1,125 responses, 5 models, 487 entities: Penti leads every cut, and how Turkish the shelf behind it looks depends on the question

Vertical: Innerwear, hosiery & beachwear — Turkey
Method: A single point-in-time study of 1,125 responses — five large language models (Claude, Gemini, ChatGPT, Perplexity, Grok), each asked 45 Turkish-language questions five times
Collection date: 21 September 2026
Edition: 2 of a quarterly series. Edition 1 (June 2026) asked the same 45 questions; two of its five models were replaced for this edition.

Key terms: AI visibility, AI Visibility Score, innerwear brands, hosiery brands, beachwear brands, Turkish (yerli) underwear/swimwear brands, LLM brand recommendations, ChatGPT/Gemini/Claude/Perplexity/Grok brand recommendation, Generative Engine Optimization (GEO).

This report reproduces the real Turkish-language questions consumers ask AI assistants about underwear, socks, tights and swimwear, and measures how often, in what order, and across how many models five large language models name each brand, retailer and marketplace. Entities are ranked separately within six cuts of the same questions — all questions, origin-constrained, origin-neutral, and three question types — defined in Chapter 2. It is a single point-in-time study and should be read within the limitations in Chapter 11.

How to read this report — visibility is not quality

This report measures one thing: how often, how prominently, and across how many AI assistants a brand is named when people ask about innerwear, hosiery and beachwear in Turkey. A high score means these systems currently have a lot of information available about a brand and surface it readily — not that the brand is better, more comfortable, cheaper or higher quality. A low score, or a zero, means a brand is currently less visible to these systems — not that it is inferior. AI visibility reflects information availability and discovery, not product quality, fit, price or trustworthiness. Readers and shoppers should not treat an AI recommendation — or this report — as a verdict on whether a brand is “good” or “bad.” This study does not measure the accuracy, quality, or sentiment of any recommendation.

1. Executive Summary

One-sentence takeaway: Penti leads Turkey’s innerwear, hosiery and swimwear answers on every cut, named in 50.49% of 1,125 responses, and the shelf behind it is almost entirely domestic when the question asks for Turkish brands and much closer to even when it does not.

• Penti leads, and leads alone. Penti tops the general board at 97.27, named in 50.49% of all 1,125 responses (568) and by all five models. Dagi (70.16), Suwen (66.34), Ayyıldız (64.55) and Dündar Çorap (63.32) complete the top five.
• 26 of 487 tracked entities qualify on the general cut — 19 domestic and 7 foreign — with an average score of 55.05. The seven foreign qualifiers are Decathlon, Intimissimi, Calzedonia, Marks & Spencer, MOEVA, Nike and Oysho. Two marketplaces, Trendyol (#21) and Hepsiburada (#26), rank alongside the brands, because they are what the assistants name.
• The question decides how Turkish the shelf is. On the 23 origin-constrained questions (575 responses), domestic entities take 97.38% of qualifying mentions and foreign entities 2.62%, with origin on file for 2,103 of 2,103. On the 22 origin-neutral questions (550 responses), domestic entities take 55.31% and foreign entities 44.69%, with origin on file for 1,931 of 1,931 — and 16 of the 28 entities in that share are foreign.
• One foreign brand clears the bar on questions that ask for Turkish brands. MOEVA, a swimwear brand run from London, is named in 55 of the 575 origin-constrained responses (9.57%) and ranks 14th of the 32 qualifiers there. Every other qualifier on that cut is domestic.
• Origin shares cannot be set against June’s. This edition assigns origin by brand home; June required a brand to be founded and headquartered in Turkey. Under the new rule OYE Swimwear, Sand&Blue and Alvina are domestic and Lovable is Italian. That is the rule and the research changing, not the market, and no origin share in this report is compared with June’s.
• The assistants agree on the leader, not on how long an answer is. Every model names Penti more often than any other entity, but Grok 4.3 names 11.48 entities per answer and Gemini 3.6 Flash 9.51, against 4.19 for Claude Haiku 4.5 and 4.21 for ChatGPT 5.6 Luna Pro. The two marketplace ranks are largely Grok’s: it accounts for 108 of Hepsiburada’s 124 general-cut mentions and 134 of Trendyol’s 172.
• Grounding was complete for four models and partial for one. 87.11% of all 1,125 responses were web-grounded. Four models grounded 100.00% of their 225 answers; Gemini 3.6 Flash grounded 35.56%, and the other 145 of its answers returned no grounding metadata at all — which is not evidence that it answered from memory.
• Against June, Kom rose most and Ayyıldız fell most — on a panel that changed. On the same 45 questions, with June’s answers counted the way this edition counts and with Gemini 2.5 Flash-Lite and GPT-4o-mini replaced by Gemini 3.6 Flash and ChatGPT 5.6 Luna Pro, Kom moved from 5.46% of 1,117 responses to 19.02% of 1,125 (+13.56 points) and Ayyıldız from 30.44% to 19.56% (-10.88); Kom’s gain sits mostly in the two models that joined (Chapter 8).
• Brand and marketplace sites lead the sources. penti.com is cited in 80 of 225 question-model cells (35.56%), ahead of ayyildiz.com.tr (49) and trendyol.com (46). Claude Haiku 4.5 returned exactly 5 citations on every one of its answers, so its domain counts are capped and do not compare with the other models’.

Why it matters. Shoppers increasingly ask AI assistants where to buy underwear, socks and swimwear, and whether a brand appears in those answers is becoming a discovery channel in its own right. June’s edition set the baseline; this one repeats its 45 questions word for word, which makes a first reading of movement possible — with the confounds stated beside it, because two of the five assistants answering are not the ones that answered in June. The clearest finding is not a movement but a structure: the same assistants, on the same day, produce a near-exclusively Turkish shelf or a contested one depending only on whether the question asks for Turkish brands. A reminder: the numbers below describe visibility and information availability — not which brands are best.

2. Methodology

One-sentence takeaway: Five models answered 45 Turkish questions five times each, giving 1,125 responses across 225 question-model cells; 487 entities were scored within each cut, and since June two models, the counting rule and the origin rule have changed.

2.1 Scope and models

The study queried five large language models through their APIs with identical Turkish-language questions. Each of the 45 questions was put to each model five times, giving 225 question-model cells and 1,125 responses analyzed, all collected on 21 September 2026. The questions are in Turkish and models infer the market from them. The deployment’s system prompt and locale parameter were not recorded with this run, so this report makes no statement about either.

2.2 Web-search configuration

Every model was asked to search the web. Whether it did is a separate question, and the table below keeps the two apart: the first column is what the deployment asked of each model; the rest is what each model returned.

Overall, 87.11% of the 1,125 responses were web-grounded. Four models grounded 100.00% of their answers. Gemini 3.6 Flash grounded 35.56%: 80 answers report a search, and the other 145 carry no grounding metadata of any kind. An “unknown” answer is not a memory answer. The model was asked to search, and it is not known whether it did; the answer may rest on retrieval that was not reported, on the model’s own knowledge, or on both. This report never treats those 145 answers as answers from memory.

That is also why this edition, like June’s, has no search-versus-memory chapter, though for a different reason. Every model was asked to search, so there is no memory segment by design; and the only ungrounded answers are one model’s, of unknown provenance. Splitting the board by grounding would compare Gemini 3.6 Flash with the other four models, not search with memory.

A citation cap, not a measurement. Claude Haiku 4.5 returned exactly 5 citations on every one of its 225 answers. Its domain counts are truncated at that ceiling and do not compare with an uncapped model’s; Chapter 9 reads them that way.

2.3 Questions and cuts

All 45 questions are the ones asked in June, carried verbatim: none was added, none retired and none reworded. Each is typed by what it asks — discovery (a general “which brand?”), attribute (a specific property such as seamless or affordable) or use_case (a need, such as a postpartum swimsuit) — and scoped by whether it constrains origin. A constrained question asks for Turkish brands (“iyi bir yerli iç giyim markası var mı” — is there a good Turkish innerwear brand?); a neutral one does not (“dikişsiz belli olmayan külot önerisi” — a seamless brief that doesn’t show).

Every metric is computed separately within each of six cuts:

June called the first three cuts main, local and open. The all-questions cut is referred to in this report as the general cut.

Every question also records the subcategory it was written for — innerwear (İç giyim), hosiery and socks (Çorap), or swimwear and beachwear (Mayo/Bikini) — and the appendix shows it against each question. Under this series’ rules, a subcategory is published as a cut only when the study declares it before collection. This study did not, so June’s subcategory leaderboards are not repeated and no subcategory figure appears in this report. Chapter 7 returns to this.

2.4 Entity resolution

Every name in every answer is matched against Herm.io’s brand registry. A name that matches a registered brand or one of its aliases is counted directly and never becomes a candidate. Names that do not match are reviewed as candidates: 428 this period. Of those, 407 were resolved to a brand, 15 were given a verdict by a language model, and 6 were answered by a stored verdict from earlier work. These figures count the work, not the field, and do not reconcile to the 487 entities tracked.

Six candidates were rejected as not a brand, and they show what the rules do.

A rejected name is not always a lost mention. Three of the six name tracked brands. “Tutku / Tutku Elit” (6 responses) was credited to Tutku. “Kom & Ayyıldız” and “Kom & Dagi” (3 responses each) were split, and each brand they name was credited.

Shops are not brands. plajperisi.com.tr (5 responses) and glingerie.com.tr (4) were judged not to be brands and are not counted.

A web address is not a brand name. ayyildiz.com.tr, written out in 3 responses, is Ayyıldız’s own domain. It was judged not to be a brand name, so the 3 responses that wrote out the address were not credited to Ayyıldız for it.

One name is overruled before any verdict is asked. “sunset” means a different brand in this study than in the global registry: it is redirected to Sunset Swimwear, a Turkish swimwear brand, not Sunset Magazine (US), because the corpus mentions describe a swimwear maker. A redirect moves mentions without changing the entity count; no name was suppressed this edition.

Resemblance is not identity. Among entities named in at least three answers, nine near-name pairs were flagged — Trendyolmilla and Trendyol, Pazarama and Zara, and Vakkorama and Vakko among them. None of the nine shares a website on file, and each pair is tracked as two entities.

The result is 487 entities: 394 brands, 39 brands that are also retailers, 26 retailers, 14 sub-brands, 9 marketplaces and 5 publishers. Origin is on file as domestic for 266 and foreign for 174; 47 have no origin on file. No entity is provisional. No entity has a product category on file, so scope could not be resolved for any of them, and all 487 are eligible to rank.

2.5 AI Visibility Score (0–100)

Score = 0.45 × Mention + 0.3 × Position + 0.25 × Breadth

Mention is the share of the cut’s responses that name the entity. Position is the mean reciprocal rank (MRR) of the entity’s first appearance in each answer that names it — first is 1, second 0.5, and so on. Breadth is the number of models, of five, that name it. Each component is scaled to the highest value among qualified entities in that cut and is never clipped. An entity qualifies for ranking on a cut if it is named in at least 5% of that cut’s responses; there is no absolute mention floor.

Worked example — Penti, general cut. Penti is named in 568 of 1,125 responses, a mention rate of 50.49%, the highest among qualified entities in the cut, so its mention component is 100.00. Its MRR of 0.5303 against the cut maximum of 0.5835, held by Dündar Çorap, gives a position component of 90.89. It is named by 5 of 5 models, so breadth is 100.00. The composite is 0.45 × 100.00 + 0.3 × 90.89 + 0.25 × 100.00 = 97.27. The position component is computed from the unrounded inputs (MRR 0.530297646 against 0.583473167); dividing the four-decimal figures printed on this page gives a slightly different number.

Penti scores 97.27 rather than 100 because the highest MRR in the general cut belongs to Dündar Çorap, not to the leader. Because the maxima come from qualified entities only, an unqualified entity can exceed 100 on a component: 48 of 1,978 entity-cut rows do (2.40%), all of them unqualified — 2.70% of the 1,791 unqualified rows. A value above 100 says the entity beats the cut leader on that one component while being named too rarely to rank.

A score belongs to one cut’s scale. The scale moves whenever the leader moves, so an AI Visibility Score is never compared across cuts, across editions or across verticals. This report shows no score change since June. Where it compares editions (Chapter 8), it uses raw mention rate on the general cut only. The maxima behind every table are published in the report’s data file.

2.6 What changed since June, and what is comparable

Three changes matter for any comparison with June. The panel: three models are carried and two were replaced, and nothing in this report corrects for that. The counting rule: June’s page counted any answer whose text contained a name; this edition counts only names identified as a recommended brand. For ordinary brands the two rules give nearly the same counts; for the two marketplaces, whose names appear in where-to-buy phrases and links, they do not (Chapter 8). So that both sides of every comparison count a brand the same way, June’s levels in this report are June’s answers, counted the way this edition counts — not the figures June’s page printed. The origin rule: see 2.7.

2.7 Origin, neutrality and self-exclusion

Origin is the brand home: the country the brand itself is operated from, which is not always its owner’s. Fendi Casa is Italian though LVMH is French; Cetaphil is American though Galderma is Swiss. Where no separate brand home can be established, the parent’s headquarters stands in. Local-market naming variants are folded into their parent before origin is read, so a brand named after its market subsidiary is not recorded as domestic; where the parent is not tracked in this edition, the variant carries its own researched origin. Tables show ISO country codes; shares use a domestic/foreign rollup. Entities with no origin on file stay in every table, are excluded from every origin share, and are counted — never defaulted.

June used a different rule. June’s edition counted a brand as Turkish only if it was founded and headquartered in Turkey. Under this edition’s rule OYE Swimwear, Sand&Blue and Alvina are domestic and Lovable is Italian. These are changes in the rule and in the research, not in the market. Within this edition the rule is published here beside every origin figure, which is what makes a domestic-versus-foreign comparison safe; across the change it is not, and no origin share in this report is compared with June’s.

Marketplaces and retailers rank in every leaderboard, because they are what the assistants actually name. The brand-only view in section 3.3 removes the two qualifying marketplaces, Hepsiburada and Trendyol, by name and changes nothing else. Retailers with their own labels — Decathlon, Cottonhill and LC Waikiki among the qualifiers — are retained everywhere.

Origin shares are read on a brand-only, qualified basis. The headline share is computed over qualified entities with origin on file, with marketplaces and publishers left out — a marketplace is where a brand is bought and a publisher is where it is written about — and own-label retailers kept in. It is always printed with its coverage: the qualifying mentions with origin on file. The same arithmetic over every tracked entity, qualified or not, is published as a sanity check; it describes the long tail, not the shelf a reader sees. This basis applies to shares only; it does not change who ranks.

Neutrality and self-exclusion. This is a market-wide study with no focus brand. Herm.io’s own domain is excluded from the source data before analysis, so the company’s own content neither appears among the cited domains nor influences any figure.

3. Overall Visibility Leaderboard

One-sentence takeaway: Twenty-six of 487 tracked entities qualify on the general cut; Penti stands alone at 97.27, and below it a brand’s rank depends as much on how early it is named as on how often.

The list below ranks the 26 entities named in at least 5% of all 1,125 responses, by AI Visibility Score. As a reminder, it ranks visibility, not quality.

Average score of the 26 qualified entities: 55.05. Nineteen are domestic and seven foreign.

3.1 Tier narrative

The leader (97.27). Penti holds the highest mention rate in the cut, 50.49% (568 of 1,125), and is named by all five models. Its only shortfall is position: an MRR of 0.5303 against Dündar Çorap’s 0.5835.

The second tier (70.16–64.55). Dagi (70.16), Suwen (66.34) and Ayyıldız (64.55) are each named in roughly a fifth to a third of all responses — 30.49%, 22.58% and 19.56%. Ayyıldız is named least often of the three but earliest, with an MRR of 0.4303 against Dagi’s 0.3499.

Position specialists, and Kom (63.32–59.94). Dündar Çorap ranks fifth on a 9.33% mention rate (105 of 1,125) because, when it is named, it is named earlier than any other qualifier — the cut’s highest MRR, 0.5835. Bolero (60.97) follows the same route: 9.60% (108 of 1,125) with an MRR of 0.5332. Kom is named in 214 responses (19.02%), against 105 and 108 for those two, yet ranks below both at 59.94 because it tends to appear later (MRR 0.3497).

The packed middle (57.69–49.35). From Decathlon (#8) to DeFacto (#19), with no clean origin boundary. Decathlon, a French retailer with its own labels, is the highest foreign entry, at 11.82% (133 of 1,125). Intimissimi reaches 10th on a 5.16% mention rate (58 of 1,125), the joint lowest of any qualifier with Nike’s, on an MRR of 0.5077. Blackspade is the reverse case: 14.13% (159 of 1,125), but an MRR of 0.3062 holds it at 14th. Marks & Spencer (53.97) and Tutku (53.94) sit a few hundredths apart; gaps of that size here carry no meaning.

The tail of the board (48.89–35.82). Cottonhill, Trendyol, Nike, NBB, LC Waikiki, Oysho and Hepsiburada. The two marketplaces show the board’s sharpest split between frequency and position: Trendyol is named in 15.29% of responses (172 of 1,125) and Hepsiburada in 11.02% (124 of 1,125), but with MRRs of 0.1739 and 0.1166 — when they appear, they appear late, typically as where to buy rather than what to buy. Every qualifier except Oysho and Hepsiburada is named by all five models; those two are named by four. The remaining tracked entities sit below the threshold, and the appendix lists them all.

3.2 The same board by mention rate

Scores reward prominence as well as frequency, so the ranking by raw mention rate is a different list, and the difference is informative.

The first four are the same either way. The marketplaces move furthest: Trendyol is 6th by volume and 21st by score, Hepsiburada 10th and 26th. Blackspade (7th by volume, 14th by score), DeFacto (11th and 19th), Tutku (9th and 13th) and Kom (5th and 7th) all lose ground to position. Decathlon and Marks & Spencer hold the same place on both lists.

3.3 Brand-only view

Removing the two qualifying marketplaces — Hepsiburada and Trendyol, named explicitly, with nothing else changed — leaves 24 entities. Scores are unchanged, because the scale is still the general cut’s; the entities ranked below Trendyol each move up a place. Decathlon, Cottonhill and LC Waikiki stay: they are retailers with their own labels, not marketplaces.

4. Two Shelves: Origin-Constrained and Origin-Neutral Questions

One-sentence takeaway: On the 23 origin-constrained questions domestic entities take 97.38% of qualifying mentions and foreign entities 2.62%, and on the 22 origin-neutral questions they take 55.31% and 44.69% — with origin on file for all 2,103 and all 1,931 qualifying mentions respectively.

The 45 questions divide into 23 that ask for Turkish brands and 22 that do not. Each set has its own leaderboard, its own denominator and its own scale. Neither is the market on its own, so the two figures are always quoted together, each with the number of questions behind it.

4.1 Origin share on both cuts

Origin-constrained (23 questions, 575 responses): domestic entities take 97.38% of qualifying mentions and foreign entities 2.62%, with origin on file for 2,103 of 2,103 qualifying mentions. Of the 30 qualified entities in the share, 29 are domestic and one, MOEVA, is foreign.

Origin-neutral (22 questions, 550 responses): domestic entities take 55.31% and foreign entities 44.69%, with origin on file for 1,931 of 1,931. Of the 28 qualified entities in the share, 12 are domestic and 16 foreign.

The assistants clearly read “yerli” as a constraint and answer it with domestic names. Without it, foreign entities outnumber domestic ones among the qualifiers, but domestic entities still take the larger share of mentions, because the most-named entities on the neutral shelf — Penti, Suwen and Dagi — are domestic.

The sanity check points the same way. Across every tracked entity, qualified or not, domestic entities take 90.34% and foreign entities 9.66% of mentions on the constrained questions (origin on file for 3,706 of 3,794), and 52.22% and 47.78% on the neutral questions (3,598 of 3,720). On the constrained questions, the long tail is more foreign than the shelf — a reminder that a constrained question narrows what qualifies, not everything that is named.

Left out of every share, and still ranked. Trendyol and Hepsiburada are left out of both shares (91 and 69 mentions on the constrained cut, 81 and 55 on the neutral cut). Across all tracked entities, 11 marketplaces and publishers with 207 mentions are left out on the constrained cut, and 11 with 179 on the neutral cut. Both marketplaces are domestic, so including them would raise the domestic share for a reason that has nothing to do with brands. They still rank in every leaderboard.

4.2 The origin-constrained shelf

32 entities qualify on the 23 origin-constrained questions (575 responses). Thirty enter the origin share — 29 domestic, 1 foreign — and the two marketplaces are left out of it.

Penti leads at 88.21, named in 48.87% of these responses (281 of 575), with Dagi second at 68.53 (33.39%, 192 of 575). Below them, this shelf rewards specialists. Woolona ranks third on a 6.09% mention rate (35 of 575) with an MRR of 0.8595; Haşema ranks eighth on 6.43% (37 of 575, MRR 0.7030); Norfolk twelfth on 5.91% (34 of 575, MRR 0.7005, named by 4 of 5 models). Dündar Çorap, fifth overall, is fourth here with a higher mention rate (14.26%, 82 of 575) and an MRR of 0.6426. Ranks built mainly on position and thin mentions should be read cautiously.

MOEVA is the one foreign qualifier. It is named in 55 of the 575 responses (9.57%), by all five models, and ranks 14th at 47.44. Its brand home is London, so it is recorded as GB. That makes it a genuine case of the assistants answering a question that asks for Turkish brands with a foreign one. OYE Swimwear, 17th here at 8.87% (51 of 575), is recorded as domestic under this edition’s rule; under June’s rule it was foreign, which is why no comparison with June’s constrained shelf is drawn.

Hepsiburada qualifies on this cut on mentions from two models only (69 mentions), the lowest breadth of any qualifier here; Chapter 5 shows why.

4.3 The origin-neutral shelf

Penti leads this shelf too, at 95.22 and 52.18% (287 of 550). The difference is what follows it. Decathlon is second at 64.91, named in 21.27% of these responses (117 of 550) — a French retailer with its own labels. Suwen (24.18%) and Dagi (27.45%) come next; Dagi is named more often but later. Bolero reaches fifth on 8.36% (46 of 550) with an MRR of 0.6408.

Foreign entities run through the rest of the board. Marks & Spencer is seventh at 18.73% (103 of 550); Intimissimi (10.55%) and Calzedonia (14.91%) are eighth and tenth. Speedo qualifies at eleventh on 5.09% (28 of 550) with an MRR of 0.5604, named by 4 of 5 models. Nike, Victoria’s Secret, Adidas, Oysho and Pierre Cardin all place in the top twenty. Two Turkish retailers with their own labels, Cottonhill and LC Waikiki, are 12th and 15th, both at 7.09% (39 of 550).

4.4 What this chapter does not compare

No origin share here is compared with June’s, because June assigned origin by a different rule (section 2.7). The constrained figure is never read on its own: 97.38% domestic and 2.62% foreign on 23 questions that ask for Turkish brands describes the question as much as the market. And the question-type cuts in Chapter 6 each blend constrained and neutral questions, so their counts of domestic qualifiers are not origin measurements.

5. Differences Between Models

One-sentence takeaway: Every model names Penti more often than any other entity, but Grok 4.3 names 11.48 entities per answer to Claude Haiku 4.5’s 4.19, and the two marketplace ranks rest mainly on one model.

5.1 Per-model behavior

All five lead with Penti, at very different rates. Penti is named in 76.89% of Gemini 3.6 Flash’s answers, 60.44% of Grok 4.3’s and 53.78% of Perplexity Sonar’s, but in 32.89% of ChatGPT 5.6 Luna Pro’s and 28.44% of Claude Haiku 4.5’s. The two models with the shortest answers name even the leader least often.

Answer length is the largest difference. Grok 4.3 names 11.48 entities per answer and reaches the widest repertoire, 291 distinct entities. Gemini 3.6 Flash names 9.51 per answer (220 distinct), Perplexity Sonar 5.73 (196). ChatGPT 5.6 Luna Pro (4.21, 198 distinct) and Claude Haiku 4.5 (4.19, 174 distinct) give short, selective answers. For an entity at the edge of visibility, being named at all depends heavily on which assistant answers.

5.2 Each model’s most-named entities

The models agree on the core: Dagi or Ayyıldız is second for every model except Grok 4.3. They differ beneath it. Grok 4.3 puts the two marketplaces second and fourth — Trendyol at 59.56% and Hepsiburada at 48.00% — and is the only model with either in its top ten. Claude Haiku 4.5’s top ten includes Marks & Spencer (10.22%), Calzedonia (9.33%) and MOEVA (8.44%). ChatGPT 5.6 Luna Pro has Trendyolmilla fourth (14.22%) and Norfolk sixth (12.44%), names no other model’s top ten carries. Gemini 3.6 Flash has Oysho ninth (18.67%).

5.3 One model’s habit

Three qualifying entities on the general cut draw at least three in five of their mentions from a single model:

The marketplace ranks are Grok’s. Grok 4.3 accounts for 108 of Hepsiburada’s 124 general-cut mentions (87.10%) and 134 of Trendyol’s 172 (77.91%). Grok names Trendyol in 59.56% of its 225 answers, almost as often as it names Penti. The two ranks — Trendyol 21st and Hepsiburada 26th — describe one assistant’s habit of pointing to where to buy, not a consensus across the panel. On the origin-constrained cut, Hepsiburada qualifies on two models only.

Oysho’s rank is largely Gemini’s. Gemini 3.6 Flash supplies 42 of Oysho’s 63 mentions (66.67%). Gemini 3.6 Flash joined the panel this edition, so this is also a reminder of how a panel change can move a single entity.

5.4 What only one model names

Every model names entities no other model names at all: 82 for Grok 4.3, 52 for Gemini 3.6 Flash, 42 for ChatGPT 5.6 Luna Pro, 25 for Perplexity Sonar and 23 for Claude Haiku 4.5.

6. Question-Type Ownership

One-sentence takeaway: Penti leads all three question types; discovery questions qualify 49 entities against 25 each for attribute and use-case questions, and the position specialists change with the type.

Discovery — “which brand?” — spreads visibility widest: 288 distinct entities, 49 qualifiers and the lowest average score of the three, 40.41. Penti is named in 61.07% of these responses (229 of 375), its highest rate on any cut, and the top five are all domestic: Penti, Dagi, Suwen, Ayyıldız and Kom. Falke holds the cut’s highest MRR, 0.7692.

Attribute — a specific property — narrows the field to 25 qualifiers from 257 distinct entities, with the highest average score, 53.19. Penti leads at 48.80% (183 of 375). Dündar Çorap is second at 66.97 on a 14.13% mention rate (53 of 375), carried by an MRR of 0.7120. Speedo holds the cut’s highest MRR, 0.7381.

Use case — a need — also qualifies 25, from 284 distinct entities, with an average score of 48.79. Penti leads at 41.60% (156 of 375), its lowest rate of the three types. Woolona is second at 67.98 on 12.00% (45 of 375) with an MRR of 0.8519, the highest of any question type. Decathlon is third at 67.77 and the only foreign entity in any of the three top fives, named in 20.00% of use-case responses (75 of 375).

Each rests on the same number of responses, 375. The differences between them are not origin findings: the cuts differ in their mix of constrained and neutral questions, and Chapter 4 is where origin is measured.

7. Attribute Axes

One-sentence takeaway: No attribute axis can be published this edition: no entity has a product category or audience on file, and no subcategory was declared before collection, which this series requires before publishing one as a cut.

An attribute chapter ranks entities by something they are — the product categories they make, or the audience they serve — rather than by the questions asked. This edition’s data carries no such attribute: none of the 487 tracked entities has a product category on file, and no audience attribute is recorded. A product-shelf chapter cannot be written from this export, and this report does not substitute an estimate for one.

The questions themselves record a subcategory — innerwear, hosiery and socks, or swimwear and beachwear — and the appendix shows it. That is a property of the question, not of the entity, and publishing it would make it a cut. Under this series’ rules a subcategory is published as a cut only when the study declares it before collection. This study did not, so June’s subcategory leaderboards have no counterpart here.

8. What Changed Since June

One-sentence takeaway: On the same 45 questions, against June’s answers counted the way this edition counts, Kom gained 13.56 points and Ayyıldız lost 10.88 on the general cut — with two of the five models replaced between editions, which no figure here corrects for.

8.1 How this comparison is built

This chapter compares raw mention rate on the general cut and nothing else; AI Visibility Scores do not travel between editions. All 45 questions were asked in both editions, so the change on the shared questions is the plain difference between the two levels: June’s 1,117 responses against September’s 1,125.

June’s levels are recounts. Every June figure in this chapter is June’s answers, counted the way this edition counts — re-read with this edition’s brand registry so both sides count a brand the same way. They are not the figures June’s page printed. June’s published page counted every answer whose text contained a name — including “buy it on Trendyol” and trendyol.com links — while this edition counts only names identified as a recommended brand; for ordinary brands the two counts land close together, and only the two marketplaces diverge materially: June’s published page showed Trendyol in 369 answers and Hepsiburada in 209, against 240 and 135 when June’s answers are counted the way this edition counts.

The panel is the confound, and it is declared, not corrected. Claude Haiku 4.5, Perplexity Sonar and Grok 4.3 answered in both editions. Gemini 2.5 Flash-Lite and GPT-4o-mini answered in June only; Gemini 3.6 Flash and ChatGPT 5.6 Luna Pro answered in September only. Rebasing adjusts for the question set, which did not change; nothing adjusts for the panel. Every number below carries that confound: the direction of a large move is better supported than its exact size.

Origin in this chapter is this edition’s coding, applied to both editions. June’s own origin rule is not re-read.

Of the 326 entities named in both editions, 174 rose, 130 fell and 22 were unchanged. Domestic entities account for 181 of the 326, of which 92 rose; foreign entities for 121, of which 71 rose; 24 have no origin on file and are neither. A further 161 entities were named only in September’s answers and 92 only in June’s.

8.2 The movers

Gains. Kom moved from 5.46% of June’s recounted answers (61 of 1,117) to 19.02% (214 of 1,125), +13.56 points. Penti gained 9.49 points, from 41.00% to 50.49%, and Blackspade 8.85, from 5.28% to 14.13%. Decathlon (+6.27, from 5.55% to 11.82%) is the only foreign entity among the ten largest gains.

Falls. Ayyıldız fell 10.88 points, from 30.44% (340 of 1,117) to 19.56% (220 of 1,125), and is still fourth on the general board. Trendyol fell 6.20 points, from 21.49% to 15.29% — measured against June’s recount of 240, not the 369 June’s page printed. Marks & Spencer fell 6.07, from 16.11% to 10.04%, and Sand&Blue 4.49, from 6.80% to 2.31%, coded domestic on both sides. Skechers and Minca, each down to 8 mentions (0.71%), rest on too few mentions to read as more than direction.

8.3 Where the gains sit

Kom’s gain sits mostly in the two models that joined. Kom is third in Gemini 3.6 Flash’s answers (48.89%) and third in ChatGPT 5.6 Luna Pro’s (14.67%); of the three models that answered in both editions, only Grok 4.3 has it in its top ten, in tenth place (18.22%). Blackspade’s profile is similar: fifth in Gemini 3.6 Flash’s answers (39.11%), tenth in ChatGPT 5.6 Luna Pro’s (6.67%), and in none of the three carried models’ top tens. This report publishes no shared-model comparison, so it does not say how much of either gain the panel change accounts for — only that the panel change cannot be ruled out as a large part of it.

9. The Discovery Ecosystem: Where AI Learns About Brands

One-sentence takeaway: penti.com is cited in 80 of the 225 question-model cells (35.56%), and the top of the list is brand and marketplace sites; one model’s citations are capped and another’s grounding is partial, so the map is uneven by model.

When an assistant answers with web search, it draws on whatever is available to it at that moment. The cited addresses were reduced to domains and counted by question-model cell — one cell is one question put to one model, holding that question’s repeats — never by response. Three units appear in this chapter, and they are not interchangeable:

• Citations recorded — 10,216. The citation entries the models returned.
• Link rows — 3,540. Attribution rows, one per model, entity, domain and attribution rule; one mention can be attributed to several domains.
• Mentions attributed — 10,155. Those rows’ mention counts, summed. Because one mention can sit on several rows, this is not a count of distinct mentions and does not compare with any mention rate’s numerator.

Across all answers, 296 domains were cited, 136 of them in a single cell only.

9.1 The most-cited domains

penti.com is cited in 80 of the 225 cells (35.56%), followed by ayyildiz.com.tr (49 cells, 21.78%), the marketplace trendyol.com (46, 20.44%) and dagi.com.tr (45, 20.00%). Every one of the fifteen has an owner among the tracked entities — mostly brands on the leaderboards, plus both marketplaces. oggusto.com is the outlier in the attribution column: cited in only 17 cells (7.56%) but carrying 458 attributed mentions, a reminder that attribution rows count every entity a cited page is credited with.

9.2 Entities’ own sites

201 of the 296 domains are the own site of some tracked entity. Tracked is not ranked: marketplaces, publishers and entities that qualify in no cut have sites too, so this count does not reconcile with the leaderboard, and it should not be read as brands citing themselves. Rows on those domains account for 4,138 of the 10,155 attributed mentions — attribution-row units, not distinct mentions.

That figure is understated. Eleven cited domains look like a tracked entity’s own site but are not attributed to it, because the registry holds a different website or none: adidas.com.tr, amazon.com.tr, ankamedikal.com.tr, corapfabrikam.com, gowith.com.tr, jojomamanbebe.com, lavcorap.com.tr, marksandspencer.com.tr, n11.com, steppeberry.com and tesmay.com. They are therefore not counted among own sites.

In 62 rows, the rule that assigns a mention to the entity’s own domain and the rule that assigns it by position in the answer disagreed about which citation the mention belongs to. Those conflicts are surfaced in the data, not resolved.

9.3 Why the map is uneven

The models do not contribute equally. Claude Haiku 4.5 returned exactly 5 citations on every one of its 225 answers — 1,125 in all — so its domain counts stop at a ceiling and do not compare with an uncapped model’s. Perplexity Sonar alone returned 6,103 of the 10,216 citations. Gemini 3.6 Flash returned 712 and grounded 35.56% of its answers. A domain’s place in this table therefore reflects which models cite at length as much as what any one model reads.

This edition carries no source-type classification of domains and no per-response citation counts, so June’s source-type mix has no counterpart here. Nothing in this chapter tests whether being cited causes visibility or merely accompanies it.

10. What the Patterns Suggest

One-sentence takeaway: Five neutral readings, not recommendations: the leader holds on every cut, origin follows the question, frequency and position are separate routes onto the board, the marketplace ranks are one model’s habit, and answer format decides who is seen in the long tail.

This chapter describes patterns in the data. It is not advice, and visibility remains separate from quality.

The leader holds on every cut. Penti leads the general cut (97.27), the origin-constrained cut (88.21), the origin-neutral cut (95.22) and all three question types (93.50, 88.87, 87.68), and it is the most-named entity for all five models. No other finding in this report is as stable across the ways the data can be cut.

Origin follows the question. On the 23 origin-constrained questions domestic entities take 97.38% of qualifying mentions and foreign entities 2.62%; on the 22 origin-neutral questions, 55.31% and 44.69% — with origin on file for every qualifying mention on both. The assistants read “yerli” as a constraint. Without it, the shelf is still led by domestic names but is shared.

Frequency and position are separate routes onto the board. Dündar Çorap is named in 9.33% of all responses and ranks fifth because it is named earliest (MRR 0.5835). Trendyol is named in 15.29% and ranks 21st because it is named late (MRR 0.1739). On the constrained cut, Woolona ranks third on a 6.09% mention rate with an MRR of 0.8595. “Named often” and “named early” are different signals, and the score weighs both.

The marketplace ranks are one model’s habit. Grok 4.3 supplies 108 of Hepsiburada’s 124 general-cut mentions and 134 of Trendyol’s 172. A marketplace in an answer usually means where to buy, and in this panel it mostly means one assistant’s answering style.

Answer format decides who is seen in the long tail. Grok 4.3 names 11.48 entities per answer, reaches 291 distinct entities and names 82 that no other model does; Claude Haiku 4.5 names 4.19 per answer, 174 distinct and 23 alone. For an entity at the edge of the board, whether it is seen depends on the assistant’s format as much as on the entity.

11. Limitations and Notes

One-sentence takeaway: This is one market on one collection day, measured on a panel and under an origin rule that both changed since June; read every figure with the notes below.

• The panel changed. Gemini 2.5 Flash-Lite and GPT-4o-mini were replaced by Gemini 3.6 Flash and ChatGPT 5.6 Luna Pro. Chapter 8 declares this confound beside every figure; it does not correct for it, and at least part of the largest gain sits in the models that joined.
• The counting rule changed. June’s page counted a name anywhere in the answer text; this edition counts names identified as a recommended brand. June’s levels in this report are therefore recounts of June’s answers, not the figures June printed. The difference is material only for the two marketplaces.
• The origin rule changed. June required a brand to be founded and headquartered in Turkey; this edition uses brand home. OYE Swimwear, Sand&Blue and Alvina are domestic under the new rule and Lovable is Italian. No origin share is compared across the change.
• Origin classification involves judgment. Brand home is a researched decision, and brands run from one country under an owner in another are genuinely arguable cases. 47 entities have no origin on file; they appear in every table and in no share.
• Some local-market names could not be folded. A few names carry a brand’s Turkish market subsidiary in them while the brand itself is not tracked in this edition, so there was nothing to fold them into. Each carries its own researched origin, and each was named in a single response.
• Scope was not resolved. No entity has a product category on file, so the study could not separate core innerwear, hosiery and swimwear entities from adjacent ones, and every entity is eligible to rank. An entity whose main business lies outside the vertical can rank if the assistants name it often enough.
• One model’s answers have unknown grounding. 145 of Gemini 3.6 Flash’s 225 answers returned no grounding metadata. They are not memory answers, and nothing in this report treats them as such.
• One model’s citations are capped. Claude Haiku 4.5 returned exactly 5 citations per answer, so its source counts are truncated (Chapter 9).
• Own-site figures are understated. Eleven cited domains look like a tracked entity’s own site but are not attributed to it, because the registry holds a different website or none.
• Three qualifying positions rest on two models. Hepsiburada on the origin-constrained cut (69 mentions), and Oysho (27) and Hepsiburada (28) on the discovery cut. Their visibility is model-specific, not panel-wide.
• Rank built on position. Woolona, Haşema, Norfolk and Dündar Çorap rank high on some cuts mainly because they are named early when named at all. That is the score working as designed, but it is a different kind of visibility from being named in half of all answers.
• Small score gaps are not meaningful. In the packed middle of each board, differences of a point or two should not be over-interpreted.
• Components can exceed 100. Because each component is scaled to the qualified leader and never clipped, 48 of 1,978 entity-cut rows (2.40%) exceed 100 on a component — all of them unqualified.
• Near-name pairs. Nine pairs of similar names among entities named in at least three answers were flagged: Direct Message / Essa, DoReMi / Dore, Dore / Dorea, Doreanse / Dore, Doreanse / Dorea, Emay / Emayo, Pazarama / Zara, Trendyolmilla / Trendyol and Vakkorama / Vakko. None shares a website on file, and each pair is tracked as two entities. The outcome is in the data; the reasoning behind each merge or non-merge is not.
• System prompt and locale. Neither was recorded with this run.
• Single, point-in-time measurement. All responses were collected on 21 September 2026. Five repeats per question-model cell capture variation within a model, not drift from day to day.
• Models are probabilistic. The same question can produce different answers; repeats reduce this variability but do not remove it.
• Web search depends on momentary conditions. Grounded answers depend on what was indexed and served at the time of collection.
• Precision over recall. Only names identified as a recommended brand are counted. Brands referred to indirectly, or only by a web address, are not.
• Visibility is not quality. This report measures mention, position and breadth only. It does not assess whether a brand was described positively, whether a recommendation was accurate, or whether a product is good. A score is not an endorsement.
• Model versions date quickly. These findings are specific to the model versions in section 2.1, as of September 2026.

12. Appendix & Data

One-sentence takeaway: All 45 questions and all 487 tracked entities behind this report are listed below, and the underlying data is available on request.

12.1 All 45 questions

12.2 All tracked entities (487)

The table lists all 487 entities tracked in this study, ordered by general-cut mentions, and marks the 26 that qualify on the general cut. 61 entities qualify in at least one cut. Brand pages are not linked in this edition.

12.3 Data availability

The data behind this report is available on request: per-entity metrics across all six cuts, the qualified leaderboards and their brand-only variants, the entity registry with origins, the per-model breakdowns, the origin-share tables on both bases, the source attribution tables, the raw-layer trend file, and the published scale registry.

All figures reflect a single collection day (21 September 2026) and are subject to the limitations in Chapter 11. This is the second edition; the study is repeated quarterly.

Frequently asked questions

About Herm.io & disclosure

Herm.io is a consumer behaviour and marketing data company. We study how people discover and choose brands to help businesses reach the right customers. This report forms part of our public research and is conducted quarterly.

Disclosure & neutrality

The rankings and metrics in this report are entirely impartial and are based on the methodology described within the report. No brand can pay or provide sponsorship to be included in the report, improve its ranking or influence how it is described in the content.

A brand’s score in the report is not an endorsement or assessment of quality; it is solely a measure of its current visibility within AI models. No service or product offered by Herm.io is sold for the purpose of directly changing the results of this report or guaranteeing a ranking. To preserve objectivity, all references to Herm.io’s own domain are removed from the source data before analysis. This ensures that the company does not appear in, measure itself through or benefit from its own study.

Brands seeking to understand their position in the data can schedule a consultation for an impartial assessment of the findings. The consultation is advisory and free of charge.

Written by

Mert Can Elkaya

Contributor

I'm a product builder working at the intersection of product, fintech, and growth. From martech and venture capital to leading product at a proptech platform and co-founding a fintech startup, I help teams—and shoppers—make smarter, more confident decisions.

More from Mert →