Clothing is Turkey's biggest online category. Turkey is the world's most ChatGPT-reliant market for AI referrals. So which apparel brands do the assistants actually name?

Series: Turkish Ready-to-Wear Brands AI Visibility Report
Cadence: Quarterly
Latest edition: July 2026 (baseline)
Publisher: Herm.io — consumer-behaviour and marketing-data research

This is a recurring, neutral study of a question nobody had measured: when a Turkish shopper asks an AI assistant what to wear, which brands does it name?

We ask five large language models the same Turkish-language questions real people ask, repeat each one five times, and record how often, how early, and across how many assistants each brand appears. The result is an AI Visibility Score — a measure of findability, not of quality. A brand ranks highly here because these systems have information about it and surface it readily. That is all it means.

Why this series exists

Three facts sit next to each other, and the gap between them is the reason for this research.

1. Turkish apparel is a serious industry. Turkey’s ready-to-wear and garment sector exported US$16.77bn in 2025 on the exporters’-association registration basis — about 6.1% of Turkey’s total merchandise exports — with 510,628 people employed as of October 2025 (İHKİB, Güncel Durum, January 2026, drawing on SGK statistics; a parallel customs-basis figure from the Ministry of Trade puts exports at US$18.61bn, on a broader product scope — the two are not interchangeable). Germany, the Netherlands and Spain are the largest destinations.

2. Clothing is Turkey’s largest online category — by a distance. In 2025, clothing, footwear and accessories generated TRY 428.7 billion online, up 42.3% year on year, making it the country’s biggest e-commerce product category: 17.45% of all retail e-commerce, at roughly 25.4% online penetration of the category’s total commerce. Those are official ETBİS / Ministry of Trade figures published on 12 May 2026 — not estimates.

3. Turkey is unusually dependent on one AI assistant. By GWI’s Q2 2025 cross-country wave, 39.7% of Turkish internet users had used ChatGPT in the previous month, placing Turkey 11th of 54 markets. Turkey’s official statistics office puts general generative-AI use at a lower 19.2% of all 16–74-year-olds (TÜİK, 2025) — a different population and question, which is why the two figures differ and should not be averaged. Most strikingly, the Digital 2026 report (using Statcounter data) found that 94.49% of Turkey’s AI-originated web referrals came from ChatGPT — the highest share of any market measured, against a global average of 80.9%.

The gap. A large, digitally mature apparel market. A population that has adopted AI assistants quickly and concentrated on one of them. And, until now, no public measurement of which clothing brands those assistants actually recommend. Brands can see their Google rank. They cannot see their AI rank. That is what this series publishes.

What we found first: AI visibility is not market share

The July 2026 baseline edition produced one finding that reframes everything else — and it emerges only when you put our data next to the market’s.

Uniqlo is the clearest case. Fast Retailing’s official European network contains no Turkey operation — the brand does not retail here. Yet in our menswear cut, Uniqlo is named in 6.6% of answers by every one of the five models, enough to qualify above several Turkish brands with hundreds of stores. Nike, meanwhile, suspended direct orders through Nike.com in Turkey following customs changes, and still ranks #15 in the open cut, named in all five models.

The lesson is not that the models are wrong. It is that AI visibility is a separate asset from distribution, revenue or market share — one that is built out of web presence, content structure and the way a brand is written about, and that can be held by a company with no shop on the street. A brand that dominates its category commercially can be near-invisible to an assistant; a brand with no local footprint can be named instantly.

What we measure, and how

• Five models, queried directly. Claude Haiku 4.5, GPT-4o-mini, Gemini 2.5 Flash-Lite, Perplexity Sonar and Grok 4.3 — via API, no system prompt, provider-default temperature, no locale set. The questions are in Turkish; the models infer the market from language alone.
• Real questions, repeated. Each edition uses Turkish-language questions of the kind people actually type, asked five times each to capture run-to-run variation.
• Human-validated brands. Candidate brands are extracted automatically, then reviewed by hand before any metric is computed — false positives dropped, aliases merged, origin labelled, common-word brand names (Only, Next, GAP, Roman, Civil, Network…) matched case-sensitively so ordinary Turkish words are not counted as brands.
• A three-component score. AI Visibility Score = 0.45 × Mention + 0.30 × Position + 0.25 × Breadth — how often a brand is named, how early it appears, and how many of the five assistants know it.
• Multiple cuts. Every edition is read across segments (women’s / men’s / children’s), origin scopes (questions that demand Turkish brands vs. questions that don’t), and behavioural types (discovery / attribute / use-case). Turkish-vs-foreign comparisons are made only on origin-neutral questions — the only fair basis.

What we do not measure. Quality, accuracy, sentiment, price, fit or durability. A score is not an endorsement. This study cannot tell you whether a brand is good — only whether an AI assistant currently knows about it.

Editions

→ Read the July 2026 edition

The wider context: is AI discovery real yet?

We think the honest answer is “real, growing fast, and smaller than the vendors say — but fashion is behind the curve.” The evidence, weighed:

AI answers are demonstrably suppressing traditional clicks. The strongest independent measurement comes from Pew Research Center, which observed 68,879 Google searches by 900 US adults in March 2025: when an AI summary appeared, users clicked a traditional search result in only 8% of visits, against 15% when no summary was present, and ended their session 26% of the time versus 16%. This is behavioural data, not a vendor claim.

AI referral traffic to retail is growing very fast from a very small base. Adobe reported AI-referred retail traffic up 138% year on year in May 2026, converting 54% better than non-AI traffic; Similarweb estimated ChatGPT-referred e-commerce visits converting at 11.4% versus 5.3% for organic search. These are vendor-published figures — Adobe sells commerce and AI analytics; Similarweb sells AI-visibility products — and their releases naturally emphasise growth. The direction is credible; the magnitudes are not settled. An independent working paper analysing 973 sites and $20bn of revenue found ChatGPT traffic performing better than paid social but worse than most traditional channels, which is a useful corrective.

AI’s influence is larger than its clicks. A 2026 working paper matching browsing histories to users’ actual ChatGPT, Claude and Gemini conversations found that when an assistant recommended an unfamiliar brand, the user’s probability of subsequently Google-searching that brand rose 4.3 percentage points, and of visiting the brand’s own site 2.4 points. Much of the journey is search-mediated — meaning referral statistics understate AI’s real effect on which brands people go looking for. (Not yet peer-reviewed.)

But fashion trails. A Vogue Business survey of 251 readers (April 2026) found only 14% used AI often for fashion or beauty, 54% had never used it for those categories, and just 24% trusted its recommendations. Adobe’s early data placed apparel among the weaker-converting AI-referred categories. Clothing involves taste, fit and visual identity — the things assistants are worst at.

Which is precisely why a baseline matters now. The measurement is most valuable before the behaviour becomes mainstream, not after. If AI-assisted fashion discovery in Turkey grows the way general AI shopping has, the brands that are visible when it happens will not be the ones that started optimising afterwards.

What the research actually says about how models pick brands

We are careful not to sell certainty here, because the academic literature does not supply it.

There is no established model of how commercial assistants decide which brands to name. What exists is evidence for individual mechanisms: brand bias (Kamruzzaman et al., EMNLP 2024, found global brands disproportionately associated with positive attributes and local brands with negative ones — directly relevant to a Turkish-vs-foreign study); retrieval dominance (a brand that is not retrieved and not in parametric memory has little chance of appearing at all); position bias (Liu et al., Lost in the Middle, TACL — models use evidence at the start and end of a context better than the middle); and content-level optimisation (Aggarwal et al., GEO, KDD 2024, showed that adding citations, quotations and statistics to a source raised its visibility in generated answers by roughly 30–40% on one metric — though its Perplexity experiment supplied sources as uploaded files rather than testing the open web).

What none of this establishes is which mechanism dominates in a live ChatGPT or Gemini shopping answer, or that more AI mentions cause more sales — for which there is almost no independent evidence at all.

So this series does the one thing that can be done rigorously: measure the output. We do not claim to explain why a model names a brand. We record that it does, how often, and how early — and we publish the method so anyone can check us.

Data, caveats and sources

• Point-in-time. Each edition is a snapshot from a single run. Model versions and web indexes shift; trend claims become possible only from the second edition onward.
• Visibility is not quality. Repeated because it matters.
• Market figures in this hub carry different confidence levels. Export, employment, e-commerce and demographic figures are official or official-derived (İHKİB/TİM, SGK, ETBİS/Ministry of Trade, TÜİK). Market sizes, segment splits, brand market shares and AI-conversion statistics are commercial market-research or vendor-published estimates (MarketLine, Euromonitor, ECDB, Adobe, Similarweb, Salesforce) and are labelled as such where used. Turkish-lira growth rates should be read cautiously: in a high-inflation economy, nominal growth is not real growth.
• Some things are simply not public. There is no official split of Turkish online clothing spend between marketplaces and brand-owned sites; ETBİS publishes only a seller survey (88.3% of surveyed businesses sell via marketplaces, 60.5% via their own site, with heavy overlap). Trendyol publishes no audited domestic GMV. We flag these gaps rather than filling them with estimates.

About Herm.io & disclosure

Herm.io is a consumer-behaviour and marketing-data company. We study how people discover and choose brands so that brands can reach the right customers.

Disclosure & neutrality. Herm.io does not sell SEO or GEO (search/AI-ranking) services, and this research does not recommend any. No brand paid to be included, ranked, or described. A brand’s score is not an endorsement or a judgment of its quality — it is a measure of visibility only. To keep the analysis objective, all citations to Herm.io’s own domain are excluded from the source data before analysis, so the company does not appear in, measure, or benefit from its own study.

Brands that want to understand their position in the data are welcome to book a call for a neutral walkthrough of the findings; this is advisory and free.

Quarterly series · Baseline edition published 13 July 2026