1,125 responses, 5 models, 45 Turkish questions: the baseline AI-visibility snapshot of Turkey's shoes and handbags market

Vertical: Shoes and handbags (ayakkabı ve çanta), Turkey
Method: A single point-in-time study of 1,125 responses across five large language models (Claude, Gemini, ChatGPT, Perplexity, Grok), each asked 45 Turkish-language questions five times
Collection window: 21 August 2026

Key terms: AI visibility, AI Visibility Score, shoe brands, bag brands, ayakkabı markaları, çanta markaları, yerli ayakkabı, yerli çanta, hakiki deri çanta, sırt çantası, spor ayakkabı, LLM brand recommendations, ChatGPT/Gemini/Claude/Perplexity/Grok brand recommendation, Generative Engine Optimization (GEO).

This report reproduces the real Turkish-language questions people ask AI assistants about shoes, boots, sneakers, leather bags and backpacks, and measures how often, in what order, and across how many models five large language models name each brand. Because the same questions can be read with or without an origin restriction, and because footwear and bags turn out to be genuinely different markets rather than two halves of one, the entities are ranked in eight separate cuts, 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

Every number here measures whether an AI assistant names a brand, and how early. It does not measure whether the shoes are comfortable, whether the leather lasts, whether the price is fair, or whether the recommendation was accurate. A brand can be highly visible and mediocre, or excellent and invisible. Read the leaderboards as a map of findability in a new discovery channel — nothing more.

1. Executive Summary

One-sentence takeaway: Turkey’s shoe and bag brands are highly visible when a question asks for Turkish brands and a minority of the answer when it does not — the first vertical in this series where the origin-neutral shelf belongs to foreign names.

• Skechers leads the general leaderboard at 83.96, appearing in 18.13% of all 1,125 responses — 204 answers — and named by all five models. Forelli (81.22) and Derimod (81.03) follow, with Desa (79.65) and Kinetix (76.80) completing the top five. This is the first edition in the series in which a foreign brand tops the overall table.
• 50 entities qualified for the general ranking, the widest field the series has measured. Of 424 tracked entities, 50 cleared the 5% mention threshold across all 45 questions: 26 Turkish and 24 foreign, an almost even split. Their average score is 57.8 and 44 of the 50 are named by all five models.
• The market has two separate shelves. On the 22 questions that ask for yerli brands, 31 of the 32 qualifiers are Turkish and domestic entities take 93.9% of mentions. On the 23 origin-neutral questions, that inverts: 46 of 70 qualifiers are foreign, foreign entities take 70.2% of mentions, and the entire open-cut top six — Skechers, New Balance, Nike, Samsonite, Adidas, Asics — is foreign. The highest-placed Turkish entity in the open cut is Derimod at #7.
• Footwear and Bags are different markets. Only 7 entities qualify in both segments out of 96 qualifying places. Skechers leads Footwear (89.46) ahead of Forelli and Kinetix; Manu Atelier leads Bags (91.79) ahead of Desa and Derimod. Bags is also the more crowded shelf, producing 56 qualifiers on 525 responses against Footwear’s 47 on 600.
• Question shape decides how many brands get named at all. Discovery questions produce 90 qualifiers, attribute questions 42 and use-case questions 27. Open “which brands are good” prompts spray names; task-shaped prompts such as “15.6 inç laptop alan sağlam sırt çantası” converge on a handful, and the leaders change with them — Samsonite tops use_case at 93.98, a cut Skechers only reaches third in.
• The models differ enormously in how they answer. Grok names 14.08 entity mentions per answer against Claude Haiku’s 6.14, and Grok alone answers with marketplaces: it names Trendyol in 73.8% of its responses and Hepsiburada in 51.6%, ahead of every brand. Trendyol’s whole general-cut total is 180 responses, of which 166 are Grok’s — roughly 92% of its visibility comes from one model.
• One model barely searched, and it changed which brands it named. Overall 86.67% of answers were web-grounded, but Gemini grounded only 33.3% of its own; all 150 ungrounded responses in the study are Gemini’s. Restricting the comparison to the 20 questions it answered both ways, 16 of the 20 largest gains from searching go to Turkish brands — Mehry Mu, Manu Atelier, Misela, Mack, Mlouye, Semender Leather — while 14 of the 20 largest gains from answering off memory go to foreign ones, led by New Balance and Adidas. Search pulls this shelf domestic; memory pulls it global.
• Sources are brand-led but unusually fragmented. Brand-owned sites account for 41.12% of citations across 238 domains, and forelli.com.tr is the second most-cited domain in the whole study at 18.6% of responses. But 19.81% of citations come from an “other” bucket spread across 349 domains — the most scattered source base the series has recorded.

Why it matters. Shoppers are shifting discovery from search engines and marketplaces to AI assistants. Independent survey work suggests roughly four in ten Turkish internet users were using ChatGPT monthly in mid-2025 (GWI), while TÜİK’s first measurement of the subject put self-reported generative-AI use at 19.2% of people aged 16–74 in 2025 — different definitions, both indicating a channel that now exists at scale. Footwear and leather goods are categories where Turkey has real manufacturing depth and long-established consumer brands, which makes the split shelf in this data the interesting result: the assistants clearly know which Turkish brands exist, and largely wait to be asked. This is the first (baseline) snapshot of that shelf. One reminder: the numbers describe visibility and information availability, not which brands are best.

2. Methodology

One-sentence takeaway: Five large language models were asked 45 Turkish questions five times each, with no system prompt, producing 1,125 responses. Entities were identified by alias matching plus three rounds of human review and ranked with a three-component (45/30/25) score, computed separately within each of eight cuts.

2.1 Scope and models

The study queried five large language models directly via API with identical questions. All models received only the user question. No system prompt was used, temperature was not set manually (provider default), and no region or locale parameter was defined. Because the questions are in Turkish, the models infer the Turkish market from language alone.

Scale: 45 questions × 5 repeats × 5 models = 1,125 responses. Every cell of the grid existed exactly once, and every response returned a completed status with a non-empty body — so collected equals analysed, and no denominator required adjustment. Denominators are computed per question-model cell from the response table rather than assumed from the design. Of the 1,125, 1,120 named at least one tracked entity; five named none. Single collection window, 21 August 2026.

The final entity dictionary carries 424 entities across 790 aliases, of which 213 are Turkish and 211 foreign. Unlike previous editions, no entity is left at unknown origin: the seven candidates the automated rule could not resolve were settled by manual review. That near-even dictionary is worth holding in mind against the leaderboards that follow — this is a vertical where global brands genuinely compete for the same answer.

The eight cuts. Every metric is computed within a cut, against that cut’s own denominator. Rates from different cuts are not comparable, and neither are scores.

2.2 Web-search configuration

Web search was offered as an enabled tool to all five models; each decided for itself whether to use it (Perplexity’s Sonar is search-grounded by design). Whether a response drew on web search or the model’s own knowledge was inferred from whether the provider returned any citation or source payload: a citation means “web-search,” none means “own knowledge.” This is a reasonable proxy, not direct proof of the model’s internal process.

In this run 86.67% of answers were grounded, but the average is misleading. Four of the five models grounded on every single answer; Gemini grounded on 33.3% of its own, leaving 150 responses drawn from model knowledge alone. Those 150 are counted in the denominator like any other valid response. Any model-level comparison in this report has to be read with that asymmetry in view, and Chapter 6 treats it as a finding rather than a footnote.

2.3 Questions and segments

The 45 questions reproduce the real Turkish-language queries people bring to an assistant about footwear and bags: general discovery (“en iyi ayakkabı markaları”, “yerli çanta markası önerir misin”), specific attributes (genuine leather, durability, comfort, waterproofing, orthopaedic support, price-performance), and use cases (standing all day at work, walking, wide feet, a wedding, a 15.6-inch laptop, cabin-size travel, a gift for one’s mother). They are balanced globally across the three behavioural types — 15 discovery, 15 attribute, 15 use_case — and split 22 local against 23 open. No question was rewritten for this edition.

Unusually for this series, the design is balanced within segments as well as globally:

That balance removes a caveat that limited the previous editions: because each segment carries an equal number of discovery, attribute and use-case questions, a difference between question types cannot be explained away as a difference in which products those questions happened to cover. The remaining design asymmetry is size — Footwear has 24 questions and Bags 21 — which is why distinct-entity counts should not be compared across the two segments even though rankings within each are unaffected.

Of the 45, 22 are “local” (they explicitly ask for Turkish / yerli brands) and 23 are “open” (no origin restriction). The Turkish-versus-foreign comparison in Chapter 8 rests solely on the 23 open questions, the only fair basis for comparing origins, since the local questions cannot surface foreign brands by construction.

2.4 Entity extraction and classification

Entities were extracted case-insensitively from URL-stripped answer text, word-boundary anchored, with Turkish-aware case folding — İ and ı are normalised explicitly, since naive lowercasing corrupts them. Candidates then went through three rounds of human review, resolving to the final 424 entities and 790 aliases. Every rejected candidate is retained with a stated reason.

Two review rounds were caused by defects, and both are worth stating. The candidate filter measured how often a name appears in lowercase in order to detect ordinary words — but it counted URL text, so every boyner.com.tr citation made the brand look like a common noun. Recomputing on URL-stripped text recovered 107 candidates, of which 30 were confirmed entities including Boyner (24 responses), Özözler (20), Mavi, Meskanto, Mounbag, PAEN, Nomatic and Deichmann. Separately, four domain-derived rows (boyner, barcin, paen, ciceksepeti) were duplicating properly-named entities and took their counts, leaving the correctly-named entity scoring zero; these were merged. A third correction restored Rossea, which had been merged into a target that was itself subsequently dropped — a merge into a dropped target deletes the source silently.

Origin follows the trademark, not the operator, derived from the trademark registration against Türkiye. Foreign marks operated in Turkey under Turkish licence are classified Foreign and tagged Licensed-brand so they can be isolated — Nine West and Reebok are the clearest cases here. Twelve entities were corrected by manual review where the register disagreed with ownership. Two of those corrections matter for the leaderboards: Manu Atelier and Mlouye are Istanbul-founded bag houses whose marks are registered in the United Kingdom and the United States respectively, and both are recorded as Turkish. Lumberjack is not: it is an Italian marque now owned and operated by a Turkish group, and it is recorded as foreign, which is why it is the single foreign entity to clear the threshold in the Turkish-only cut.

Entity types. Six classes are tracked: Brand (317), Brand+Retailer (56), Retailer (25), Sub-brand (16), Marketplace (5) and Licensed-brand (5). Sub-brands matter more in this vertical than in previous ones: Kinetix, Polaris and Lumberjack are group own-labels named independently often enough to qualify on their own.

What the brand-only views remove. Only the five marketplaces: Amazon, Hepsiburada, N11, Trendyol and Çiçeksepeti. Vertically integrated own-label retailers such as FLO, Boyner, Decathlon and Barçın are retailers rather than marketplaces and are retained in every table.

Scope fit. 313 entities are classed Core, 106 Adjacent and 5 Out-of-scope. Adjacent entities — fashion houses, luggage specialists and outdoor brands with genuine footwear or bag ranges — are included in all cuts, because they answer questions the study deliberately asked. The five out-of-scope entities (Lenovo, Pegasus, MediaMarkt, Vatan and one non-brand domain) appear in no leaderboard and no origin share; two of them, Lenovo and Pegasus, would otherwise have cleared the threshold, because models named a laptop maker and an airline in answer to questions about laptop backpacks and cabin luggage. They remain in the full entity index in Chapter 12, marked out of scope. No excluded entity sets a component maximum in any cut, so removing them leaves every published score unchanged.

Alias restrictions. Nineteen short all-caps names (CAT, FLO, YDS, MP, PAEN, ÇÇS and others) are matched case-sensitively, and three collision-prone names match only in multi-word form — Mavi Jeans, Panda Safety, Crash Galata — as does On, which matches only as On Running. Those four consequently score zero this edition, a deliberate precision-over-recall trade.

Definitions.

• Mention: a validated entity (or alias) appears in the answer text, counted once per response (presence, not frequency).
• Position (MRR): mean reciprocal of the entity’s first-appearance rank within each answer (first = 1.0; second = 0.5; …), averaged only over the responses where it appears.
• Breadth: number of the 5 models (0–5) that mention the entity within that cut.

2.5 AI Visibility Score (0–100)

Score = 0.45 × Mention + 0.30 × Position + 0.25 × Breadth

Each component is divided by the highest value observed among qualified entities within that cut, so the cut’s leader on any one component scores 100 on that component. Those maxima live in a single scale registry that every output reads and none recomputes.

Worked example, Skechers in the general cut. 204 responses of 1,125 gives a mention rate of 0.1813; the cut maximum is 0.2276 (Derimod), so the mention component is 100 × 0.1813 ÷ 0.2276 = 79.69. Its MRR of 0.5383 against the cut maximum of 0.6991 (Misatra) gives a position component of 77.00. It appears in all five models, so breadth is 100. The score is 0.45 × 79.69 + 0.30 × 77.00 + 0.25 × 100 = 83.96.

The leader scores 83.96, not 100, and that is the expected behaviour rather than a defect. A score of 100 requires one entity to top all three components at once. In this edition no entity does so in the general cut — Derimod leads on volume, Misatra on position and forty-four entities tie on breadth — so the overall leader is a compromise between the three.

Two consequences follow, and both are load-bearing:

1. Scores are not comparable across cuts, editions or verticals. The scale moves whenever the leader moves, so 100 in the open cut is a different absolute quantity from 100 in the local cut. For any comparison beyond this edition, use the raw layer — mention rate, MRR and breadth — which is published for every entity in Chapter 12.
2. An unqualified entity can exceed 100 on a component. Unqualified entities are scored against a scale they did not help set, so 48 of the 2,304 entity-cut rows (2.08%) carry at least one component above 100. The clearest case is Karınca, whose position component reaches 118.7 in the general cut on 44 responses. These are reported and never clipped. No qualified entity can exceed 100 by construction, and none does.

Qualification (≥5% rule): only entities mentioned in at least 5% of that cut’s responses are ranked. This limits the risk that a rare brand which happens to appear first in a few answers inflates the position component. It does not eliminate it — see §3.1 and Chapter 11. Unfiltered metrics for all 418 matched entities appear in Chapter 12.

2.6 Neutrality and self-exclusion

This is a market-wide, neutral study with no focus brand. To prevent any conflict of interest, all citations to Herm.io’s own domains were excluded from the source data before analysis, so the company’s own content neither appears among the most-cited domains nor influences any figure.

3. Overall Visibility Leaderboards

One-sentence takeaway: The general leaderboard is split almost evenly between Turkish and foreign brands — but that balance is an average of two very different shelves, and the local cut shows the domestic one.

As a reminder, these rank visibility, not product quality, durability, price or reliability.

3.1 General leaderboard: 50 qualified entities

Average score of the 50 qualified entities: 57.8. Origin split: 26 Turkish, 24 foreign.

Reading. Skechers tops the table at 83.96 on an 18.13% mention rate — 204 of 1,125 answers — from all five models. It does not lead on volume: Derimod is the most-mentioned entity in the study at 22.76% (256 responses) and finishes third, because it is named comparatively late in the answers where it appears (MRR 0.2571 against Skechers’ 0.5383). Forelli splits the difference at second, with a lower mention rate than either but the strongest position score of the three.

The top ten is seven Turkish to three foreign, and the two groups behave differently within it. The Turkish entries are specialists in the classical sense — Forelli, Derimod, Desa, Kinetix, Manu Atelier, Greyder, Lescon — leather houses and own-label sportswear. The three foreign entries are global athletic brands, Skechers, New Balance and Nike, with Adidas (#12) and Asics (#14) close behind them. There is no foreign leather-goods presence at the top of this table and no Turkish athletic brand above Kinetix.

Fifty qualifiers is the widest field this series has measured, and the reason is visible in the tail. Places 20 to 50 contain a long run of global luxury and outdoor names — Hermès, Timberland, Chanel, Louis Vuitton, Prada, The North Face, Longchamp, Coach, Gucci, Bottega Veneta — that clear the 5% bar without approaching the leaders. Their presence is a property of the questions: asking about handbags in Turkish reliably surfaces the international luxury set alongside the domestic one.

One entry rests on two models. Misatra at #35 qualifies on 5.87% (66 responses) but a breadth of only 2, and its MRR of 0.6991 is the highest in the cut — which means it also sets the position scale that every other entity here is measured against, including the 77.00 position component in the Skechers worked example above. Chapter 11 returns to this; it is the single most consequential thin entry in the edition.

3.2 The same cut, ranked by mention rate

Identical data, ranked by volume alone rather than by the weighted score:

The reordering is instructive. Trendyol climbs from #22 to #7, the largest move in the table, because it is named in 16.00% of all answers but almost always late and by only four models. Greyder rises from #7 to #3 and Tergan from #13 to #8 on the same logic. Moving the other way, Kinetix falls from #5 to #9 and Skechers and Forelli each drop three places, because both are named early when they appear.

Neither ranking is more correct than the other: the first asks how prominently is this brand named, the second asks how often. The AI Visibility Score deliberately blends the two, and readers should hold both pictures in mind for the rest of the report.

What the threshold leaves out is also worth one line. Karınca, a Turkish work-shoe maker, scores 68.35 in the general cut — enough for eleventh place, between Nike and Samsonite — but it is named in only 3.91% of general-cut answers and therefore does not qualify. It appears twice elsewhere in this report, at #12 in Footwear and #4 in use_case, because in those cuts its mention rate clears the bar. The 5% rule is doing real work here, and it cuts both ways.

3.3 Brand-only view (marketplaces removed)

Two of the 50 qualified entities are marketplaces rather than brands: Trendyol (#22) and Hepsiburada (#39). They are kept in the headline ranking because that is how many shoppers meet the shelf, but removing them gives a pure brand-versus-brand picture:

Neither marketplace sits high enough to disturb the top of the table: the leading twenty-one are otherwise unchanged, with Skechers, Forelli, Derimod, Desa and Kinetix still first through fifth. Retailers stay in — FLO at #25 and Beymen at #50 are own-label and multi-brand retailers respectively, not marketplaces — and §2.4 explains why. The brand-only view matters more in the model chapter than here: as Chapter 5 shows, almost all of the marketplace visibility in this study comes from a single model.

3.4 Local leaderboard: 32 qualified entities (Turkish-only questions)

When the 22 questions that explicitly ask for Turkish/yerli brands are isolated, the shelf changes completely:

Average score of the 32 qualified local entities: 60.0. Origin split: 31 Turkish, 1 foreign.

Reading. Forelli leads at 89.09 on 22.36% of local-cut responses, with Desa (86.21) and Manu Atelier (83.35) behind it. Greyder is the most-mentioned entity in this cut at 25.09% and places fourth on the same volume-versus-position trade seen in §3.1. Skechers, Nike, Adidas, New Balance and Asics — the five names that dominate the origin-neutral shelf — appear nowhere in this table at all.

The origin filter is respected almost absolutely. 31 of 32 qualifiers are Turkish and domestic entities take 93.9% of the cut’s mentions. The single exception is Lumberjack at #16, an Italian marque owned and operated by a Turkish group, which by this study’s trademark rule counts as foreign. That the assistants treat it as a Turkish answer is defensible from a shopper’s point of view and is the kind of edge the trademark rule is designed to expose rather than hide.

Two structural points sit underneath the table. First, the local cut produces 32 qualifiers against the open cut’s 70 — asking for Turkish brands narrows the field by more than half, which is the mirror image of the finding in Chapter 8. Second, the bottom of this table is thin: Blinq (#32) is named by two models and Cakard, Gön, Bago and Balkan Deri Çanta by three, against five for most of the top fifteen. Breadth is doing the work at the bottom of this table, and small differences there should not be over-read.

4. Segment Leaderboards

One-sentence takeaway: Footwear and bags are not two halves of one market — only seven entities qualify in both, and each segment has a different leader, a different origin balance and a different shape.

The two segments partition the same 45 questions by product category, and both carry a balanced 8/8/8 and 7/7/7 split of question types (§2.3). Because Footwear has 24 questions and Bags 21, counts of distinct entities should not be compared across segments; rankings within each are unaffected.

4.1 Footwear (24 questions, 600 responses, 47 qualifiers)

Skechers takes it at 89.46 on a 34.00% mention rate, ahead of Forelli (84.83) and Kinetix (80.98). Greyder is again the volume leader at 33.83% and places fourth. The top eleven alternate almost perfectly between Turkish and foreign — Skechers, Forelli, Kinetix, Greyder, New Balance, Lescon, Nike, Adidas, Asics, Lumberjack, Scooter — which is a fair summary of the segment as a whole: 25 of the 47 qualifiers are Turkish, but foreign brands take 50.2% of mentions.

Two entries reward a closer look. Karınca at #12 qualifies on a 7.33% mention rate with an MRR of 0.830 — the highest position score in the segment. It is a work-and-comfort shoe specialist that the models reach for on standing-all-day and workplace questions and rarely otherwise, which is exactly the profile the position component is designed to surface. YDS at #13 shows the same pattern more mildly (11.00%, MRR 0.600). Neither is a general-purpose answer; both are the answer to a specific question.

4.2 Bags (21 questions, 525 responses, 56 qualifiers)

Manu Atelier leads at 91.79, level with Desa on mention rate (30.10% each) and ahead on position. Derimod (82.23) is third, Samsonite (78.61) fourth and Tergan (75.26) fifth. This is the more domestic of the two segments on volume — Turkish entities take 56.8% of mentions here against 49.8% in Footwear — and the more crowded: 56 qualifiers on 525 responses.

The composition differs from Footwear in kind, not just in degree. Where footwear pits Turkish own-labels against global athletic brands, bags pit Turkish leather houses — Desa, Derimod, Tergan, Misela, Mehry Mu, Mlouye — against international luxury: Hermès (#9), Chanel (#11), Louis Vuitton (#13), Prada (#14). Between the two sits a functional layer of luggage and backpack specialists led by Samsonite and Eastpak. Three distinct competitive sets share one leaderboard, and a brand’s rank means something different depending on which it belongs to.

Misatra at #15 is the thin entry here: 12.57% of responses, an MRR of 0.699 that is the highest in the segment, and a breadth of two. As in the general cut, it sets the position scale that every other entity in this segment is measured against.

4.3 What crosses between them

Only seven entities qualify in both segments, out of 96 qualifying places: Derimod, Desa, Tergan, Hotiç, The North Face, Trendyol and Hepsiburada. Two of those seven are marketplaces and one is an outdoor brand, which leaves four — Derimod, Desa, Tergan, Hotiç — Turkish leather houses whose ranges genuinely span both categories.

That is the clearest structural finding in this edition. A brand’s AI visibility in Turkish footwear tells you almost nothing about its visibility in Turkish bags, and the two leaderboards should be read as two markets that happen to share a retail aisle. It also explains the general cut: Skechers tops it not because it dominates the category but because Footwear carries three more questions than Bags, and because the entity that dominates Bags — Manu Atelier — does not appear in footwear answers at all.

5. Differences Between Models

One-sentence takeaway: The five models broadly agree on the leading Turkish leather houses, but they differ enormously in how many names they give, and one of them answers with marketplaces rather than brands.

5.1 Per-model behaviour summary

The clearest split is breadth of answer. Grok names 14.08 entities per answer and Gemini 11.24, while Claude Haiku names 6.14 — less than half. ChatGPT (7.80) and Perplexity (7.98) sit between. Unlike the previous edition, breadth per answer and breadth of vocabulary move together here: Grok also has the widest vocabulary at 288 distinct entities and Claude the narrowest at 196.

This matters for interpretation. A specialist in the long tail has a materially different chance of being named depending on which assistant the shopper opens. In Grok or Gemini, an answer has room for fifteen names and small Turkish leather workshops appear in it; in Claude, an answer listing six names rarely goes beyond the best-known houses — with one striking exception, covered next.

5.2 Each model’s most-named entities

The tables below give each model’s five most-mentioned in-scope entities (rate = % of that model’s 225 answers naming the entity).

No entity is the most-named in more than two models. Derimod leads Gemini (36.0%) and Perplexity (24.4%), Desa leads ChatGPT (26.7%), Forelli leads Claude (32.0%) and Trendyol leads Grok. That is a markedly lower level of agreement than the previous editions of this series, where one brand topped three or four models — and it is consistent with a fragmented category rather than a concentrated one.

Claude is the outlier on the domestic side. Its top five includes Misatra at 18.7% and Balkan Deri Çanta at 12.4%, two small Turkish makers that barely register elsewhere. Misatra’s entire study-wide total of 66 responses comes from two models; 42 of them are Claude’s. The model with the narrowest vocabulary spends a disproportionate share of it on a handful of specialists — which is why Misatra, on two models, ends up setting the position scale in three of the eight cuts.

Grok is the outlier on the retail side. It names Trendyol in 73.8% of its answers and Hepsiburada in 51.6%, putting both ahead of every brand; Derimod, its highest-ranked brand, reaches 29.8%. The effect on the study-wide numbers is large and worth stating plainly: Trendyol appears in 180 general-cut responses, 166 of which are Grok’s — about 92% — and Hepsiburada in 125, of which 116 are Grok’s. Both marketplaces qualify in the general cut essentially on the strength of one model’s habit of answering “where to buy” rather than “what to buy.” The brand-only views in §3.3 exist for readers who consider that a different question.

A note on what these tables exclude. Each model also named entities that appear in no leaderboard — component trademarks, corporate parents and adjacent-category brands classed out of scope. The pattern is not evenly spread. Gemini named Gore-Tex in 14.2% of its answers and Vibram in 9.3%; Grok named Ziylan Grup in 9.8% and FLO Grubu in 4.4%. The first pair are materials, not brands a shopper can buy; the second pair are the holding companies behind FLO, Kinetix, Polaris and Lumberjack. Both are real differences in answer usefulness that the leaderboards, by design, do not show — and the Ziylan mentions are the reason that name is excluded from the entity set entirely (§2.4).

6. Search vs. Memory

One-sentence takeaway: One model searched on only a third of its answers, and comparing its grounded answers against its ungrounded ones on the same questions shows a clear directional effect: search surfaces Turkish brands, memory surfaces global ones.

The aggregate grounding rate of 86.67% conceals a near-binary split between models.

All 150 ungrounded responses in the entire study are Gemini’s. That fact governs how this chapter can be read. An all-models comparison of grounded against ungrounded answers would not measure grounding at all — it would compare Gemini against the other four models and call the difference a search effect. The only defensible comparison is within Gemini.

Even that is not quite enough. Gemini’s grounding is not random across the question set: it searched on all five repeats of 7 questions, on none of the five for 18 questions, and variably on the remaining 20. Comparing all 75 grounded answers against all 150 ungrounded ones would therefore partly compare questions rather than grounding. This chapter uses only the 20 questions Gemini answered both ways — 40 grounded responses against 60 ungrounded — which holds the question mix constant and isolates the effect of searching.

6.1 What changes when Gemini searches

Reading. The direction is unambiguous and it runs along origin. Sixteen of the twenty largest gains from searching go to Turkish brands — Mehry Mu (+15.8pp), Manu Atelier (+14.2pp), Misela (+14.2pp), Classone and Mack (+13.3pp), Mlouye (+12.5pp) — and two of them, Semender Leather and Rossea, are never named at all when Gemini does not search. In the other direction, fourteen of the twenty largest gains from memory go to foreign brands: New Balance (+11.7pp from memory), Adidas (+10.8pp), Skechers (+8.3pp), Clarks, Geox, Asics. Across the full comparison, Turkish entities outnumber foreign ones 54 to 30 among those that gain from searching, and foreign outnumber Turkish 40 to 20 among those that gain from memory.

The reading that fits is the simple one: a model’s stored knowledge of this category is weighted toward globally-marketed brands, and the live Turkish web is where the domestic specialists live. When Gemini does not look, it answers a Turkish question with an international shelf. When it does look, the Turkish leather houses appear.

This is a single-model result on 100 responses and should not be generalised to the other four assistants, all of which searched on every answer. But it is the first time in this series that grounding has produced a directional shift rather than a marginal one, and it has a practical implication worth stating plainly: for a smaller Turkish brand, being findable on the live web is not a supplement to being known — in this data it is the entire mechanism by which one model names you at all.

The other four models ground on every answer, so this report largely measures what they retrieve and select, not what they know unprompted — and retrieval is sensitive to the live web, which changes independently of the models. The grounding rate is measured every edition.

7. Question-Type Ownership

One-sentence takeaway: How a question is phrased changes not only who is named but how many brands get named at all — discovery questions produce 90 qualifiers, use-case questions 27.

The 45 questions are read through three behavioural types: discovery (general “best / recommended”), attribute (specific qualities such as genuine leather, durability, comfort or waterproofing) and use_case (situation-driven, such as standing all day, a wedding, a 15.6-inch laptop or a gift). Each is a cut in its own right with an identical denominator of 375 responses, and — unusually for this series — each is balanced within both segments, so the differences below are not a product of question mix.

Reading. The instability is the finding, and it is the reverse of the previous edition. Each question type has a different leader — Manu Atelier on discovery, Skechers on attribute, Samsonite on use_case — and only sixteen entities qualify in all three cuts. There is no brand in this vertical that owns the conversation regardless of how the question is put.

The width of the answer changes even more than its content. 90 entities clear the 5% threshold on discovery questions, 42 on attribute and 27 on use_case. A question like “en iyi çanta markaları” produces a long list in which many brands get a mention; “kabin boy seyahat sırt çantası hangi marka iyi” produces a short one in which a few do. Fifty-six entities qualify on discovery questions and on neither of the other two types — they are visible only in the broadest possible prompt.

That concentration also reshapes the top. Use_case is where position scores run highest in the entire study: Forelli’s MRR of 0.811, Karınca’s 0.848 and Samsonite’s 0.678 are all first-or-second-name-mentioned averages. When an assistant answers a specific task, it leads with one recommendation rather than surveying the market, and the brands that own those tasks are named first or not at all.

Three entities are worth pulling out of the specific cuts. Karınca places #4 on use_case at 11.47% with the highest MRR in the study (0.848), carried by standing-all-day and workwear questions. Mack places #6 on use_case (13.33%) and #7 in Bags, a school-and-work backpack maker that surfaces on task questions and disappears on discovery ones. Moving the other way, Hotiç is #8 on discovery at 29.87% and does not reach the use_case top ten at all — a well-known name that comes up when the question is “which brands are good” and not when it is “what should I buy for this.”

8. Open Market: Turkish vs. Foreign

One-sentence takeaway: Across origin-neutral questions, foreign brands take 70.2% of mentions and the entire top six — the first vertical in this series where the domestic industry does not lead the unrestricted shelf.

This chapter relies only on the 23 open questions with no origin restriction (575 responses), the only fair basis for comparing Turkish and foreign brands. A blended figure across all 45 questions would be misleading, because 22 of them constrain the answer to domestic brands by construction.

8.1 Origin share

Origin share can be measured two ways, and both are reported here because they answer different questions:

The two measures point the same way and diverge slightly in magnitude: Turkish brands hold about a third of the positions on the origin-neutral shelf but under 30% of the volume, because the largest entities in this cut are foreign. Skechers alone accounts for 202 open-cut mentions, half as many again as Derimod, the most-mentioned Turkish entity in this cut, and nearly four times Manu Atelier’s total.

The local cut inverts it. On the 22 questions asking for Turkish brands, domestic entities take 93.9% of mentions and 31 of 32 qualifying places. The gap between 29.8% and 93.9% is 64 points — the widest origin gap this series has recorded, and the defining number of this edition. Both figures must be stated together; neither is meaningful alone.

8.2 Open-market top 15 entities (any origin)

8.3 The pattern: two shelves, not one

The individual ranking shows what the aggregate hides. The open-market top six is entirely foreign — Skechers, New Balance, Nike, Samsonite, Adidas, Asics — and the first Turkish name is Derimod at #7, on a 23.65% mention rate. Kinetix (#10) and Forelli (#11) follow, and then the table returns to global names for another ten places.

The composition of that foreign set is the real finding, and it splits cleanly in two. In footwear questions, the origin-neutral answer is global athletic: five of the top six are sportswear brands with no particular Turkish connection. In bag questions, it is a mix of functional luggage (Samsonite, Eastpak, Osprey) and international luxury (Hermès, Chanel, Louis Vuitton, Prada). Turkish brands are present in both, but as the second tier rather than the first.

This is a description of model behaviour, not a verdict on the industry — and here the two are harder to reconcile than in previous editions. Turkey has substantial footwear and leather-goods manufacturing, and several of the entities on this leaderboard are long-established consumer brands with their own retail networks. The assistants clearly know they exist: given a yerli filter, they produce 32 Turkish qualifiers without hesitation and place foreign names almost nowhere. What the open cut shows is not ignorance but default. When nothing in the question points toward Turkey, the default answer for shoes is a global sportswear brand and the default answer for a luxury bag is a European house.

That is a different problem from invisibility, and it has a different shape. A Turkish shoe brand in this data is not competing to be known — it is competing to be the answer to a question that never mentions Turkey. The gap between 29.8% and 93.9% is the size of that gap, and it is the number this series will track across future editions.

9. The Discovery Ecosystem: Where AI Learns About Brands

One-sentence takeaway: Brand-owned sites are the largest single source of what these assistants read — and one brand’s own website is the second most-cited domain in the entire study.

When an assistant answers with web search, it draws on the content available to it at that moment. We collected the source addresses returned by the models, reduced them to domains, and ranked them by how many distinct answers cited each, giving a map of where these systems read about shoes and bags. Across the run the models cited 691 distinct domains.

9.1 The 15 most-cited domains

The most-cited domain is the marketplace trendyol.com at 21.0% of responses, but the entry immediately behind it is the more interesting one. forelli.com.tr is cited in 18.6% of all responses — 209 answers, from all five models — making a single brand’s own website the second most-read source in the study, ahead of Hepsiburada, Çiçeksepeti and every editorial site. Eight of the fifteen most-cited domains are brand-owned, and six of those eight belong to entities in the general leaderboard.

Two other patterns are worth naming. Editorial sources are stronger in this vertical than in previous ones: oggusto.com (#3, 12.9%) and eniyimarka.com.tr (#5, 10.2%) are both Turkish editorial and “best of” sites cited by three to five models. And forums are genuinely present: eksisozluk.com is cited in 9.6% of responses and forum.donanimarsivi.com in 5.4% — user-generated discussion appearing above most brand sites, which is not something the home-textiles edition showed.

One caveat sits on the eighth row. misatra.com is cited in 8.8% of responses but by only two models, mirroring the entity’s breadth-2 profile in the leaderboards. A domain cited heavily by a minority of models is a different signal from one cited moderately by all five, and the Models column is there to let readers separate the two.

9.2 Source-type mix

9.3 The fragmentation finding

Brand-owned sites account for 41.12% of all citations across 238 domains — the largest single class, but well short of a majority, and notably lower than the 52.82% recorded in the home-textiles edition. The commercial classes together — brand-owned, retailer and marketplace — reach 56.02%. Editorial contributes 18.86% across just 67 domains and forum/social 5.31% across only 11.

The residual is the story. 19.81% of citations come from an “other” bucket spread across 349 domains — more than half of every domain cited in the study, contributing a fifth of the reading. These are small independent retailers, regional news sites, and commerce pages with no tracked entity behind them: ayakkabidunyasi.com.tr, marmarisyenisayfa.com and hundreds like them, most cited in a handful of answers each.

Two things follow. First, this is the most fragmented source ecosystem the series has measured. No small set of intermediaries controls what the assistants read about Turkish shoes and bags; the long tail is genuinely long. Second, the concentration that does exist is unusually top-heavy within its class: 238 brand-owned domains produce 41.12% of citations, but a single one of them — forelli.com.tr — produces nearly a fifth of all responses’ worth of citations on its own.

Read as a description rather than advice: in this market the assistants are reading a wide, shallow web, and the brands that appear most often in it are the ones with substantial, current content on their own domains. That is not a claim that publishing more will raise a score — it is an observation about where these systems currently read.

10. What the Patterns Suggest

One-sentence takeaway: The visible names share breadth across all five models and a clear product context; the widest gap in the data is between being known and being the default.

This chapter is a neutral reading of the data. It describes patterns associated with visibility; it is not advice, a service, or a recommendation, and visibility remains separate from quality.

1) Breadth is the baseline condition. 44 of the 50 qualified general entities appear in all five models. The exceptions are instructive: Misatra (2 models), Gön (3), and Tergan, Trendyol, Hepsiburada and Beymen (4 each). Broad model coverage is what the visible entities share; an entity strong in one assistant and absent from three does not rank. Fourteen qualifying places across the eight cuts rest on fewer than three models, and every one of them sits in the lower half of its table.

2) Mention and position are different signals, and in this vertical they diverge sharply. Some entities are named often but late — Derimod (22.76%, MRR 0.257), Greyder (18.40%, 0.277), Hotiç (14.04%, 0.206). Others are named rarely but first — Karınca (3.91% general, MRR 0.848 in use_case), YDS (6.13%, 0.584), Mack (6.04%, 0.586), Samsonite (9.24%, 0.559). Both patterns are real and they describe different market positions: one is a name that comes up in a list, the other is a name that comes up as the answer.

3) Product context determines the leader, not the category. Only seven entities qualify in both Footwear and Bags. Manu Atelier leads bags and is invisible in footwear; Skechers leads footwear and does not qualify in bags. A brand’s visibility here is not a single number; it is a number per product context, and the useful question is which segment and which question type a brand owns, rather than where it sits overall.

4) The question does more work than the brand. Discovery questions produce 90 qualifiers, use_case 27 — and the leaders change with them. Fifty-six entities qualify on discovery alone. For most brands in this vertical, visibility is not a property they carry into every conversation; it is contingent on the shape of the prompt, and the brands that clear the bar in all three types (sixteen of them) are the exception rather than the pattern.

5) The gap here is default, not recognition. This is the finding that separates this vertical from the previous editions. Turkish brands are not missing from the models’ knowledge: given a yerli filter they produce 32 qualifiers and 93.9% of mentions. They are missing from the default answer: on origin-neutral questions they take 29.8% of mentions and none of the top six places. Of 424 tracked entities, 50 qualify in the general cut and 111 in any cut at all, which leaves 313 tracked entities qualifying nowhere. For most Turkish shoe and bag businesses the relevant gap is not that assistants prefer foreign brands when asked to compare — it is that nothing in an ordinary question prompts the assistant to consider a Turkish one.

How a brand can locate itself in the data. Using the dataset (available on request), a brand can read five things in order: whether it appears in all five models; how often it is mentioned within its cut; how early it is named when mentioned; which segment and question type it is strongest and weakest in; and — most diagnostic in this vertical — whether it appears in the open cut as well as the local one. These readings show where a visibility gap sits, without saying anything about whether its product is good.

11. Limitations and Notes

One-sentence takeaway: This is a single, point-in-time snapshot of one Turkish vertical, scored on a scale that does not travel between editions; read the figures within the limitations below.

• Single collection window. All data comes from one run on 21 August 2026. Model outputs drift; these figures are a snapshot, not a stable property of the models. This is the first (baseline) edition; quarterly editions will enable trend analysis.
• Scores are not comparable across cuts, editions or verticals. Each component is scaled against that cut’s maximum, so 100 in the open cut is a different absolute quantity from 100 in the local cut, and the scale moves whenever the leader moves. Any comparison beyond this edition uses the raw layer — mention rate, MRR and breadth — published for every entity in Chapter 12.
• One low-breadth entity sets the position scale in three cuts. Misatra is named in 66 responses by two models. It qualifies in four cuts — general, local, Bags and attribute — and in three of them (general, Bags and attribute) its MRR is the cut maximum, so every position component in those cuts is expressed against a scale supplied by a single two-model entity. This does not change the order of any leaderboard, but it compresses the position scores of everything below it, and a future edition in which Misatra’s breadth changes will move those numbers for reasons that have nothing to do with the other brands. For reference, the position scale is set by Mack in the local cut, Samsonite in the open cut, Karınca in Footwear and use_case, and Manu Atelier in discovery.
• The score rewards position as well as frequency. This is by design, but it produces counterintuitive placings — Karınca reaches #12 in Footwear and #4 in use_case on mention rates of 7.33% and 11.47%, on the strength of MRRs of 0.830 and 0.848. §3.2 gives the mention-rate view of the general cut for comparison; read thin-mention ranks cautiously everywhere.
• Fourteen qualifying places rest on fewer than three models. Misatra (four cuts), Hepsiburada (three), Trendyol, Gön, Semender Leather, Carfier, Polène, Blinq and Cakard account for the rest. All are in the lower half of their tables and should be read as indicative.
• Fourteen qualifying entities carry sub-High classification confidence in the entity dictionary, including New Balance, Lescon, MANC, Merrell, Chanel, Eastpak, Misatra, Karınca and İnci. Classification errors in this group would show up in a published table rather than only in the appendix.
• Never compare local against foreign across all 45 questions. The local cut constrains answers to domestic brands by construction. Chapter 8 uses only the 23 open questions, and both denominators are stated wherever an origin figure appears.
• Do not compare distinct-entity counts across the two segments. Footwear carries 24 questions and Bags 21, so Bags producing more qualifiers (56 against 47) partly reflects a smaller denominator. Rankings within each segment are unaffected.
• Two segments only. A finer product cut — boots, sneakers, backpacks, luxury handbags — puts every segment below the eight-question floor this series uses, so Footwear/Bags is the whole product axis this edition. Several findings above (the luxury cluster in bags, the athletic cluster in footwear) would benefit from a finer cut and cannot get one yet.
• Grounding is not uniform, and one model supplies the entire ungrounded set. All 150 ungrounded responses are Gemini’s. Chapter 6 therefore compares Gemini against itself, and further restricts the comparison to the 20 questions it answered both ways, because its grounding is not randomly distributed across the question set. The result is a single-model finding on 100 responses and should not be read as a general property of web search.
• Marketplace visibility is largely one model’s habit. Trendyol and Hepsiburada qualify in the general cut, but 166 of Trendyol’s 180 mentions and 116 of Hepsiburada’s 125 come from Grok. Readers who consider marketplaces out of place in a brand ranking should use the brand-only views in §3.3; readers who do not should still be aware that this part of the leaderboard is not a five-model consensus.
• Out-of-scope entities are excluded, and two of them would otherwise rank. Five entities appear in no leaderboard and no origin share. Lenovo and Pegasus both cleared the 5% threshold in the open cut — an airline and a laptop maker named in answer to questions about cabin-size travel backpacks and 15.6-inch laptop bags. They are tracked and visible in Chapter 12, marked out of scope. No excluded entity sets a component maximum, so no published score changes as a result of the exclusion.
• Origin follows the trademark, which has consequences. Twelve entities were corrected by manual review where the register disagreed with ownership. Manu Atelier and Mlouye are recorded as Turkish despite UK and US trademark registration; Lumberjack is recorded as foreign despite Turkish ownership and operation, and is the only foreign entity to qualify in the Turkish-only cut; Neri Karra is recorded as foreign despite a Turkish mark, on the basis of Bulgarian operation. Readers who would draw those lines differently should note that Manu Atelier’s classification alone moves the Bags leader between origins.
• Sub-brands under-count where models name the parent. Kinetix, Polaris and Lumberjack are named independently often enough to qualify; smaller own-labels are not. Separately, Ziylan Grup was classified a non-brand corporate group and its 39 responses are not credited to FLO, Kinetix, Polaris or Lumberjack — a decision that reduces those brands’ measured visibility relative to a reading that treats a group mention as a brand mention.
• Six dictionary entities have zero matches and are carried for completeness: Crash, Mavi and Panda are alias-restricted collisions that match only in multi-word form, and dalkilicspor, genclercanta and modakozmetikcanta are organisations found only through cited domains and never named in any answer.
• Precision over recall in matching. Nineteen short all-caps names are matched case-sensitively and four names are restricted to multi-word forms, so brands mentioned indirectly or in unusual casing may be undercounted. Diga and Özder share an identical response set — both are sub-brands of Özözler Deri always named in the same sentence — and were verified as co-mentions rather than a split entity.
• Components above 100 are reported, not clipped. 48 of 2,304 entity-cut rows (2.08%) carry a component above 100 because unqualified entities are scored against a scale they did not set. The highest is Karınca’s position component at 118.7 in the general cut.
• Turkish-language prompts only. Results may not transfer to the same questions asked in English about the Turkish market.
• Small score gaps are not meaningful. In the packed middle of each leaderboard, differences of a point or two should not be over-interpreted; only larger gaps are robust. Mehry Mu and Brooks are separated by 0.11 points in the general cut, and Chanel and Mlouye by 0.10.
• Visibility is not quality. This report measures only mention and position. 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.
• Models are probabilistic. The same question can produce different answers; five repeats reduce but do not eliminate this.
• Source-type mix is derived, not independently classified. A domain’s type comes from the entity dictionary via root-website ownership, falling back to a name match on the domain label. Treat the mix as indicative, particularly the 349-domain “other” bucket, which is a residual rather than a category. Source coverage also varies by model.
• Output-length limits and model versions differ and date quickly; findings are specific to the versions in Chapter 2 as of August 2026.

Frequently Asked Questions

12. Appendix & Data

12.1 All questions (45) and their categories

Distribution: 15 discovery, 15 attribute, 15 use_case; 22 local and 23 open; across two product segments (Footwear 24, Bags 21), balanced 8/8/8 and 7/7/7 by question type within each. No question was rewritten for this edition.

12.2 All tracked entities (418 matched)

12.3 Data availability

The study tracked 424 entities, of which 418 matched at least once; 50 cleared the 5% threshold in the general cut, 32 in the local cut, 70 in the open cut, and 111 in at least one of the eight cuts. To support scrutiny and reproduction, the underlying data is available on request: response-level entity mentions with grounding flags, all eight qualified leaderboards and their brand-only variants, full unfiltered entity metrics, per-model and per-question-type breakdowns, the open-market breakdown, the full 691-domain source list, the Gemini grounded-versus-memory comparison, the entity dictionary with its decisions log and every rejected candidate with a reason, and the methodology and data dictionary.

All figures reflect a single point-in-time run (21 August 2026) and are subject to the limitations in Chapter 11. This is the first (baseline) measurement; the study will be repeated quarterly.

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.