AI Search for Saunas in Helsinki (2026):the first city where every AI engine agrees
TL;DR: We asked five AI engines 24 sauna questions in English and Finnish, through US and Finnish proxies — 475 of 480 captures landed. For the first time in ten studies, the two languages returned the same top-5 on the headline question: 100% overlap, where the last two cities measured 0%. Löyly is named in 379 of 475 answers — the strongest single-venue consensus the series has recorded — and ChatGPT cites sauna-owned websites 45.5% of the time, a series high that lands exactly where the registry’s 78.4% own-domain density predicts. Meanwhile ChatGPT’s Reddit habit, 17–20% everywhere since Seoul, drops to 1 citation in 242.
Executive Summary
Ten cities in, every market this series visited split somewhere — English against the local language, engine against engine, mention counts against citation counts. Helsinki is the city where the splits close.
We picked saunas as the missing end of a dose-response curve. Seoul reframed ChatGPT’s own-website citation share as a function of how much ownable web a vertical has, and Istanbul confirmed the thin end: a 22.1%-density registry produced a 21.3% share. Nobody had tested the thick end, and almost nothing on Google Maps is as thoroughly domained as Finnish saunas: of 125 real venues in our registry, 116 (92.8%) carry a website and 98 (78.4%) an own-domain site. The prediction held. ChatGPT cited sauna-owned sites 45.5% of the time (110 of 242) — above the previous high (Amsterdam bikes, 42%) and clear of the old 32–35% “service baseline”.
The result we did not predict is the convergence. “Best public saunas in Helsinki” and “Helsingin parhaat yleiset saunat” return the identical five venues — Löyly, Kotiharjun Sauna, Allas Pool, Kulttuurisauna, Sompasauna — a 100% top-5 overlap on a metric whose last two readings (Istanbul, Mexico City) were literally zero. Three of 24 templates hit 100%, the study median is 43%, and the language→TLD coupling that gave Tokyo its 5× headline reads 0.92× here — no coupling at all. One canon, machine-readable in both languages.
The grid: 475 of 480 captures, 3,493 cited URLs, 3,017 extracted venue mentions at 94.5% resolution — the series’ best. Perplexity is 5 captures short after a targeted refire (disclosed in the methodology, nothing imputed), and exactly 1 answer in the whole study recommended nothing.
Terminology, as on every page in this series: citations are the URLs an engine attaches under its answer, mentions are the venue names inside the answer text. Source analysis uses the former, the leaderboard the latter.
Ask in Finnish, get the same answer
This metric has been the series’ sharpest divider. Istanbul’s English and Turkish barber lists shared nothing; Mexico City’s English and Spanish taco lists shared nothing. Helsinki’s English and Finnish sauna lists are the same list.
| Prompt template | EN vs FI top-5 overlap |
|---|---|
| control (best public saunas) | 100% (5/5) |
| ice swimming (avanto) | 100% (5/5) |
| sea-swim saunas | 100% (5/5) |
| district: central Helsinki | 67% (4/6) |
| district: Kallio | 67% (4/6) |
| most famous saunas | 67% (4/6) |
| tourist persona | 67% (4/6) |
| winter saunas | 67% (4/6) |
| affordable saunas | 67% (4/6) |
| spas with saunas (bleed) | 43% (3/7) |
| wood-heated saunas | 43% (3/7) |
| open late | 43% (3/7) |
| trendiest modern saunas | 43% (3/7) |
| quiet, non-touristy saunas | 43% (3/7) |
| most traditional saunas | 43% (3/7) |
| island saunas | 29% (2/7) |
| bookable online | 25% (2/8) |
| saunas by the sea | 25% (2/8) |
| family-friendly | 25% (2/8) |
| first-timer persona | 25% (2/8) |
| smoke sauna | 25% (2/8) |
| most upscale experience | 25% (2/8) |
| hotel saunas (bleed) | 11% (1/9) |
| private group rental | 11% (1/9) |
Overlap = shared venues ÷ distinct venues across both top-5 lists (Jaccard). Top-5 taken in answer order from the NER mention rows — ChatGPT’s map carousel, which earlier studies read this metric from, appeared in 1 of 96 captures this run (section 9).
Where the agreement comes from is visible in the table’s shape. The prompts that touch the famous venues converge hard: the control, ice swimming and sea-swim saunas at 100%, the famous-saunas, Kallio, winter and tourist prompts at 67%. The prompts that step outside them diverge: private group rentals and hotel saunas sit at 11%, because rental saunas and hotel spas are exactly the corners of this market the shared list doesn’t reach. Language stops mattering where the subject is famous enough; it still matters at the edges.
The TLD measurement agrees. In Tokyo, Japanese prompts pulled .jp entity domains five times harder than English ones; Istanbul’s Turkish prompts ran 1.9×. Helsinki: 0.92× — English prompts cite .fi sauna domains essentially as often as Finnish prompts do (49 vs 45 citations across 14 domains). There is no separate tourist web to couple to. Finnish sauna sites publish in English, the city’s guide publishes in English, and both languages’ answers draw from the same pool.
One familiar split does survive, shrunken: FI-language answers lean a little more local (Kotiharju first, Sompasauna higher), EN answers put Löyly first. But that is a reordering of a shared list — in Istanbul it was two different cities wearing the same name.
45.5%: the density law passes its high-end test
Whether ChatGPT cites a business’s own website turns out to have little to do with what the business sells. It tracks how many businesses in the market own a domain — and Helsinki finally supplies the data point at the top of the curve.
The arc across the four cities where we measured registry density directly now reads: Mexico City 12.5% density → 0% share. Seoul 13% → 2.3%. Istanbul 22.1% → 21.3%. Helsinki 78.4% → 45.5%. Share rises with supply through its whole range, and Helsinki adds the observation the hypothesis was missing: it keeps rising well past the old 32–42% service band, which retroactively looks like a density plateau rather than an engine ceiling. What Helsinki also shows is that the relationship bends — at 78.4% density the share is 45.5%, not 78%, because even here ChatGPT spends most of its remaining citations on the editorial layer (section 5).
The rest of the engine ladder holds its shape at altitude. Copilot, the most website-hungry engine in every study, reaches 71% (368 of 518); Gemini 28.7%; Perplexity 17%; AI Mode 9% plus its 52.4% of self-citations. Every engine cites more owned web here than in any food or retail city — supply lifts all boats, each to a different waterline.
Löyly, in four answers out of five
The waterfront sauna in Hernesaari is named in 379 of 475 captured answers — 79.8% of everything five engines said about Helsinki saunas, in either language, through either proxy. No venue in nine prior studies came near that: Mexico City’s El Vilsito, the previous consensus record, appeared in 34.8% of its city’s answers. ChatGPT names Löyly in 91 of its 96 captures.
Per-engine mention matrix
| # | Sauna | AI Mode96 captures | ChatGPT96 captures | Perplexity91 captures | Gemini96 captures | Copilot96 captures |
|---|---|---|---|---|---|---|
| 1 | Löyly Helsinki | 72.9%(70) | 94.8%(91) | 72.5%(66) | 70.8%(68) | 87.5%(84) |
| 2 | Kotiharjun Sauna | 49%(47) | 81.2%(78) | 61.5%(56) | 53.1%(51) | 66.7%(64) |
| 3 | Sompasauna | 52.1%(50) | 80.2%(77) | 28.6%(26) | 27.1%(26) | 63.5%(61) |
| 4 | Allas Pool | 59.4%(57) | 62.5%(60) | 33%(30) | 61.5%(59) | 19.8%(19) |
| 5 | Kulttuurisauna | 32.3%(31) | 76%(73) | 17.6%(16) | 36.5%(35) | 65.6%(63) |
| 6 | Lonna | 30.2%(29) | 61.5%(59) | 13.2%(12) | 35.4%(34) | 42.7%(41) |
| 7 | Uusi Sauna | 22.9%(22) | 61.5%(59) | 5.5%(5) | 14.6%(14) | 42.7%(41) |
| 8 | Sauna Hermanni | 10.4%(10) | 54.2%(52) | 11%(10) | 18.8%(18) | 40.6%(39) |
| 9 | Saunasaari | 10.4%(10) | 31.2%(30) | 7.7%(7) | 6.2%(6) | 13.5%(13) |
| 10 | Hyvän tuulen sauna | 1%(1) | 40.6%(39) | 12.1%(11) | 0(0) | 12.5%(12) |
| 11 | Furuvik Seaside Sauna | 2.1%(2) | 35.4%(34) | 5.5%(5) | 1%(1) | 20.8%(20) |
| 12 | Helsingin Saunalautta | 1%(1) | 35.4%(34) | 2.2%(2) | 0(0) | 14.6%(14) |
| # | Sauna | Answer mentions | Citation score | Engines (of 5) |
|---|---|---|---|---|
| 1 | Löyly Helsinki | 379 | 162 | 5 |
| 2 | Kotiharjun Sauna | 296 | 123 | 5 |
| 3 | Sompasauna | 240 | 53 | 5 |
| 4 | Allas Pool | 225 | 43 | 5 |
| 5 | Kulttuurisauna | 218 | 46 | 5 |
| 6 | Lonna | 175 | 49 | 5 |
| 7 | Uusi Sauna | 141 | 39 | 5 |
| 8 | Sauna Hermanni | 129 | 37 | 5 |
| 9 | Saunasaari | 66 | 24 | 5 |
| 10 | Hyvän tuulen sauna | 63 | 9 | 4 |
| 11 | Furuvik Seaside Sauna | 62 | 0 | 5 |
| 12 | Helsingin Saunalautta | 51 | 12 | 4 |
For once, the two metrics this series keeps side by side tell one story. Löyly tops the mention count (379) and the citation score (162, via loylyhelsinki.fi — at 161 raw citations, the biggest single-venue domain in any study so far). Kotiharjun Sauna, the wood-heated Kallio institution, doubles at #2 on both (296 mentions, score 123). That makes Helsinki the third city where mentions and citations crown the same winner, and the first where no artifact row — no Instagram-keyed brand, no keyword-stuffed listing — had to be excluded anywhere near the top of the board. When venues own real domains, citation counting finally measures what mention counting measures.
Below the top two, the consensus list is broad and stable: Sompasauna, the volunteer-run, always-open harbour sauna (240 mentions), Allas Pool by the market square (225), Kulttuurisauna (218), the island saunas Lonna and Saunasaari, and the sauna rafts. Ten of the twelve are named by all five engines. The interesting absentee class: the 203 private apartment-saunas our seed crawl surfaced produce essentially no recommendations — AI answers describe the public sauna city, not the private sauna city Airbnb sees.
What five engines read about Helsinki saunas
All 3,493 cited URLs, bucketed per engine. Two features stand out against the nine earlier mixes: the owned-web column is tall on every engine, and the social column has almost nothing in it.
Integer percentages via largest-remainder rounding — each column sums to exactly 100. Raw counts per engine: AI Mode 1,444 citations, Perplexity 868, Copilot 518, Gemini 421, ChatGPT 242. International press and review aggregators are folded into “Other”.
Two buckets, 86% of the diet
45.5% sauna websites + 41% official tourism and city editorial leaves almost no room for anything else: 1 Reddit citation, zero Instagram, zero travel-blog rows. The narrowest, most institutional source diet ChatGPT has shown in this series.
The blog readers
The personal-blog layer the other engines skip — Helsinki expat guides, sauna travelogues, walking-tour blogs — takes 14% of Gemini’s citations and 13% of Perplexity’s. Perplexity also keeps the study’s only real OTA habit (18.5%, mostly localized hotels.com pages pulled in by the hotel-sauna prompt).
Finland’s booking platforms stay small
saunat.fi, venuu.fi and saunalista.fi — the platforms many rental saunas list as their only web presence — reach 2–7% per engine. Nothing like Berlin yoga’s 31% booking-platform surge; when venues own domains, the marketplaces stay middlemen instead of becoming the answer.
The city outranks the crowd
In every previous market, the most-cited non-Google domain was social or editorial — Reddit three times, Instagram twice, a Michelin guide once. Helsinki’s is the city government’s own tourism site.
Helsinki’s tourism board maintains what is effectively the canonical sauna database of the internet — listicles by neighbourhood, by sea access, by price — in English and Finnish at the same URLs. The engines treat it as ground truth: myhelsinki.fi is cited by all five, in both languages, on 246 rows. A city that publishes its own canon, it turns out, gets to write the AI answers about itself.
The mirror image is what shrank. Reddit — the single most-cited external domain in Amsterdam, a fixture of ChatGPT’s diet at 17–20% in three straight studies — totals 47 citations here (1.3% of the study), and ChatGPT’s share of that is one citation. Instagram totals 42, of which 41 belong to AI Mode; ChatGPT’s Instagram count is zero for the seventh consecutive city. The forum layer exists to fill gaps, and this market doesn’t have the gap.
Twelve domains, half of them Finnish
The cross-engine table, domains normalized. 9 domains reach all five engines, and five of the top twelve belong to individual saunas — a first for this table, which in Seoul was Instagram, Naver and tourism portals top to bottom.
| Domain | Engines | Cites | What it is |
|---|---|---|---|
| myhelsinki.fi | 5/5 | 246 | Helsinki’s official city guide — the study’s most-cited non-Google domain |
| loylyhelsinki.fi | 5/5 | 161 | Löyly’s own site — the highest single-venue domain count the series has recorded |
| visitfinland.com | 5/5 | 155 | The national tourism board |
| kotiharjunsauna.fi | 5/5 | 123 | Kotiharjun Sauna’s own site |
| fi.hotels.com | 4/5 | 58 | Localized OTA pages, pulled in by the hotel-sauna prompt |
| helsinki-escape.com | 4/5 | 55 | Local guide blog |
| timeout.com | 4/5 | 50 | International city editorial |
| lonna.fi | 4/5 | 49 | Lonna island sauna’s own site |
| reddit.com | 4/5 | 47 | Community threads — exactly one of these is ChatGPT’s |
| stgeorgehelsinki.com | 5/5 | 45 | Hotel St. George — leader of the parallel hotel-sauna set |
| allaspool.fi | 4/5 | 43 | Allas Sea Pool’s own site |
| uusisauna.fi | 4/5 | 38 | Uusi Sauna’s own site |
The table omits google.com — those 757 rows are AI Mode pointing at its parent’s surfaces, and section 9 covers them with the series context.
A hotel crashes the sauna table
stgeorgehelsinki.com is cited by all five engines (45 rows) — more than the island sauna Saunasaari — almost entirely on the hotel-sauna probe prompt. The hotel set it belongs to (St. George, Clarion, Kämp) is real but separate: those venues are excluded from the sauna registry by design, and the hotel prompt is where the EN/FI overlap falls to its 11% floor.
A sauna manufacturer, doing content marketing
Among the industry-layer citations sits solix.no — a Norwegian sauna maker whose “Helsinki sauna guide” landed 24 citations across four engines. Publishing the reference guide for a city you don’t operate in is, on this evidence, a working AI-visibility strategy for a hardware brand.
Where the recommended saunas actually are
The first map is the registry (395 rows, 125 real venues, recovery additions included); the second isolates the twelve consensus venues. The consensus hugs the water: the Hernesaari and Katajanokka waterfronts, the harbour islands (Lonna, Uunisaari, Saunasaari), the sauna rafts moored off Hakaniemi — plus the two inland classics, Kotiharju and Sauna Hermanni. The engines also leave the city when the sauna is famous enough: Kuusijärvi’s smoke sauna (Vantaa) and Löylykontti’s Espoo container turned up often enough to be seeded with their real districts.
All 125 mapped registry saunas across Helsinki
Top 12 saunas by answer mentions — click a marker for per-engine counts
District-accuracy scoring (did the Kallio prompt return Kallio saunas?) relied on ChatGPT’s map widget in earlier studies; with the widget gone this run (section 9) the check has no data source, so we don’t report one rather than approximate it.
Ten studies, one scoreboard
Helsinki against the running series metrics. Each prior value was re-checked against the data arrays of CDMX tacos, Istanbul barbers, Seoul coffee and Berlin tattoo as published on this site. (Istanbul and CDMX published a week apart and neither carries the other’s numbers; this table is the first to hold both.)
| Metric | Prior studies | Helsinki saunas | Reading |
|---|---|---|---|
| ChatGPT own-website % | High: Amsterdam 42 · services 32–35 · Istanbul 21.3 · food/retail 0–10 (CDMX 0) | 45.5% | Series high — the density law confirmed at the top of the curve |
| Registry own-domain density | Istanbul 22.1 · CDMX 12.5 · Seoul ~13 | 78.4% | Densest owned-web market the series has measured |
| Copilot entity-website % | Peak 95–97 (yoga/bikes/tattoo) · CDMX 38.8 full-batch · Seoul 25.7 (n=27) | 71.0% | Back near its band — supply restored, habit intact |
| AI Mode google.com % | Range 53–80 across nine studies | 52.4% | Series low by a hair — still half of everything it cites |
| ChatGPT Reddit share | Band 17–20% since Seoul (Istanbul 18.1) | 0.4% (1 citation) | Collapse — the fallback layer with nothing to fall back for |
| ChatGPT → Instagram | Zero in six consecutive cities | 0 citations | Seventh consecutive zero |
| Top non-Google domain | Reddit ×3 · Instagram ×2 · guide.michelin.com (CDMX) | myhelsinki.fi (246) | First city whose own tourism board tops the table |
| Perplexity booking/aggregator % | Berlin yoga 31 · Istanbul 18.9 · tattoo 4 · bistros 3 | 6.7% | Marketplaces stay marginal where venues own domains |
| Language→TLD coupling | Tokyo .jp 5.0× · Istanbul 1.9× · Berlin yoga 1.5× · tattoo 0.87× | 0.92× | First fully neutral reading — one web serves both languages |
| EN vs local top-5 overlap (control) | Tattoo 67 · most 25 · Marseille 11 · Istanbul 0 · CDMX 0 | 100% | Series first — identical top-5 in both languages |
| Mentions #1 vs cites #1 | Clean doubles: tattoo (OMEN), Istanbul (Rimedzo); artifacts in Marseille/Seoul/CDMX | Löyly 379 · score 162 — clean double | Third clean double, and the first artifact-free board |
Engine field notes — a run with two instrument changes
ChatGPT’s map carousel is gone
Prior runs harvested 550–872 map-widget entities per city from ChatGPT’s embedded venue carousel. This run: 10 entities, from a single Finnish-language capture. The prose answers still name venues (which is what the leaderboard reads), but the structured widget that used to accompany them has all but vanished from the captured surface — worth knowing for anyone whose tooling parses it.
Search telemetry went dark mid-quarter
69 of ChatGPT’s 96 answers carry attached citations, yet the capture’s web-search flag and fan-out-query fields came back empty on 95 of 96 — a consequence of ChatGPT’s late-August switch to a new search-stream format that the scraper’s parser lags. We report no web-search trigger rate this run rather than publish a number the instrument can’t currently see.
Perplexity’s five missing answers
Bright Data returned 26 Perplexity error items (peer failures and an auth-wall detection). A targeted refire of only the missing prompt×proxy pairs recovered 21; the last 5 (2 US-proxy, 3 FI-proxy) stay absent from every number on this page. Its landed answers were unusually decisive — for the first time in the series, the engine behind the study’s single zero-recommendation capture is AI Mode, a column Perplexity had owned in every prior city.
AI Mode, at its series floor and still half Google
757 of AI Mode’s 1,444 citations point at google.com surfaces — 52.4%, a shade under its previous 53–80% range. The remainder skews official (myhelsinki.fi is its favourite external domain), and both FI-proxy batches landed in full, without the rejection behaviour the FR proxy showed in Marseille.
If you heat a sauna for a living
- An own domain pays here, measurably. loylyhelsinki.fi (161 citations) and kotiharjunsauna.fi (123) are two of the four most-cited non-Google domains in the study, and 45.5% of ChatGPT’s citations went to venue-owned sites. In this market a website is not table stakes — it is the answer surface itself.
- Get into the official canon. myhelsinki.fi (246 citations) and visitfinland.com (155) are the two most-cited editorial sources on every engine. A listing in the city’s own guide is worth more AI visibility here than any amount of social presence — the whole social layer is 89 Reddit + Instagram citations combined.
- You don’t need two web presences for two languages. The 100% EN/FI overlap and the 0.92× TLD ratio say the engines read one Finnish sauna web in both languages. Bilingual pages on one domain — the Löyly/MyHelsinki pattern — cover both audiences.
- A marketplace page is not a domain. Venues whose only presence is a saunat.fi or venuu.fi listing collect 2–7% of citations per engine, spread across the whole platform. The engines cite the venue that owns its name.
- Consensus this strong is hard to enter. Löyly appears in 80% of all answers; the top-5 is identical across languages and stable across engines. For a new sauna, the realistic path into AI answers runs through the niches where the consensus thins — the group-rental, family and district prompts where overlap drops and lesser-known venues already surface.
What convergence needed
The interesting thing about Helsinki is how unmysterious it is. Everywhere the series found fragmentation, it found a missing layer behind it: no venue domains in Mexico City, no crawlable identity in Seoul, two barber economies that never wrote about each other in Istanbul. Helsinki has none of those absences. The venues own domains (78.4% of the registry), the city publishes the canonical guide in both languages, and the canon itself is old and famous enough that Finnish and English writers describe the same five places. Feed the engines a complete web and they agree — with each other, across languages, and between their own mention and citation metrics.
For the series, Helsinki closes the density question: the own-website arc now runs 0 → 2.3 → 21.3 → 45.5 against densities of 12.5 → 13 → 22.1 → 78.4, and the old “32% service baseline” is better read as a mid-density plateau. What it opens is a question about the other end of the pipeline: if a complete first-party web produces this much agreement, the next markets worth visiting are ones where the first-party web is complete but contested — where venues own domains and still disagree about who matters. The hotel-sauna bleed probe, stuck at 11% overlap inside this otherwise-converged city, hints at what that looks like.
Study design
Data collection
- 24 prompt templates × 2 languages (EN/FI) × 2 proxy countries (US/FI) × 5 AI engines (ChatGPT, Perplexity, Gemini, Copilot, Google AI Mode) = 480 theoretical captures; 475 landed (98.9%). Captured 2026-08-26 via Bright Data, through the pipeline’s in-process fallback runner (identical payloads, tables and parsing).
- One refire, disclosed: Perplexity’s two batches returned 26 error items (502 no_peers, auth-wall detections). A targeted refire of only the missing prompt×proxy pairs — no duplication of landed captures — recovered 21; the remaining 5 are absent and nothing is imputed. All other batches landed 48/48 on the first pass.
- Totals: 475 captures · 3,493 cited URLs · 3,017 extracted venue mentions, 94.5% resolved — the series high (CDMX 86.6%, Istanbul 82.1%).
- Registry: 376 Apify Google Maps rows from one city crawl — 203 of them apartment/cottage rentals advertising a private sauna, classified out along with gyms, spas, beaches and hotels. Google’s umbrella “Sauna” category and the Finnish swimming halls (sauna+pool complexes) are kept real. Two name-targeted recovery passes added 28 rows the crawl missed — Sauna Arla, Saunasaari, the sauna-raft fleet, USVA Spa, Kuusijärvi’s smoke sauna — including 12 hotel, spa and campground rows seeded as non-registry entities so the hotel-probe mentions resolve without joining the sauna leaderboard; 8 directory-named rows were reclassified out. Final: 395 rows, 125 real sauna venues.
- Spend: Apify $1.59 (seed + two recovery passes) + Bright Data 506 items ≈$0.76 ≈ $2.35 total.
What we measured
- Venues named per answer (brand-aggregated; ranked by mentions, citation score alongside)
- Cited URLs bucketed into a sauna-specific source taxonomy (marketplaces separated from venue identity domains)
- Registry own-domain density vs ChatGPT own-website share — the density-law test this city was chosen for
- EN vs FI top-5 overlap across all 24 templates (position-ordered NER mentions; see instrument note)
- Language→TLD coupling on entity domains
- The official-tourism citation layer, new to this study
- Hotel-sauna entity bleed via a dedicated probe prompt
From prose answers to countable saunas
Engines answer in running text, and Finnish answers inflect names — the same venue surfaces as “Löyly”, “Löylyn sauna” or “Löylystä”, and Helsingin Saunalautta doubles as “Helsinki Sauna Ferry” in English. The pipeline’s NER pass reads each of the 475 answers with an LLM under fixed extraction rules (named venues only, recommendation context required, no inference beyond the text), then a deterministic resolver matches names to the registry: normalised exact match, fuzzy similarity, Finnish-diacritic folding, and a hand-checked alias table for the morphology fuzzy matching can’t bridge. Frequently recommended unresolved names were checked against Google Maps in two recovery passes. Final resolution: 94.5% of 3,017 mention rows.
Extractor disclosure: this run’s extraction was performed by a Claude model rather than the pipeline’s usual Gemini extractor (the sandbox could not reach the Gemini key) — same prompt rules, same deterministic resolution path, run in 11 parallel chunks. As a QA gate, all 3,017 extracted names were mechanically verified to appear in their source answer text; four flagged as Finnish case-inflections were verified by hand, zero came back suspect. The Finnish prompt set was not native-reviewed — the caveat this series has carried for Japanese, Korean, Turkish and Spanish.
Caveats
- Two instrument changes shaped this run (section 9): ChatGPT’s map carousel appeared in 1 of 96 captures, so the language-overlap metric is computed from position-ordered NER mentions instead of the widget top-5 — same Jaccard definition, different source; and ChatGPT’s search-telemetry fields were empty on 95 of 96 captures, so no web-search trigger rate is reported.
- Website-field counts reflect what Google Maps knows. Recovery-pass rows mostly carry websites, but the 78.4% density figure inherits Maps’ own gaps.
- The registry draws a city line the engines don’t. Kuusijärvi (Vantaa) and Löylykontti Matinkylä (Espoo) are recommended for “Helsinki” and seeded with their real districts; one recovery candidate (Kaurilan Sauna) resolved only to its retail-shop listing and was rejected, leaving its 5 mentions unresolved.
- Judgment calls in the registry: Google’s inconsistent “Sauna” umbrella category kept real; 8 platform-named directory rows reclassified out; hotels/spas seeded as non-registry rows. Each is logged in the run’s findings doc.
- NER resolution is precision-first; unresolved mentions are dropped, so venue counts are conservative lower bounds.
- Single-city, single-run snapshot (August 2026); engine behaviour moves with model updates — this run caught two such moves in the act.
- Disclosure: no personal affiliation with any Helsinki sauna.
Frequently Asked Questions
Summarize with AI
Continue Reading
The Istanbul study whose hypothesis this one completes, and the rest of the series.