1,125 responses, 5 models, 322 brands: ask for a Turkish brand and you get one — stop asking, and the shelf goes foreign

Vertical: Cosmetics, skincare, makeup & personal care — Turkey
Method: A single point-in-time study of 1,125 responses — five large language models (Gemini, ChatGPT, Claude, Grok, Perplexity), each asked 45 Turkish-language questions five times
Collection window: 4–5 September 2026
Edition: 2 of a quarterly series. Edition 1 (May 2026) covered 27 questions and 4 models.

Key terms: AI visibility, AI Visibility Score, cosmetic brands, Turkish cosmetics market, Turkish (yerli) cosmetic brands, LLM brand recommendations, ChatGPT/Gemini/Claude/Grok/Perplexity brand recommendation, Generative Engine Optimization (GEO).

This report reproduces the real Turkish-language questions consumers ask AI assistants and measures how often, in what order, and across how many models five large language models name each cosmetic brand. 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 cosmetics 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, safer, healthier or more effective. 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. Readers and consumers 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: The single strongest determinant of which cosmetic brands an AI assistant names in Turkey is not the brand — it is whether the question contains the word “Turkish.”

• The Purest Solutions leads the full 45-question board. It appeared in 33.2% of all 1,125 responses, was named by all five models, and scores 93.92. Golden Rose (89.64), Flormar (82.61), Maybelline New York (82.08) and La Roche-Posay (76.44) complete the top five. Two of that five are foreign, which is new for this series.
• The market has two entirely different shelves, and the question chooses between them. On the 22 Turkish-constrained questions, Turkish entities take 98.3% of qualifying mentions and 38 of the 40 qualifiers are Turkish. On the 23 origin-neutral questions, Turkish entities take 30.8%, only 17 of 53 qualifiers are Turkish, and the top five — Maybelline New York, La Roche-Posay, L’Oréal Paris, Estée Lauder, Bioderma — are all foreign.
• This is not a reversal of May, and should not be read as one. Edition 1’s origin-neutral comparison rested on 5 questions and 100 responses. This one rests on 23 questions and 575 responses, 18 of which are new product-level queries such as “akmayan rimel önerisi” (a mascara that doesn’t run). We did not watch a shelf move; we built a much larger one and looked at it properly.
• The five assistants do not agree, and one of them disagrees on origin. Grok names 15.2 entities per answer, Claude Haiku 6.2. Perplexity has the widest repertoire at 207 distinct entities. And Claude Haiku 4.5 is the only model whose most-named brand is not Turkish — it leads with L’Oréal Paris at 23.1%, while The Purest Solutions, the overall leader, appears in only 7.6% of its answers.
• Turkish challengers gained on the carried questions; the incumbents did not. On the 27 questions repeated verbatim from May, 43 brands rose and 60 fell. The largest gains went to Maruderm (+16.11 points) and Bioxcin (+12.41). The largest falls went to Farmasi (−10.37) and Atelier Rebul (−10.18). The leader itself slipped 3.88 points and still led comfortably.
• Gemini stopped searching. Its web-grounding rate fell from 84.4% in May to 26.2% — all 225 calls completed normally, so this is model behavior, not a collection failure. Every other model grounded between 96.4% and 100%.
• Retailers and marketplaces are now part of the shelf. Gratis (#18), Trendyol (#19), Watsons (#22), Sephora (#35) and Hepsiburada (#36) all qualify on the full question set. Four in five of Trendyol’s mentions come from a single model.
• The source ecosystem is long-tailed and increasingly brand-owned. Across 770 cited domains, 52.7% of citations come from a long tail of small sites; brand-owned sites now account for 22.6%. Only nine domains were cited by 100 or more responses.

Why it matters. Consumers are shifting product discovery from search engines to AI assistants. Whether a brand appears in those assistants’ answers is becoming a discovery channel in its own right. This edition doubles the question set and adds a fifth model, and in doing so it exposes something edition 1 was too small to see clearly: a Turkish cosmetics brand can be highly visible in the domestic conversation and nearly invisible the moment the consumer stops specifying origin. A reminder: the numbers below 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 — to produce 1,125 responses; every one of 2,832 entity candidates passed a human review gate, and scores are computed separately within each published cut against a published scale.

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, models infer the Turkish market from language alone; an explicit locale was deliberately not set.

Four of the five providers carry over from edition 1. Perplexity Sonar is new. Within the carried four, only Grok is genuinely unchanged: Gemini moved from 3.5-flash to 3.6-flash, ChatGPT from 5.5 to 5.6-luna-pro, and Claude moved from Sonnet 4.6 to Haiku 4.5 — a tier change, not a version bump. Chapter 8 explains what that does to cross-edition comparison and how this report handles it.

2.2 Web-search configuration

Web search was available to every model, and each model decided for itself whether to use it. Overall 84.5% of responses were web-grounded. The distribution is extremely uneven:

Claude, Grok and Perplexity grounded every answer. ChatGPT grounded 96.4%. Gemini grounded 26.2%, down from 84.4% in May. All 225 Gemini calls completed and returned substantive text, so this is a change in how the model chooses to answer, not a data-collection fault. It means roughly 174 of the 1,125 responses in this study were produced from model memory rather than retrieval, and about 166 of those are Gemini’s.

Citation payloads had to be read differently for each provider. Gemini returns grounding text segments rather than URLs, with a redirect wrapper on the URI, so domains were taken from the search-results title field. Grok returns citation titles as bare numerals, so domains were taken from the URL. Perplexity and Claude embed the full text of cited pages inside the citation payload; that payload was never matched against brand names, because doing so would credit a brand quoted inside a cited page but never named by the model.

2.3 Questions and cuts

45 questions, all asked exactly as a consumer would type them. 27 are carried verbatim from edition 1; 18 are new, all origin-neutral and product-level (“kalıcı mat ruj önerisi”, “beyaz iz bırakmayan yağsız SPF 50 güneş kremi hangi marka iyi?”). No question was reworded, and none was excluded after collection.

Within-type scope is deliberately uneven — discovery skews origin-neutral, use-case skews Turkish-constrained — because the 27 carried questions could not be reworded without breaking the trend. Read per-type origin splits with that in mind.

Every metric is computed separately within each of eight published cuts:

2.4 Entity resolution

Every one of 2,832 candidate names above a three-response floor was reviewed by a human. 1,449 were dropped as non-entities, 1,079 were merged into a parent, and 322 were kept — 298 brands, 12 retailers, 4 marketplaces, 5 licensed brands and 3 sub-brands. Every kept entity carries a verified website or an explicit blank, and every one carries a source URL for its origin decision.

Four parts of that process are worth stating, because each is a place where a ranking of this kind usually goes wrong.

Generic words look exactly like brands. The most frequent extracted candidate in the whole corpus was “türkiye” (672 responses), followed by “cilt” (skin, 513) and “türk” (492). None is a brand. So is “solutions,” which appears in 373 responses — every single one of them a fragment of The Purest Solutions.

Product lines are not brands. “Anthelios” appeared in 89 responses; it is La Roche-Posay’s sunscreen range, and was merged into it. “Lash Sensational” (88), “Fit Me” (77) and “Sky High” (73) are all Maybelline products and were merged into Maybelline New York. Left alone, each would have entered the leaderboard as a mid-table brand.

One brand can be written several ways. L’Oréal Paris was reassembled from three spellings — differing only by apostrophe style and the presence of “Paris” — and stands at 247 responses in the general cut. Merged entities are counted as the union of their response sets, never the sum: Make Up For Ever’s eight child names summed to 75 responses but the union is 21, because a single response frequently mentions more than one of them.

Names that resemble each other usually belong to different companies. Nine variant pairs were folded on evidence — a corporate suffix (Yves Saint Laurent Beauty into Yves Saint Laurent), an accent variant, or a shared verified website (Bade Naturel into Bade Natural, both at badenatural.com). A further fourteen near-name pairs were reviewed and deliberately left separate, because near-identical spelling is not evidence of shared ownership:

Twenty-eight entities were promoted: they were merge targets that never appeared as candidates in their own right, so they had no classification until they inherited their children’s responses. Make Up For Ever, Yves Saint Laurent Beauty and Kiehl’s all arrived that way.

2.5 AI Visibility Score (0–100)

Three components, computed within each cut:

An entity qualifies for ranking if it is named in at least 5% of that cut’s responses.

Worked example — The Purest Solutions, general cut. 373 of 1,125 responses gives a mention rate of 0.3316, which is the highest among qualified entities, so the mention component is 100.0. Its MRR of 0.4957 against the cut maximum of 0.6217 gives a position component of 79.7. Breadth of 5 out of 5 gives 100.0. The composite is 0.45 × 100.0 + 0.30 × 79.7 + 0.25 × 100.0 = 93.92.

The leader scores 93.92 rather than 100 because no single entity tops all three components at once — the highest MRR in this cut belongs to Estée Lauder, not to the leader. That is how component scaling is supposed to behave.

The scale moves with the leader, so scores don’t travel cleanly. A score of 93.92 in the general cut and 96.65 in the core27 cut are not the same quantity, because each cut has its own maxima. Scores must never be compared across cuts or across verticals. Across editions, the main leaderboard shows each brand’s score change since May as a guide to direction; the strict like-for-like comparison is Chapter 8, which uses raw mention rate only. The maxima used for every table in this report are published in the report’s data file.

Components are never clipped at 100. Because the maxima come from qualified entities only, an unqualified entity can exceed 100 on a component — 56 of the 1,924 metric rows do, which is 2.9% of all rows and 3.6% of the unqualified ones. A value above 100 is a true statement that the entity beats the cut leader on that one component while being named too rarely to rank.

2.6 What changed since edition 1, and what is comparable

The overall leaderboards in Chapter 3 rest on a wider base than edition 1’s. Edition 1’s 138-entity list was brand-only and never faced the retailer or marketplace question; this edition’s 322-entity list does, and the question set is two-thirds larger. The main leaderboard’s Δ column shows each brand’s score change since May, and marks entities that were not ranked then as New. Part of any change comes from the wider question set, the larger entity list and the changed panel rather than from the brand, so Δ shows direction, not exact size. The like-for-like comparison is Chapter 8, which uses raw mention rate on the 27 carried questions and the four carried providers.

2.7 Origin, neutrality and self-exclusion

Origin is assigned by trademark ownership, sourced to a public record — a trademark registry entry or the brand’s own corporate disclosure — for every entity. Under this rule a foreign mark operated in Turkey under licence stays foreign and is tagged as a licensed brand. Fourteen entities whose origin could not initially be established have since been confirmed Turkish and are counted as such throughout this report.

Marketplaces and retailers are included in every published leaderboard, because they are genuinely what the assistants name. A brand-only view, with Trendyol and Hepsiburada removed and nothing else changed, is published alongside the main board in section 3.3. Retailers that carry their own labels — Gratis, Watsons, Rossmann, Sephora — are retained everywhere; they are retailers, not marketplaces.

Herm.io’s own domain is excluded from the source data before analysis, as in edition 1. Herm.io does not appear in, measure itself through, or benefit from its own study.

3. Overall Visibility Leaderboard

One-sentence takeaway: Fifty-three entities cleared the threshold on the full question set; the top three are Turkish, the next two are foreign, and the list flattens quickly after the first ten.

The list below ranks all 53 entities that were named in at least 5% of the 1,125 responses, by AI Visibility Score. As a reminder, this ranks visibility, not brand quality.

Average score of the 53 qualified entities: 54.34. Thirty are Turkish, twenty-three are not.

3.1 Tier narrative

The leader, and a closer race than last time (93.92–82.08). The Purest Solutions holds the top of the board with a maximum mention component and full breadth, but its margin has narrowed: Golden Rose is 4.3 points behind and Flormar and Maybelline New York are within 12. In edition 1’s much smaller question set the leader stood 10.7 points clear of second place. Widening the question set brought competitors closer, which is what one would expect if the leader’s dominance was concentrated in the questions that name Turkey explicitly — and section 4.2 shows exactly that.

The dermocosmetic block, foreign and domestic interleaved (76.44–65.39). La Roche-Posay (76.44), Maruderm (74.89), Pastel (74.71), L’Oréal Paris (72.98), Dermoskin (70.06), Cosmed (69.25) and Bioderma (65.39) sit in one band with no clean origin boundary. Two contrasting profiles are visible inside it. La Roche-Posay is named in 17.7% of responses but sits very high when it appears, giving it a position component of 91.4. Maruderm is named far more often (26.5%) but appears late, scoring 46.5 on position. They finish two points apart by completely different routes.

The prominence specialists (65.25–58.43). Estée Lauder is the clearest case in the study: named in only 7.6% of responses, but with the highest MRR of any qualified entity in the cut — when it is named, it is named first. That single component carries it to 12th. Dior (59.56) and Benefit (58.43) follow the same pattern. These are low-frequency, high-prominence entities, and their rank should be read as “rarely mentioned, but mentioned early,” not as broad visibility.

Retail enters the board (57.36–47.06). Gratis at 18th, Trendyol at 19th, Watsons at 22nd, Sephora at 35th and Hepsiburada at 36th. Edition 1’s brand-only list could not show this. It is worth taking literally: when Turkish consumers ask an AI assistant about cosmetics, the answer names where to buy roughly as readily as it names what to buy.

The long tail (46.87–30.09). Scores fall from the high 40s to 30.09 across the last twenty positions. Most entities here are still named by all five models, but both mention and position decline together. At the bottom, S’he Vec (30.09) and three others carry a reduced breadth component because only three or four models named them. A lower rank here means lower current visibility to AI systems — not lower brand quality — and 269 tracked entities sit below the threshold entirely.

3.2 The same board by volume

Scores reward prominence as well as frequency, so the ranking by raw mention rate is a different list. It is worth publishing because the gap between the two is itself informative.

La Roche-Posay is the sharpest divergence in the study: 12th by volume, 5th by score. Farmasi runs the other way, 11th by volume and 16th by score — it is named often but consistently late in the answer. Maruderm and Cosmed both gain positions on volume that the position component takes back.

3.3 Brand-only view

Removing the two marketplaces — Trendyol and Hepsiburada, named explicitly, with nothing else changed — leaves 51 qualified entities and moves every entity below rank 19 up by one or two places. No entity in the top ten is affected.

4. Two Shelves: What Happens When You Say “Turkish”

One-sentence takeaway: The same five models, in the same week, produce a 98.3% Turkish shelf and a 69.2% foreign one depending only on whether the question specifies origin.

This is the finding that carries the edition. The 45 questions split cleanly into 22 that constrain the answer to Turkish brands (“Kaliteli ve yerli cilt bakımı markaları arıyorum”) and 23 that do not (“en iyi makyaj bazı hangi marka”). Each set has its own leaderboard, its own denominator, and its own scale. Neither can be read as the market; both together can.

4.1 Origin share on both cuts

4.2 The Turkish-constrained shelf

Forty entities qualify, thirty-eight of them Turkish.

The Purest Solutions is named in 51.8% of these 550 responses — more than one answer in two — against 33.2% across all 45. Its lead here is 10.5 points, close to the margin it held in edition 1’s smaller, more origin-constrained question set. Maruderm (42.4%), Cosmed (35.6%), Golden Rose (34.0%) and Dermoskin (34.9%) follow. Golden Rose is the position specialist of this cut: it has the highest MRR of any qualified entity here, so when a model names Turkish brands, Golden Rose tends to come first even though four brands are named more often.

Two non-Turkish entities cleared the threshold on questions that explicitly asked for Turkish brands. Bioderma appears in 5.5% of them and Watsons in 6.2%. Neither is an error in classification — Bioderma is French and Watsons’ trademark sits in the British Virgin Islands. It is a genuine, if modest, rate of models answering a constrained question with an unconstrained answer, and it is worth watching across editions.

4.3 The origin-neutral shelf

Fifty-three entities qualify, only seventeen of them Turkish, and the first Turkish entry is sixth.

Maybelline New York is named in 49.9% of these responses — a higher rate than any entity achieves on any other cut in this study. La Roche-Posay and L’Oréal Paris follow. The first Turkish entity is Flormar at sixth, then Golden Rose seventh; Gratis, a Turkish retailer rather than a brand, is tenth; The Purest Solutions — first overall, first on the constrained cut — is fourteenth here, at 15.3%.

The pattern inside the Turkish seventeen is consistent: makeup travels, skincare does not. Flormar, Golden Rose, Pastel, Note Cosmetics and Beaulis all place on the origin-neutral shelf. Turkish skincare, which dominates the constrained cut, largely does not — Maruderm falls to 31st, Cosmed to 38th, Dermoskin to 51st, and Gülsha and Cream Co. sit at the very bottom of the qualifying list. Against the constrained cut, where Turkish skincare occupies four of the top seven positions, that is a striking asymmetry.

Do not read this against edition 1’s 49.1%. May’s origin-neutral figure came from 5 questions and 100 responses; this one from 23 questions and 575. The two are not the same measurement, and the difference between them is a change in what was asked, not a change in what the models believe. The comparison that is valid is in Chapter 8.

5. Differences Between Models

One-sentence takeaway: The five assistants differ by a factor of two and a half in how many brands they name, and one of them disagrees with the other four about which brand belongs at the top.

5.1 Per-model behavior

Grok names 15.2 entities per answer; Claude Haiku names 6.2. That is not a difference in knowledge — Claude’s 200 distinct entities are within a few of Grok’s 203 — but a difference in answer format. Grok produces long enumerated lists; Claude produces short, selective ones. Both behaviors are legitimate; they simply produce very different visibility economics for a brand. A brand that is “somewhere in the list” is visible to Grok’s users and invisible to Claude’s.

Claude Haiku 4.5 is the outlier on substance. It is the only model in the panel whose most-named brand is not Turkish: L’Oréal Paris at 23.1%, followed by Maybelline New York at 21.8%. The Purest Solutions — the overall leader, named in 39.6% of Gemini’s answers and 38.2% of Perplexity’s — appears in only 7.6% of Claude’s. Its top ten splits five-five between Turkish and foreign brands; every other model’s top ten is eight or more Turkish. Section 8.3 returns to this, because Claude is also the model whose tier changed between editions.

Perplexity has the widest repertoire at 207 distinct entities, and it is the model most likely to surface small Turkish names: Misbahçe, S’he Vec and Cecile all appear in its top twenty and in no other model’s.

5.2 Each model’s most-named entities

5.3 The marketplace effect is one model

Trendyol qualifies 19th overall and Hepsiburada 36th. Both positions rest almost entirely on Grok.

Grok names Trendyol in more of its answers than it names any actual cosmetics brand. Read the marketplace ranks accordingly: they describe one assistant’s habit of pointing at where to buy, not a market-wide consensus. In the Turkish-constrained cut both marketplaces qualify on two models only — the lowest breadth of any qualifier in this study.

5.4 What only one model sees

Each model names entities no other model names once. The lists are a useful reminder of how much of the long tail is model-specific rather than market-wide.

6. Question-Type Ownership

One-sentence takeaway: Broad discovery questions spread visibility across 63 entities; specific use-case questions concentrate it into 33, and the winners change with the format.

Discovery — “which brands should I know about?” — produces the widest field: 63 qualifiers, the most of any cut. Golden Rose leads at 50.9%, with Flormar close behind at 49.3%. Both are makeup houses, and the top five is makeup-heavy. When a model is asked for names without further constraint, it reaches for the brands with the broadest name recognition.

Attribute — “vegan,” “cruelty-free,” “affordable but good” — narrows to 49 qualifiers and is the most Turkish cut in the study proportionally, with 34 of 49 qualifiers Turkish. Golden Rose leads again, but the interesting entry is Bade Natural at 7th: named in only 16.3% of these responses, yet with a position component high enough to place it above L’Oréal Paris and La Roche-Posay. It is a specialist brand that models reach for early on natural-content questions specifically.

Use case — “oily and acne-prone skin,” “a moisturizer for winter dryness” — is the narrowest and the most decisive: 33 qualifiers, the highest average score (54.05) of the three, and a top ten that is almost entirely skincare. The Purest Solutions, La Roche-Posay, Maruderm, Dermoskin and Bioderma take the first five. Estée Lauder, with the highest MRR in the study at 0.81, sits 7th on an 8.8% mention rate.

The practical reading: a brand’s visibility is not one number. A makeup brand can be highly visible on discovery questions and absent from use-case ones; a dermocosmetic brand can be the reverse. The three cuts here have identical denominators, so those differences are real and not artifacts of sample size.

7. The Product Shelf

One-sentence takeaway: Skincare and makeup have real, populated leaderboards; haircare has two qualifying entities in the entire study.

Product category is an attribute of each entity, assigned during entity resolution — not a question cut. The rows below therefore do not sum to the general cut, and an entity spanning several categories is counted once, as “multi-category.”

Skincare is where Turkish brands are strongest. 86 tracked entities, 17 qualified, and 12 of those 17 Turkish — The Purest Solutions, Maruderm, Dermoskin, Cosmed, Rosece, Bade Natural, Gülsha and others. This is the deepest domestic field in the study.

Makeup is where they are most exposed. 56 tracked entities, 13 qualified, and only 5 Turkish. Golden Rose leads it, with Flormar and Pastel behind, but Maybelline New York, Benefit, MAC Cosmetics, Essence, NYX and NARS all qualify. Set against section 4.3 — where Turkish makeup is the part of the domestic field that survives origin-neutral questioning — the picture is of a small number of strong Turkish makeup houses competing in a crowded international category, and a much larger Turkish skincare field that is visible mainly when the question invites it.

Haircare is nearly empty. 21 tracked entities and two qualifiers, Bioxcin and Urban Care, both Turkish. There is no third haircare brand at 5% anywhere in this study. Some of that is question design — no question asks about haircare in general, only about oily hair — but the scale of the gap is worth naming.

Fragrance, suncare and shaving have three, five and one tracked entities respectively and no qualifiers at all. Suncare in particular is under-represented as a category despite three questions about sunscreen, because models answer those questions with dermocosmetic brands that are classified as skincare.

8. What Changed Since May

One-sentence takeaway: On the 27 questions repeated verbatim, 43 brands rose and 60 fell, and the movement runs against the largest incumbents and toward mid-sized Turkish challengers.

8.1 How this comparison is built, and why it isn’t quite like-for-like

The Δ column on the main leaderboard compares overall scores, which also move with the changes in questions and panel. This chapter uses the stricter measure, raw mention rate only, on the core27_4prov cut: the 27 questions carried verbatim from edition 1, answered by the four providers that were also in edition 1, five times each — 540 responses, structurally identical to May’s 540.

Two things still differ, and both must be stated with any number in this chapter. Claude moved from Sonnet 4.6 to Haiku 4.5, a tier change rather than a version bump, and as section 5.1 showed, Haiku behaves very differently — it is the only model in the panel that leads with a foreign brand. Gemini moved from 3.5-flash to 3.6-flash and its grounding rate collapsed from 84.4% to 26.2%. Neither change has anything to do with the brands being measured.

We publish the comparison anyway, for a reason worth stating plainly: model churn is permanent. Providers ship new versions and new tiers continuously, and a brand’s objective is not to be visible to one particular model but to be visible whichever model happens to answer. A comparison that waits for a stable panel would never be published. So the direction of travel here is meaningful; the precise magnitude for any single brand is not.

One reassuring check. Adding Perplexity to the September figures — the same 27 questions on all five models rather than four — moves the mention rate of the average matched brand by 0.68 points, with a median of 0.3, and nineteen of the top twenty brands are the same either way. The panel expansion is far less destabilizing than the panel substitution. The table below carries both September columns so this can be checked brand by brand.

8.2 The movers

Maruderm is the story of the quarter. It moved from 23.89% to 40.00% of the carried responses, a 16.11-point gain, and it is now the second most-named brand on those questions where in May it was ninth. Bioxcin roughly doubled, from 11.85% to 24.26%. Cosmed, Cream Co., Urban Care and Note Cosmetics all gained meaningfully. These are mid-sized Turkish brands, none of them a household leader in May.

Farmasi lost the most, from 33.52% to 23.15%, and Atelier Rebul lost almost as much, from 20.37% to 10.19%. Both were top-ten brands in edition 1. Section 8.3 discusses how much of Farmasi’s fall can be attributed to the panel.

The leader held. The Purest Solutions slipped 3.88 points to 50.56% and remains first on the carried questions by more than 10 points. Pastel, Flormar and Golden Rose all drifted down by one to four points without changing position materially.

Turkish brands did better than foreign ones on these questions, which is notable given Chapter 4’s headline. Of 77 matched Turkish brands, 35 rose; of 30 matched foreign brands, only 8 did. The carried 27 questions are heavily Turkish-constrained — 22 of them — so this is the domestic shelf getting more crowded, not the foreign one retreating.

8.3 One movement we can partly explain

Farmasi’s 10.37-point fall deserves a caveat rather than a narrative. In edition 1, Claude was the single model that ranked Farmasi first. In this edition Claude is a different tier of model, and Farmasi appears in 8.9% of its answers, well below Grok’s 40.4%. A meaningful share of the fall is therefore attributable to the model that changed, not to anything Farmasi did or failed to do. We can say the direction is down; we cannot responsibly quantify how much of it is the brand.

The same caution applies in reverse to any brand that Claude Haiku happens to favor. Readers looking for a clean brand-level trend should wait for edition 3, where the panel is expected to be stable enough for the raw layer to carry more weight.

8.4 What dropped out

Twenty-six of edition 1’s 138 brands are no longer tracked: sixteen produced zero matches anywhere in the September corpus, and ten fell below the three-response review floor. None of them had qualified in May — the largest, Laneige, was named in nine of 540 responses. Their disappearance changes no published figure. It is reported because a brand vanishing from an entity list is a finding, not a housekeeping detail: eleven of the sixteen zero-match brands are small Turkish natural-cosmetics labels.

9. The Discovery Ecosystem: Where AI Learns About Brands

One-sentence takeaway: No single source shapes what these assistants know about Turkish cosmetics; 770 domains contribute, more than half of all citations come from small sites, and brands’ own websites now account for nearly a quarter.

When an AI assistant answers with web search, it draws on whatever content is available to it at that moment. Across the grounded responses in this study, the cited addresses were reduced to domains and ranked by how many distinct responses cited them — a map of where these systems are reading about cosmetic brands. Note: citation payloads differ by provider (see 2.2), so this analysis is more complete for some models than others, and Gemini contributes far less than it did in May because it searched far less.

9.1 The nine most-cited domains

Nine domains reached 100 responses. The top of the list mixes three kinds of site: a general-interest content platform (onedio.com, 201), a marketplace (trendyol.com, 193), and beauty-specific editorial (oggusto.com, wonderflaw.com, vogue.com.tr). Pharmacy and retail e-commerce — watsons.com.tr, hopi.com.tr, dermoeczanem.com — occupies the middle. That trio of source types is the character of the ecosystem: editorial lists, retail listings, and retailer loyalty content.

9.2 Source-type mix

Brand-owned sites account for 22.6% of citations — 1,807 of 8,009, spread across 154 distinct brand domains. sephora.com.tr, lorealparis.com.tr, bioderma.com.tr, maybelline.com.tr and laroche-posay.com.tr lead, but thepurestsolutions.com (63), maruderm.com (60), flormar.com.tr (49), cosmed.com.tr (45) and cyrene.com.tr (39) are all in the same band. A brand’s own site is not a marginal source in this category.

9.3 The fragmentation finding

52.7% of citations come from the long tail. Of the 770 domains, 577 were cited by fewer than ten responses and 203 by exactly one. Only nine cleared 100. No single authority determines what these systems believe about Turkish cosmetics; hundreds of small beauty sites, blogs, listicles and regional news pages add up to more than half the picture.

What this means for understanding brand discovery. The entities that rank highest tend to be present across several parts of this ecosystem at once rather than depending on one: named in Turkish beauty editorial, listed on the major retail and pharmacy platforms, and carrying a current, structured website of their own. This is a description of an observed pattern, not advice on how to rank, and this report does not test whether presence in these sources causes visibility or merely accompanies it.

10. What the Patterns Suggest

One-sentence takeaway: Five neutral readings of what the data shows, offered as observations rather than recommendations.

Visibility is question-shaped, not brand-shaped. The same brand can hold 51.8% on one cut and 15.3% on another in the same week. Any single-number description of “a brand’s AI visibility” is hiding which question was asked. This is the reason every figure in this report is published with its cut and its denominator.

Category origin matters more than brand strength on the open shelf. Turkish makeup houses survive origin-neutral questioning; Turkish skincare largely does not, despite being the deeper and more numerous domestic field. Whatever is producing that asymmetry — international category conventions, the composition of the sources, the way sunscreen and serum questions are phrased — it operates at the level of the category, not the individual brand.

Format differences between models are as consequential as knowledge differences. Claude and Grok know almost exactly the same number of entities. One names six per answer and the other fifteen. For a brand at the edge of visibility, that difference decides whether it is seen at all, and it is a property of the assistant’s answer style rather than of the brand.

Retail intermediaries occupy real space in the answer. Five retailers and marketplaces qualify on the full question set, and one model names a marketplace more often than any cosmetics brand. When consumers ask assistants about cosmetics, a meaningful part of what comes back is where to buy rather than what to buy.

Movement over a quarter runs toward the middle of the table. The largest gains on the carried questions went to mid-sized Turkish brands, and the largest falls to established ones. Whether that is a durable pattern or one quarter’s noise cannot be determined from two data points, particularly with a panel that changed. Edition 3 is the first opportunity to know.

11. Limitations and Notes

One-sentence takeaway: This is a single, one-market, point-in-time snapshot on a panel that changed between editions; the figures should be read within the limitations below.

• The model panel changed, and one change is a tier change. Edition 1 ran four models; this edition runs five. Perplexity is new, Gemini moved 3.5 to 3.6-flash, ChatGPT 5.5 to 5.6-luna-pro, and Claude moved from Sonnet 4.6 to Haiku 4.5. Only Grok is constant. This is why the leaderboard’s score changes since May are presented as direction rather than exact size, and why even the raw-layer comparison in Chapter 8 is reported with the confound named beside it.
• Gemini’s grounding collapsed. From 84.4% to 26.2%. All 225 calls completed with substantive text, so this is model behavior, but roughly an eighth of this study’s corpus reflects memory rather than retrieval, concentrated almost entirely in one model.
• No search-versus-memory chapter this edition. Because the ungrounded segment is essentially one model’s, splitting the leaderboard by grounding would produce a model comparison dressed as a behavioral finding. Edition 1 published such a split on 58 responses; that finding should not be carried forward.
• Single, point-in-time measurement. Data were collected in one run on 4–5 September 2026. Five repeats per question-model cell capture within-model variability, not day-to-day drift.
• Small score gaps are not meaningful. In the packed middle of the leaderboard, differences of a point or two should not be over-interpreted.
• Seven qualifying positions rest on two models. Trendyol and Hepsiburada in the Turkish-constrained and use-case cuts, Bionnex in the attribute cut, Luv it! in the use-case cut, and S’he Vec in the trend cut each qualified on only two of five models. Their visibility is model-specific, not market-wide.
• A high-prominence, low-frequency entity can rank surprisingly high. Estée Lauder ranks 12th overall on a 7.6% mention rate because it holds the highest MRR in the cut. That is the score working as designed, but it is not the same kind of visibility as a brand named in a third of all answers.
• No segment cuts this edition. 27 of the 45 questions carry no product subcategory, and the 18 that do average one question each — far below the floor at which a segment leaderboard means anything. The product-category chapter is an entity attribute, not a question cut, and does not reconcile to the general cut.
• Visibility is not quality. This report measures mention and position only. It does not assess whether a brand was described positively, whether a recommendation was accurate, or whether a product works. A score is not an endorsement.
• Models are probabilistic. The same question can produce different answers; five 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 in matching. Entity matching was deliberately cautious: alias-based, word-boundary anchored, Turkish-aware in its case folding, and run only on the answer text with link targets stripped. Brands mentioned indirectly, or by a product line that could not be safely attributed, may be undercounted.
• Line-level and sub-brand naming splits some entities. Where a model names a range rather than its owner, the range was merged into the parent when evidence supported it. Some legitimate sub-brand and local-variant naming remains separate, and a small number of entities are consequently counted more conservatively than a marketing team would.
• Origin classification involves judgment. Origin follows trademark ownership and is sourced for every entity, but foreign marks operated in Turkey under licence, and Turkish-founded brands later acquired abroad, are genuinely arguable cases.
• Fourteen origins were corrected after the first build. Entities whose trademark ownership could not initially be established have since been confirmed Turkish and are counted as Turkish throughout. This affects origin-share figures slightly and no leaderboard position at all.
• Source coverage varies by model. Providers expose citations differently, and one model grounded only a quarter of its answers, so the source analysis does not represent all answers equally.
• Model versions date quickly. These findings are specific to the model versions listed in 2.1, as of early September 2026.

12. Appendix & Data

12.1 All 45 questions

12.2 All tracked entities (322)

The table lists all 322 entities tracked in this study and marks the 53 that cleared the 5% mention threshold on the full question set. 176 are Turkish. The list is ordered by mention count, descending. Brand pages are not linked in this edition.

12.3 Data availability

To support scrutiny and reproduction, the data behind this report is available on request: the full per-entity metrics across all eight cuts, the qualified leaderboards and their brand-only variants, the entity dictionary with origin evidence, the complete entity-decision audit trail including every dropped and merged candidate with its reason, the per-model breakdowns, the origin-neutral market table, the full 770-domain source list, the raw-layer trend file, the published scale registry, and the methodology and data dictionary.

All figures reflect a single collection window (4–5 September 2026) and are subject to the limitations in Chapter 11. This is the second edition; the study is repeated quarterly, and edition 3 will be the first with a stable enough panel for brand-level trend to carry real weight.

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 →