AI Search for Barbershops in IstanbulAsk in English and Turkish — get two different cities

This is the ninth city study in the series and the first to leave both the food-and-retail world and the Latin-only web: barbershops, measured in a city where the craft is a national institution. The design question was specific. Across eight prior studies, ChatGPT’s willingness to cite a business’s own website looked like a service-vs-food split — yoga, bikes and tattoo held 32–42%, while bookstores, coffee and bistros collapsed to 0.9–10%. The Seoul study argued the real driver was how much own-website layer a vertical even has. Istanbul barbers are the discriminating case: a service craft whose shops mostly live on Instagram and booking apps (22.1% of the registry has an own-domain site).

The verdict landed between the two stories — ChatGPT cites barbershop websites 21.3% of the time, under every prior service vertical and well above the food floor, almost exactly matching the density of the web layer itself. But the run’s loudest result was one we didn’t design for: asked the same control question in two languages, the engines returned two disjoint sets of barbershops. Not a reshuffle. Zero common shops in the top five.

• 0% EN↔TR control overlap — series first; the previous floor was Marseille’s 11%. Six of eleven measurable templates sit at exactly 0%.
• ChatGPT searched on only 60% of captures (55/92) — its lowest trigger rate in the series (Seoul 95%, Berlin tattoo 100%) — and its citation volume shrank to 160 rows.
• The booking layer is back: 18.9% of Perplexity’s citations — and its reach is not one engine’s quirk: armut.com, fresha.com, kolayrandevu.com and onliner.tr each appear in every fired engine’s citation pool.
• Rimedzo Barber Shop is a clean dual-metric #1 — 140 answer mentions across all four engines and the top citation score (93).

Every study in this series asks its control question twice — once in English, once in the local language, from a local IP — and compares the top five businesses each version surfaces. Until now the two lists always intersected somewhere: 25% in Paris, Berlin, Tokyo and Seoul; 11% in Marseille; 67% in the Berlin tattoo control. Istanbul is the first city where the intersection is empty. “Best barbershops in Istanbul” and “İstanbul’daki en iyi berberler” produce ten distinct shops with no member in common.

The pattern holds below the control, too. Of the eleven templates with resolvable answers in both languages, six sit at exactly 0% overlap and none exceeds 25%:

Here is the metric this study was built to test. ChatGPT’s citations went to barbershops’ own websites 34 times out of 160 — 21.3%. Two reference points make that number interesting: every previous service vertical held 32–42%, and every food/retail vertical fell between 0.9% and 10%. Istanbul’s barbers split the difference.

Why? Count the websites. Only 81 of the 367 real shops in the registry — 22.1% — have an own-domain site at all (Berlin tattoo: 73%; Paris bistros: 67%; Marseille: 41%; Seoul: 13%). A fifth of the shops own a web presence, and roughly a fifth of ChatGPT’s citations land on shop websites. The service-vs-food framing that survived eight studies turns out to be a proxy: what ChatGPT’s citation mix actually reflects is how much own-web layer the vertical maintains. Paris bistros remain the residual puzzle — plenty of websites, near-zero citations — so density is the strongest single predictor we have, not yet a law.

ChatGPT ran a live web search on 55 of its 92 captures — 60%. In Seoul it searched on 95%; in the Berlin tattoo study, on every single capture. The other 37 Istanbul answers came straight from the model’s index/memory, and they are not short refusals: they average about 3,500 characters of confident, named recommendations.

The skipping is uniform — roughly 40% in each language × proxy cell (9, 8, 10 and 10 of 23) — so it isn’t a Turkish-language behavior. Whatever gate decides “do I need the live web for this?” concluded, four times in ten, that Istanbul barbershops were already covered. Our hotel-side result_source teardown measured 37.3% of hotel questions skipping search; this is the first time the Fun series catches the same reflex at comparable scale.

The knock-on effect matters for measurement: with search off, there are no citations, so ChatGPT’s citation-based numbers this run stand on 160 rows. A shop can be recommended dozens of times without a single citable trace — which is precisely why this series counts answer mentions separately from citations.

Berlin yoga made booking platforms look structural: Urban Sports Club, Eversports and ClassPass carried 31% of Perplexity’s citations there. Then the layer vanished — 4% for Berlin tattoo, ~3% for Paris bistros — and the series concluded it only exists where a class-pass economy does. Istanbul confirms that reading from the other side: barbering has a real appointment-app economy, and the layer snaps back to 18.9% of Perplexity’s 1,160 citations.

The composition is local-first: armut.com (118 citations study-wide), a Turkish everything-services marketplace, ahead of global fresha.com (79), Turkish appointment tools kolayrandevu.com (65) and onliner.tr (43). All four of those booking domains show up in every fired engine’s citation pool — the layer isn’t a Perplexity quirk, Perplexity just leans on it hardest.

Per-engine source mix, as a share of each engine’s raw citation rows (the same basis as every earlier study in the series):

The ranking below counts answers that recommend the shop — extracted from the prose of all 368 captures and resolved against the registry — because citation counting is structurally blind in a market where most shops have no domain to match. The citation score is shown alongside for contrast.

Turkish-language prompts put 8% of their citations on .tr domains; English prompts 4.2% — a coupling ratio of 1.9×, measured across all engines and all 4,211 citation rows. That lands mid-series: Tokyo’s Japanese prompts hit .jp domains at 5.0×, Berlin yoga measured 1.5× for .de, Berlin tattoo was neutral (0.87×), and Seoul couldn’t produce the ratio at all because Korean web identity lives on platform subdomains.

Istanbul has a structural ceiling of its own: among the 81 shop domains in the registry, .com outnumbers .com.tr 46 to 11. Turkish barbers who bother with a website mostly buy the global TLD, so even a perfectly local-leaning engine has limited .tr inventory to cite. The ChatGPT-entity-only variant of this metric (used when the series began) is below any reasonable threshold here — 20 rows — and is disclosed rather than charted.

The registry spans both sides of the Bosphorus — Beyoğlu (117 shops), Şişli, Fatih, Kadıköy, Üsküdar, Beşiktaş, plus Sarıyer and Bakırköy from the district passes. The twelve consensus shops cluster tightly on the tourist-and-transit spine: Sirkeci, Taksim/Beyoğlu, Beşiktaş, with Gent’s Istanbul (Kadıköy) and Cut Labs’ Caddebostan branch holding the Asian side.

• The full grid landed. All 8 platform×proxy batches returned complete (368/368) — a series first for the engines that were fired, including AI Mode over a Turkish proxy (the FR-proxy rejection that hit the Marseille run has no Turkish equivalent).
• Refusals barely exist in Turkish. 9 of 368 captures recommend nothing (AI Mode × TR 4, Perplexity × TR 3, ChatGPT 2). Seoul’s pattern — Perplexity refusing to name places in 28% of its Korean captures — has no Istanbul counterpart.
• Unresolved-mention rates by engine (share of extracted names not matching the 374-row registry): ChatGPT 21.8%, AI Mode 18.1%, Copilot 15.2%, Perplexity 11.8%. Most of the remainder are documented ambiguities — generic titles (“Barbershop (Cihangir)”), two-branch brands (Dest Hair), and a kids-salon chain the recovery pass couldn’t disambiguate.
• AI Mode ships ads in Turkish answers. Several TR captures embed shopping results (GetYourGuide tours, Harry’s razors, a keratin treatment) inside the recommendation prose. The extraction pass excludes them by rule, but it is the first time this series has seen ad units inside a local-recommendation answer.
• Google’s Turkish grooming taxonomy is blurry. Real barbershops carry “Barber shop”, “Hairdresser” and “Hair salon” labels interchangeably (“erkek kuaförü” = men’s barber). The registry keeps all three as real, mirroring the bistros study’s handling of “French restaurant” — dropping them would have inflated the unresolved rate.

1. Your armut/Fresha/KolayRandevu profile is your citable page. All four engines cite the booking layer; ChatGPT will not cite your Instagram. If the choice is between polishing the Instagram grid and completing the booking-platform listing, the listing is what AI search can actually read.
2. A cheap own-domain site still pays. The shops that own both ranking methods — Rimedzo, 2.a.berber, Gent’s — all have real websites the engines cite directly. In a market where only 22.1% of shops have one, a basic domain is a differentiator, and .com works fine (the registry’s own split is 46 .com to 11 .com.tr).
3. Decide which Istanbul you serve — or list for both. The English and Turkish recommendation sets share nothing. Tourist-facing shops win English answers with English-legible names and hot-towel content; neighborhood shops win Turkish answers through review mass on Google. Bilingual presence (site + listing text in both languages) is the only route into both lists this data suggests.
4. Reddit remains the one social lever ChatGPT reads — a fifth of its citations here, as in every city measured. Threads asking “barber in Istanbul?” are retrievable assets in a way Instagram posts are not.

We came to Istanbul to break a tie between two explanations of ChatGPT’s own-website baseline, and the tie broke cleanly: the citation share follows the web layer a vertical actually maintains — 22.1% density, 21.3% share — with “being a service” worth at most a premium on top, and Paris bistros still filed as the unexplained outlier.

The result we didn’t come for is the bigger one. Nine studies in, the engine personalities have survived every vertical, language and continent we’ve thrown at them; what fractured instead is the city itself. One prompt, translated, retrieves two non-intersecting Istanbuls — a tourist city assembled from English tour content and fade-shop SEO, and a Turkish city assembled from review mass and appointment apps. AI search didn’t blend them. It picked one per language and never looked across.

Next open questions for the series: does a second thin-web service vertical (nail salons? laundromats?) reproduce the density line, and does the language fork widen further in cities whose local web is even less English-adjacent?