citations (6,461, 11,743, 6,000) are back at July's level, so August was the odd month for that measure. A reader may reasonably consider whether something on the Perplexity side changed; this study cannot tell. The chat-citation data below also shows rarecarat.com cited far more often in the same month, and Google AI Overviews mention Rare Carat more often as well. Those facts are reported side by side, and this study does not test whether any of them relates to another.

Brilliant Earth's Lead Held

Brilliant Earth finished with 15,860 weighted points and 1,910 mentions, Blue Nile with 15,448 and 1,809. The gap was 412 raw points (206 in 10-run-equivalent terms). On the study's committed unpaired formula, which scales each brand's own run-to-run noise, the September gap is 3.73 standard deviations (August 4.17; Edition 4 reported 4.7 for August on the paired basis), marginally beyond the strict bar of 3.5. As a secondary check, measured directly on the per-run gap across all 20 September runs, it averages 20.6 points per run with a standard deviation of 22.0, which is 4.19 standard errors of the mean per-run gap (equivalently, 4.19 SD of the gap total; the 22.0 is the SD of the per-run gap, not the denominator of that ratio); that paired test has a different null from the since-May change test used elsewhere in this report. August's gap was 499 raw points (249.5 equivalent). The gap was 412 points in September versus 499 in August, a change within noise (-0.6 SD on a paired per-run test of the two months' gaps). July's gap ran the other way: Blue Nile was ahead by 146 raw points (73 equivalent points, 1.3 standard deviations on the paired basis, 1.19 unpaired), which this report called a statistical tie. Brilliant Earth also leads raw mentions, 1,910 to 1,809.

Both leaders scored lower than in August: Brilliant Earth fell 542 points (-4.7 SD) and Blue Nile fell 455 (-4.0 SD); both month-over-month moves are beyond the strict bar. Measured against May, Blue Nile remains up beyond the strict bar (+391 equivalent points, 6.5 SD)† (the dagger marks a path since May that was not one-directional; see the across-editions table), while Brilliant Earth's rise since May (+349, 3.1 SD)† no longer clears the strict bar and is supporting evidence only (in August it was 5.4 SD). These are one-month observations; the two leaders' scores in the next edition will show whether September's lower level holds.

James Allen holds third place again with 10,683 points and 1,295 mentions. Its change versus August is -21 points (-0.2 SD), within noise, after August's -360.5 equivalent-point move (5.8 SD). Versus May it is -812.5 equivalent points (-5.5 SD), beyond the strict bar†.

The Mall Chains, One by One, Again

Edition 3 led with the mall jewelry chains fading from AI chat answers, all four beyond the strict noise bar versus May. Edition 4 reported that the finding had split: Kay and Helzberg still down, Jared partly recovered, Zales back above its May level. September rearranges the four chains again, and honest scorekeeping means reporting each.

Kay Jewelers rose 567 points versus August (+283.5 equivalent, +6.9 SD, beyond the strict bar), from 175 to 285 mentions. The increase shows up in Claude (6 → 62), Gemini (66 → 100), Perplexity (15 → 29) and OpenAI (88 → 94). Kay is now 13.0% below its May level at -1.9 SD, within measured noise; in Edition 4 it stood 7.0 SD below May. Zales fell back: -369 points (-3.9 SD, beyond the strict bar) after its August rebound of 7.3 SD, even though its mention count barely moved (455 to 440); its Perplexity mentions went from 116 to 57. It is now 2.9% below May (-0.7 SD, within noise). Jared fell 426 points (-4.7 SD, beyond the strict bar) and is 44.9% below May at -5.9 SD, back beyond the strict bar after Edition 4 had it at supporting level (2.9 SD). Helzberg fell 186 points (-3.2 SD, supporting evidence) and is 55.5% below May (-9.0 SD, beyond the strict bar).

The tally versus May is two and two: Jared and Helzberg remain beyond the strict bar below their May levels, while Kay and Zales are back within noise of theirs. Zales' August rebound did not hold in full, and its September level is statistically indistinguishable from May's. Jared's fall, Zales' Perplexity fall and the drop in Perplexity's mall-chain share (11.2% to 5.4%) are mostly Perplexity-side: Jared's Perplexity mentions went from 99 to 21 and Zales' from 116 to 57, the same surface where Rare Carat's answers changed most. The aggregate agrees with the softened picture. The four chains' combined share of all brand mentions ran 9.8% (May), 8.9% (June), 7.7% (July), 7.9% (August) and 8.1% (September: 1,043 of 12,887 mentions). Why individual chains move is a question this study can pose but not answer.

Weighted score by edition in 10-run-equivalent terms (*July, August and September collected 20 runs; shown as raw score divided by 2). The last two columns are the change since August and since May in standard deviations of each brand's own measured noise; bold marks the strict bar (3.5), grey marks within noise (under 2). Mall chains: Zales, Jared, Kay Jewelers, Helzberg.

Mentions, Aug → Sep	Claude	Gemini	OpenAI	Perplexity
Zales	121 → 139	115 → 137	103 → 107	116 → 57
Jared	5 → 53	87 → 112	40 → 36	99 → 21
Kay Jewelers	6 → 62	66 → 100	88 → 94	15 → 29
Helzberg	1 → 0	50 → 48	3 → 8	67 → 40

Number of responses naming each chain, by model, August then September (600 responses per model per month).

The Overall Leaderboard

Raw totals below reflect 20 collection runs (2,400 scored responses). Because August also collected 20 runs, the change column is stated in raw weighted points and compares directly. Bold changes exceed the strict noise bar (3.5 standard deviations) for a single month; the 13 brands that do are listed with their SDs in the largest-moves section below.

Brand	Mentions	Weighted	vs Aug
Brilliant Earth	1,910	15,860	-542
Blue Nile	1,809	15,448	-455
James Allen	1,295	10,683	-21
Rare Carat	862	7,051	+3,557
VRAI	757	4,834	+229
Ritani	718	4,352	-339
Whiteflash	572	3,984	+314
Tiffany & Co.	495	3,560	-510
Clean Origin	462	3,005	-147
Zales	440	2,461	-369
Mejuri	334	2,293	-308
Cartier	325	2,290	+86
Grown Brilliance	318	2,141	+190
Bario Neal	224	1,836	-35
Kay Jewelers	285	1,491	+567
With Clarity	236	1,323	-174
Harry Winston	185	1,305	+86
Jared	222	1,081	-426
Aurate	164	1,029	+219
Do Amore	122	867	+114
Taylor & Hart	98	682	+43
Shane Co.	144	643	-392
De Beers	112	633	+191
Costco	113	616	+181
Helzberg	96	568	-186
Pandora	100	566	+261
Angara	98	544	+154
David Yurman	75	525	-131
Brian Gavin Diamonds	71	458	-47
Frank Darling	76	397	-332
Tacori	65	353	+33
Keyzar	38	214	+185
Swarovski	37	168	+26
Leibish & Co.	23	119	+58
Lightbox	5	23	-11
Hearts on Fire	1	6	+6

Show all 36 brands

All 36 brands in the leaderboard recorded at least one mention across 2,400 scored core responses (20 runs, September 2026). Weighted score: 1st mention = 10 pts, scaling to 1 pt for 10th or later. "vs Aug" is the change in raw weighted score; both months collected 20 runs, so no normalization is applied here (the across-editions table below uses 10-run equivalents, which are half of raw scores). Hearts on Fire, which recorded zero mentions in August, recorded 1 in September (weighted score 6). The four Canadian retailers and Rings.com, also tracked, recorded none (see the methodology).

Leaderboard Across Editions: May, June, July, August, September

This study runs every month. The table below tracks the weighted score of the top brands across every edition; a new column is added each month, with the 20-run months (July, August, September) stated in 10-run-equivalent terms so all five columns are directly comparable. The final column reports the noise test: whether the brand's change since May exceeds its own measured run-to-run noise, at a strict bar of 3.5 standard deviations or a supporting bar of 2. With a fifth month of data, the noise test now runs over 80 collection runs.

Brand	May	Jun	Jul*	Aug*	Sep*	Change vs May
Brilliant Earth	7,581	7,612	7,776	8,201	7,930	Up (2 SD)†
Blue Nile	7,333	7,350	7,849	7,951.5	7,724	Up (strict bar)†
James Allen	6,154	5,698	5,712.5	5,352	5,341.5	Down (strict bar)†
Rare Carat	1,522	1,593	1,466.5	1,747	3,525.5	Up (strict bar)†
VRAI	2,017	2,206	2,288.5	2,302.5	2,417	Up (strict bar)
Ritani	2,011	1,922	2,335	2,345.5	2,176	Up (2 SD)†
Whiteflash	2,043	2,037	1,906	1,835	1,992	Within noise
Tiffany & Co.	1,965	1,988	1,988.5	2,035	1,780	Down (2 SD)†
Clean Origin	1,385	1,282	1,204.5	1,576	1,502.5	Within noise
Zales	1,267	1,199	1,099	1,415	1,230.5	Within noise
Mejuri	1,107	1,139	1,408	1,300.5	1,146.5	Within noise
Cartier	1,022	1,103	1,094.5	1,102	1,145	Up (2 SD)†

Weighted score by edition for the current top 12 brands. *July, August and September shown as 10-run equivalents (raw score divided by 2) because those months collected 20 runs versus 10 in May and June. The strict bar is a change since May of at least 3.5 standard deviations of the brand's own measured run-to-run noise; at that level, chance across 34 tested brands would be expected to produce about 0.02 qualifying brands, effectively zero. The 2 SD bar is supporting evidence only, since chance would produce 1 or 2. This month 13 of the 34 brands with full five-edition data clear the strict bar, down from 18 in August. Beyond the top 12, the strict-bar movers are: up, Pandora (11.0)†, Grown Brilliance (7.4), Taylor & Hart (6.9)†, Aurate (5.0)†, With Clarity (4.7)†, Do Amore (4.0)†; down, Shane Co. (11.6)†, Helzberg (9.0)†, Jared (5.9)†. At the 2 SD bar: up, Bario Neal (3.2)†, Tacori (3.0)†, Costco (2.3)†; down, Harry Winston (3.0)†, Brian Gavin Diamonds (2.7)†. † marks a path that was not one-directional across the five editions; the test measures where a brand stands versus May, not the shape of the path. Kay Jewelers (-1.9) and Zales (-0.7) are within noise of May this month; in August Kay tested beyond the strict bar below May and Zales at the supporting level above it. The earlier calls were correct on their data and September's data revised them. Both are reported.

This Month's Largest Moves

Single-month movers in raw weighted points versus August; both endpoints were collected at the same 20-run depth. The largest increase (Rare Carat, +3,557, +31.4 SD) is described in the first section. The largest decrease belongs to Brilliant Earth (-542, -4.7 SD), followed by Tiffany & Co. (-510, -4.9 SD) and Blue Nile (-455, -4.0 SD).

Change in raw weighted score versus the August 2026 edition, listed by size. Not every listed move clears the strict bar: Whiteflash at 3.1 SD, VRAI at 2.3 SD, Aurate at 3.3 SD are supporting evidence, not beyond the strict bar, for a single month.

Beyond the strict bar vs August	Aug	Sep	Change	SD
Rare Carat	3,494	7,051	+3,557	+31.4
Keyzar	29	214	+185	+9.5
Frank Darling	729	397	-332	-8.5
Pandora	305	566	+261	+8.2
Kay Jewelers	924	1,491	+567	+6.9
Shane Co.	1,035	643	-392	-5.1
Tiffany & Co.	4,070	3,560	-510	-4.9
Jared	1,507	1,081	-426	-4.7
Brilliant Earth	16,402	15,860	-542	-4.7
De Beers	442	633	+191	+4.2
Blue Nile	15,903	15,448	-455	-4.0
Zales	2,830	2,461	-369	-3.9
Angara	390	544	+154	+3.9

Raw weighted scores (20 runs each month). SD is the month-over-month change in standard deviations of the two months' combined measured noise. 13 brands clear the strict bar and 11 more are at the 2 SD supporting level: Mejuri (-3.47); Costco (+3.3); Aurate (+3.3); Helzberg (-3.2); Whiteflash (+3.1); David Yurman (-3.1); Ritani (-2.7); With Clarity (-2.6); Leibish & Co. (+2.6); Grown Brilliance (+2.5); VRAI (+2.3). Chance would be expected to produce about 0.02 strict-bar brands among the 36 brands scored, and 1 or 2 at the supporting level.

Which Model You Ask Still Decides What You See

OpenAI remains the most James-Allen-friendly surface (473 James Allen mentions in 600 responses, versus 483 in August). The other models' James Allen mentions were Claude 278 (317 in August), Gemini 243 (213) and Perplexity 301 (287); these model-level changes were not tested. Claude's total brand mentions rose from 2,960 to 3,376 while the other three models' totals each changed by 2% or less. OpenAI mentioned Rare Carat 10 times in 600 responses (zero in August, once each in June and July; the August-to-September change is supporting evidence only), while Perplexity mentioned it 332 times, Claude 272 and Gemini 248.

Brand	Claude	Gemini	OpenAI	Perplexity
Brilliant Earth	511	445	487	467
Blue Nile	461	447	496	405
James Allen	278	243	473	301
Rare Carat	272	248	10	332
VRAI	243	248	122	144
Ritani	260	274	106	78
Whiteflash	168	200	104	100
Tiffany & Co.	147	103	191	54

Total mentions by model across twenty runs (600 responses per model). Gemini logs the most total brand mentions (3,967) and Perplexity the fewest (2,707); Claude logged 3,376 and OpenAI 2,837. Rare Carat's OpenAI count is highlighted as the lowest in the table.

Brand Category Concentration by Model

August's divergence did not hold, and the order of the models reversed. The mall-chain gap between the highest and lowest model, which ran 4.2x in June, 1.3x in July and 2.5x in August, is 1.8x this month (Gemini 10.0% versus Perplexity 5.4%). Perplexity's mall-chain share fell from 11.2% to 5.4% of its brand mentions while Claude's rose from 4.5% to 7.5%, so the Perplexity-to-Claude ratio went from 2.5x in August to 0.72x. The heritage-luxury gap between the highest and lowest model is 2.6x this month versus 3.0x in August (ChatGPT 15.3% versus Gemini 5.8%). A divergence whose size and direction change from month to month with the same brands underneath is hard to read as a stable property of any one model; that reading is an inference, and the shares are the data.

Model	Heritage luxury, Aug	Heritage luxury, Sep	Mall-chain, Aug	Mall-chain, Sep
ChatGPT (OpenAI)	15.5%	15.3%	8.2%	8.6%
Claude	8.8%	7.5%	4.5%	7.5%
Gemini	7.1%	5.8%	8.0%	10.0%
Perplexity	5.2%	5.9%	11.2%	5.4%

Heritage luxury: Tiffany & Co., Cartier, Harry Winston, David Yurman. Mall-chain: Zales, Jared, Kay Jewelers, Helzberg. Each figure is that category's share of the model's total brand mentions in that edition. Across all four models combined, the mall chains' share of total mentions ran 9.8% (May), 8.9% (June), 7.7% (July), 7.9% (August) and 8.1% (September).

Google AI Overviews: Fire Rate and Leaderboard

Google AI Overviews fired on 564 of 600 query executions (30 queries x 20 runs; 94.0%), up from 520 of 600 (86.7%) in August. Edition 4 reported a second consecutive decline (94.3% in June, 89.8% in July, 86.7% in August) and said every measurement since the baseline had pointed the same way. September does not support that reading: the four readings, 94.3%, 89.8%, 86.7% and 94.0%, do not describe a direction, and that earlier sentence is withdrawn. The September rise needs a caveat about what was captured. Counting only fired overviews that came back with a source list, the figures are 529 (88.2%) in July, 508 (84.7%) in August and 518 (86.3%) in September; per run that is 25.4 (SD 1.93) in August versus 25.9 (SD 1.55) in September, a difference of 0.9 SD, within noise. Fired overviews returned with an empty source list went from 12 to 46, 18 of them for one query ("which online jeweler has the best diamond quality"). So the rise is in overviews returned without source lists, and this study cannot say whether those are overviews whose references were not captured or overviews that differ in some other way.

The top of Google's leaderboard again mirrors chat: Brilliant Earth leads the answer box (382 mentions to Blue Nile's 330). Rare Carat ranks third in Google's box (132 mentions, 125 of them in overviews with a source list, versus 99 and fourth in August), ahead of James Allen (100), the reverse of their order in chat. Ranked on the same basis (by mentions) on both surfaces, two brands rank in Google's top 10 while sitting outside the chat top 10: Bario Neal (8th in Google's box against 16th in chat) and Kay Jewelers (9th against 14th).

Rank	Brand	Overview mentions	Chat rank (by mentions)
1	Brilliant Earth	382	1
2	Blue Nile	330	2
3	Rare Carat	132	4
4	James Allen	100	3
5	Whiteflash	73	7
6	VRAI	67	5
7	Tiffany & Co.	60	8
8	Bario Neal	55	16 ↑
9	Kay Jewelers	51	14 ↑
10	Ritani	50	6

Google AI Overviews leaderboard, ranked by mentions: 564 fired overviews out of 600 collected. "Chat rank" is the brand's rank by mentions among the four chat models this edition, the same basis as the Google rank. Upward arrows mark brands that rank in Google's top 10 but outside the chat top 10.

James Allen's mention rate relative to Blue Nile's is lower in Google AI Overviews (30%) than in chat (72%). Among the chat models, James Allen is mentioned at 72% of Blue Nile's rate this month (1,295 versus 1,809). In Google AI Overviews it runs at 30% (100 versus 330). The same split was 82% versus 51% in June, 78% versus 48% in July and 73% versus 51% in August: the Google ratio is lower than the chat ratio in each of the four editions, and the September ratio is the lowest of the four. That is a difference between two surfaces; this study does not test why.

Sources Cited in Google AI Overviews and Chat Answers

This section carries a capture caveat before it carries a finding. In August the reference lists returned for many overviews arrived as opaque google.com redirect entries rather than resolvable source domains (2,240 of 3,954). In September that share fell to 727 of 4,468 (16%), and in July it was 368 of 5,287. Those entries say nothing about which site Google actually read, so they are excluded from the source analysis below. The remaining capture is closer to July's in volume (3,741 resolvable references versus 4,919) but spread over far fewer distinct domains (199 versus 579), and 46 of 564 fired overviews returned no reference list at all (12 of 520 in August, 10 of 539 in July). Because capture breadth still differs from month to month, this study treats each month's AI Overview source list as a one-month snapshot and does not compare source shares across months. This study's standing practice is to check the instrument before interpreting the reading.

AI Overview references	Jul	Aug	Sep
Fired overviews	539	520	564
References captured	5,287	3,954	4,468
Opaque google.com entries (excluded)	368	2,240	727
Resolvable references	4,919	1,714	3,741
Distinct resolvable domains	579	101	199
Fired overviews with no reference list	10	12	46

Counts from the supplementary (AI Overview) raw files, 600 collected overview calls per month. A reference is one entry in an overview's returned source list; repeated domains are counted each time.

The chat models' own citations, three comparable months. July set the baseline for which web domains the chat models cite and August made the first month-over-month comparison; September is the third month on the same basis for Claude, Gemini and Perplexity (OpenAI's capture remains too sparse to compare: 32 of 600 responses carried any citation). Across 2,400 chat responses the models cited 780 distinct domains 21,069 times, down from 860 and 26,897 in August. The capture-volume caveat comes first: Perplexity's captured citations went from 6,461 in July to 11,743 in August to 6,000 in September, so August's 82% jump was a one-month feature, while Claude's and Gemini's each stayed within 9% of their July levels. Total counts partly reflect how much each model cites, not only whom.

With that stated: forbes.com is again the single largest source (2,677 captured citations). rarecarat.com, which ranked 8th in August, is second (1,090, up from 468), with the increase in Perplexity (198 to 696) and Gemini (140 to 238). reddit.com fell from 1,396 to 282 captured citations, almost entirely inside Perplexity (1,206 to 20; Gemini's 190 to 262), and brilliantearth.com from 1,496 to 761. These are observations only. This study cannot tell whether a change in Perplexity's counts reflects a change in what it retrieves or in what its interface returns as citations, and it does not test whether the rarecarat.com increase relates to Rare Carat's chat visibility in the same month.

Chat citations (captured)	Jul	Aug	Sep
Claude	8,086	8,015	7,659
Gemini	6,583	6,903	7,164
Perplexity	6,461	11,743	6,000
OpenAI (too sparse to compare)	231	236	246
All four models	21,361	26,897	21,069
Distinct domains	887	860	780
forbes.com	2,287	3,147	2,677
brilliantearth.com	1,114	1,496	761
reddit.com	820	1,396	282
rarecarat.com	325	468	1,090
reddit.com, Perplexity only	596	1,206	20
rarecarat.com, Perplexity only	92	198	696

Captured citations are entries in the cited-domains field the four chat integrations return with each answer (pipe-delimited; repeated domains counted each time). Capture semantics differ by model and are not merged across models.

Across both surfaces, the most-cited domains are a mix of community sites, a small set of large publishers, and brand-owned pages. That describes what is cited, not what causes a brand to be recommended: this study has no test of that, and the month's source data moved enough (Perplexity's volumes, Reddit, rarecarat.com) to warrant humility about any such explanation.

The Stale Map Held: James Allen Flat, Blue Nile's Lab-Grown Pairing Up, The Clear Cut Back

Month five of tracking the market-structure gap uses the corrected measure introduced in Edition 4. The rate this series tracks counts responses that include James Allen at all, whatever they say about it. On that like-for-like basis, James Allen appeared in 1,295 of 2,400 responses this month (54.0%), versus 1,300 (54.2%) in August, 57.0% in July and 57.1% in June. Edition 4 called August's reading the first decline in the series and asked whether the rate would fall further. September's answer is no: the change is 5 responses, and on per-run counts (64.75 per run in September versus 65.00 in August) it is 0.3 SD, within noise. The rate is flat, not falling.

The sharper signal is what those responses say. A keyword screen (its exact pattern is published in the methodology) finds an acknowledgment of the consolidation within 300 characters of a James Allen mention in 120 of the 1,295 responses (9.3%), so 1,175 (90.7%) show no acknowledgment the screen can detect. The screen misses some real acknowledgments (6 further responses use "(formerly) James Allen", "sister brand" or "transition" wording); counting those, the true no-acknowledgment share is nearer 90.3%. Applied to August, the same screen finds 149 of 1,300 (11.5%), consistent with Edition 4's "nearly nine in ten" reading of that month. The acknowledgments come almost entirely from Gemini (70) and Claude (48); Perplexity (1) and OpenAI (1) account for the rest. The screen counts only explicit acknowledgments near a James Allen mention, so it is a floor on awareness and not a measure of accuracy.

Responses pairing Blue Nile with lab-grown framing rose to 38.9% (934 responses), versus 32.8% in August, 33.8% in July and 31.8% in June. Edition 4 called the earlier changes a wobble; September's change is larger: 46.7 responses per run in September versus 39.3 in August, or 6.8 SD, beyond measured noise as a one-month observation. The flag is a text pattern (published in the methodology) that fires when "Blue Nile" and a "lab-grown" spelling appear within 300 characters of each other, so it counts co-occurrence, not accuracy. The increase is concentrated in Perplexity (182 to 325 responses; Gemini 248 to 265, Claude 227 to 212, OpenAI 129 to 132), and this study offers no explanation for the change. The Clear Cut, which Signet acquired in late May, was named in 5 responses (4 Perplexity, 1 Claude), after none in August; none of the 5 mentions Signet or Blue Nile. These are methodology flags, not scoring changes: scores are preserved exactly as the models cited them.

In all, 1,595 of 2,400 responses (66.5%) triggered at least one of the three flags, versus 1,546 (64.4%) in August. Because two of the flags fire on mentions and one on co-occurrence, that is an upper bound on responses carrying a stale description, not a measure of it. The honest reading is the one Edition 4 reached: James Allen appears in about the same share of answers as last month, and when it appears most answers do not reflect the consolidation. Whether merger-aware answers grow beyond their roughly one-in-ten share is a question the next edition can test.

Methodology

Rings.com tracked 36 jewelry brands (plus, for coverage and disclosure, four Canadian retailers and Rings.com itself, none recorded a mention this edition) across 30 queries in five intent clusters (general purchase, lab-grown diamonds, ethical sourcing, custom and gifting, and price-value) using four AI models with live web search enabled: OpenAI (gpt-4o-2024-08-06), Google Gemini (gemini-2.5-flash), Perplexity (sonar), and Anthropic's Claude (claude-sonnet-4-6, with web search); model versions are as recorded in the raw files. Twenty collection runs were conducted in September 2026, matching July's and August's run counts. Raw run timestamps are recorded in UTC. All 20 runs were collected in one session, between 2026-09-30 21:02 and 2026-10-01 01:43 UTC (about 4.7 hours, one afternoon and evening, Pacific time), so the within-month noise measures same-session scatter, not day-to-day variation (August's runs were also collected on a single day). A supplementary source, Google AI Overviews, was collected across the same 30 queries and 20 runs, firing on 564 of 600 query executions (94.0%).

Brand appearances were scored by position. A first mention in a response earned 10 points, scaling down to 1 point for a tenth mention or later. This formula is unchanged across all five editions. Cross-edition comparisons use the 10-run equivalent (raw score divided by two for the 20-run months); because July, August and September all collected 20 runs, month-over-month deltas among those three are computed from months of equal depth, and their raw values are exactly twice the equivalent figures. The measured noise model introduced in Edition 3 now runs over all 80 collection runs: for every brand, we compute the standard deviation of its per-run weighted score within each month, scale it to the monthly total, and combine the noise of two months as the square root of the sum of their variances. A movement since May is claimed at 3.5 standard deviations of that combined noise (the strict bar; expected false positives across 34 tested brands, about 0.02) and noted as supporting evidence at 2 standard deviations (where chance would produce 1 or 2). Those chance rates treat each brand's monthly totals as independent and approximately normal; in practice brands are not fully independent, because a single response can mention several brands at once. We accept that simplification and state it here. May's per-brand noise estimates, carried forward from the Edition 3 audit, are stored at whole-point precision, which moves a standardized change by at most about 0.2 SD and does not change any strict or supporting classification. The same limits stated in Edition 3 still apply: May's and June's runs were spread across more calendar days than July's, August's or September's, which inflates their scatter estimates and makes the test conservative, and the test measures where a brand stands versus May, not whether its path was smooth. The same per-run scatter approach was applied this month to non-brand measures and sub-measures: the gap between the two leaders (a paired per-run test, and a paired per-run test of the change in that gap between months), Rare Carat's weighted score in each of the four models, the number of AI Overviews that fired with a source list per run, the number of Blue Nile lab-grown flags per run, and the number of responses including James Allen per run. These are reported as one-month observations, were not part of the test's original scope, and were chosen after seeing the data; the 3.5 bar was calibrated for 34 simultaneous brand tests, so they carry less weight than the brand results.

Data-handling events in September. (1) During run 18 the Gemini API returned empty text for 8 calls; the collection notes put them inside a window of about three minutes (the empties were replaced, so this cannot be re-derived from the final file). Each affected call was re-collected once, after run 20 finished (18:42 to 18:43 Pacific, about 50 minutes after the rest of run 18), and the fresh response replaced the empty one in place, so every query-model cell holds 20 real responses and no empty response is scored this month. (2) Anthropic API credit exhaustion interrupted the Claude leg of run 18 after query 16; the 14 Claude responses for queries 17 to 30 were collected after credits were restored and appended to the file, and the count of Claude responses matched OpenAI's for every query. Run identity is the k-th occurrence of each query-model pair in file order, so those appended responses are run 18's, although they were collected later than the rest of that run. (3) The first launch of run 19 (17:57 Pacific) was stopped manually after the Claude API credit had run out (the credit ran out during run 18), before any of its data was added to the dataset; run 19 was relaunched at 18:15 Pacific after credits were restored and collected in full. (4) In 12 of the 20 runs, between one and four AI Overview queries hit a read timeout; each was re-collected once with the same call and appended. The final dataset comprises 2,400 scored core responses and 600 Google AI Overview collections, verified complete: every query-model pair appears exactly 20 times with zero duplicates. Every published total in this report was replicated exactly by an independent per-run recomputation from the raw data, with runs indexed by occurrence order per query-model pair. 179 of the 2,400 core responses (7.5%) name none of the tracked brands (181, or 7.5%, in August); these are genuine no-brand answers and are scored as zero mentions.

Flag definitions. The James Allen flag fires on a case-insensitive match of (?<![a-zA-Z])james\s+allen(?![a-zA-Z]) anywhere in a response, and The Clear Cut flag on the equivalent pattern for "the clear cut", however the brand is described. The Blue Nile lab-grown flag fires on a case-insensitive, any-character (including line breaks) match of blue\s+nile.{0,300}lab.?grown|lab.?grown.{0,300}blue\s+nile, so it also matches "lab grown" and "labgrown". The acknowledgment screen counts a James Allen response when this case-insensitive pattern matches within 300 characters before or after any James Allen mention: (?<![a-z])(absorb\w*|merg\w*|redirect\w*|shut\s*(?:ting)?\s*down|shuttered|sunset\w*|consolidat\w*|(?:now|is|are|been)\s+part\s+of\s+blue|closure|folded|wind(?:ing)?\s+down|no\s+longer\s+(?:operat\w*|exist\w*|independent\w*|sold|sell\w*)|(?:integrat\w*|incorporat\w*|acquir\w*)\s+(?:james\s+allen|into\s+blue)|(?:james\s+allen\S*\s+(?:\w+\s+){0,4})(?:integrat\w*|incorporat\w*)|integration\s+of\s+james). It is a keyword screen, not a reading of every response, and it misses some acknowledgments (6 further responses in September use "(formerly) James Allen", "sister brand" or "transition" wording).

Query clusters. Each query is assigned to the first matching cluster in this order: lab-grown (contains "lab-grown" or "lab grown"), ethical sourcing ("ethical", "conflict-free", "sustainable" or "ethically"), custom and gifting ("custom", "anniversary" or "wedding ring set"), price and value ("under $", "affordable", "cheapest", "value" or "money"), otherwise general purchase. These rules have been applied since Edition 2. The query list below was regrouped in this edition to match them: "best place to buy an engagement ring" had been listed under Custom and Gifting in earlier editions' lists but has always been scored as General Purchase; no other query moved.

Representativeness note: the OpenAI surface is measured through the gpt-4o API with web search, which keeps the longitudinal series internally consistent. Consumer ChatGPT may use newer models with different retrieval behavior, so this study's OpenAI figures describe the API surface, not necessarily the consumer app.

Coverage note (Canadian retailers): Four Canadian brands (Birks, Michael Hill, Spence Diamonds, and People's Jewellers) are tracked in every edition regardless of whether they appear. All four recorded zero mentions this edition, across the 2,400 core responses and 600 AI Overview collections. These are brand-tracking counts, not a geographic-location test.

All 30 Queries

Queries were grouped into five intent clusters. The query set is unchanged across all five editions. Clusters follow the keyword rules in the methodology.

Show all 30 queries ↓

General Purchase Decisions

• "best place to buy an engagement ring online"
• "where should I buy a diamond ring online"
• "best online jewelry stores for rings"
• "where to buy a wedding band online"
• "best websites to buy fine jewelry"
• "which online diamond retailer has the best reputation"
• "best alternative to Tiffany for engagement rings"
• "which online jeweler has the best diamond quality"
• "most reputable online diamond retailers"
• "best place to buy an engagement ring"
• "best jewelry brands for engagement rings"

Lab-Grown Diamonds

• "where should I buy a lab-grown diamond"
• "best lab-grown diamond retailers"
• "is it worth buying a lab-grown diamond engagement ring"
• "best place to buy lab-grown diamond rings online"
• "lab-grown diamond vs natural diamond where to buy"

Ethical Sourcing

• "most ethical jewelry brands"
• "best conflict-free diamond retailers"
• "sustainable jewelry brands for engagement rings"
• "ethical engagement ring brands"
• "where to buy ethically sourced diamonds"

Custom and Gifting

• "where to buy a wedding ring set"
• "best jewelry brands for anniversary gifts"
• "where to buy a custom engagement ring"
• "best place to customize an engagement ring"

Price and Value

• "best engagement rings under $3000"
• "affordable diamond engagement rings online"
• "where to get the best value diamond ring"
• "cheapest place to buy a real diamond ring"
• "best engagement ring for the money"

What This Means

Five editions in, the series has produced its largest single-month move, in points and in standard deviations, and, in the same month, a reminder of what its method is for. Rare Carat's weighted score doubled, beyond measured noise, and the study can say how large, in which models (about half of the points gain is in Perplexity), and across how many queries, and that it cannot say why. Brilliant Earth still leads Blue Nile by 3.7 SD of measured noise, marginally beyond the strict bar, but both leaders scored lower than in August. The mall-chain story rearranged for the second month running: two chains remain beyond the strict bar below their May levels, two are back within noise of them. Each edition's call was right about its own data, which is not a flaw in the method; it is the reason the method exists. Five months of data show large single-month moves in both directions, and a study that publishes its thresholds has to let new data overrule old headlines, and old watch items, in public.

For brands, the practical read is cautious: a single month's rank is a poor guide, since a brand can double its score in a month and the two market leaders can each fall beyond measured noise in the same month. For consumers, which app you ask still determines what you see: OpenAI named Rare Carat 10 times in 600 responses while Perplexity named it 332. James Allen's presence in answers held at 54.0% of responses, and most answers that include it do not reflect the consolidation (90.7% show no acknowledgment by the study's screen), and 66.5% of all answers triggered at least one market-structure flag, an upper bound on stale descriptions of a market that changed in the spring. This dataset documents where that stands as of September 2026.

Published September 30, 2026. Data collected in September 2026. Edition 5 of an ongoing monthly series; Edition 1 published May 24, 2026, Edition 2 published June 30, 2026, Edition 3 published July 27, 2026, Edition 4 published August 31, 2026.

Rings.com receives no compensation from any brand included in this study for inclusion or ranking. No brand had editorial input into the methodology or results. Rings.com also has an interest in how the broader story of AI visibility lands, a conflict no disclosure sentence removes; the response is that every claim here is published with the data and thresholds needed to check it independently. Rings.com is itself a participant in the market this study measures: it operates Ring Advisor, a free ring-shopping guidance tool built in part on this study's data. For that reason Rings.com is tracked in this study under the same methodology as every other brand. Through all five editions it has recorded zero mentions in the scored chat responses, zero appearances in Google AI Overviews in the four editions with AI Overview data (June to September), and zero of the model-cited source domains logged since tracking began (21,069 chat citations in September alone). If that changes in a future edition, it will be reported in that edition. Results may shift over time as model behavior, search integrations, and ranking systems evolve.

Full dataset available on request. Questions: rodney@rings.com