AI Search for Tango Schools in Buenos Aires (2026):the canon is bilingual
TL;DR: Five AI engines, 24 tango questions, English and Spanish, US and Argentine proxies — 392 of 480 captures landed. The last two tourist-heavy cities in this series (Istanbul, Mexico City) each answered its two languages with two disjoint lists; Buenos Aires answers them with largely the same one: 43% top-5 overlap on the control prompt, and ask for the best milongas and both languages return the identical five venues. La Viruta is named in 240 of 392 answers, ChatGPT’s own-website share lands at 25.3% — the fifth city measured on the density curve, and the one that fills its empty middle — and its onetime 17–20% Reddit habit is now 0 citations out of 368.
Executive Summary
A pattern had started to look like a law: in cities where tourists and locals search in different languages, the engines serve each group its own city. Then we pointed the methodology at the most tourist-saturated vertical we could think of, and the pattern didn’t show up.
This is the series’ first South American city, and its first classes-and-lessons vertical since the Paris and Berlin yoga studies back in May — you leave with an ocho, not an espresso. Tango school prompts in English and Spanish share 43% of their top-5 ( Istanbul barbers and Mexico City tacos both measured 0%), the milonga prompt returns the identical five venues in both languages, and the language→TLD coupling is a mild 1.33× — Spanish prompts put 24.8% of citations on .ar domains, English prompts 18.6%. The venues that carry the agreement are exactly the ones with global names: La Viruta, La Catedral, El Beso, Club Gricel.
The series’ central law also picked up its missing middle point. The registry’s own-domain density is 37.7% — between Istanbul’s 22.1% and Helsinki’s 78.4% — and ChatGPT’s own-website citation share came in at 25.3% (93 of 368), between their 21.3% and 45.5%. Five cities measured, five in rank order. And a new engine personality surfaced: 24.2% of Gemini’s citations go to dinner-show ticketing and tango tour agencies — ask it where to learn, it answers partly with where to buy a ticket.
By the numbers: 392 of 480 captures landed, carrying 2,025 cited URLs and 2,281 extracted venue mentions, 83.8% of them resolved against a 441-row registry. Perplexity is the disclosure this time: 8 captures of 96 landed after both of its batches returned error items and one targeted refire — every Perplexity number on this page is a raw count rather than a percentage, and the 88 missing captures stay missing instead of being filled in.
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.
Two languages, one milonga list
Mexico City’s taco answers and Istanbul’s barber answers described two different cities depending on the prompt language. Buenos Aires was supposed to be the extreme case of that split — no vertical is more tourist-facing than tango. Instead the split half-closed.
| Prompt template | EN vs ES top-5 overlap |
|---|---|
| best milongas (bleed control) | 100% (5/5) |
| district: Palermo | 100% (5/5) |
| district: downtown/centro | 67% (4/6) |
| district: San Telmo | 67% (4/6) |
| tourist persona | 67% (4/6) |
| control (best tango schools) | 43% (3/7) |
| tango shows (bleed control) | 43% (3/7) |
| milongas with a beginner class | 43% (3/7) |
| oldest academies | 43% (3/7) |
| district: Almagro | 40% (2/5) |
| group classes | 25% (2/8) |
| most famous schools | 25% (2/8) |
| classes taught in English | 25% (2/8) |
| one-week intensive | 25% (2/8) |
| salon-style tango | 25% (2/8) |
| prácticas | 11% (1/9) |
| complete beginners | 11% (1/9) |
| where porteños learn | 11% (1/9) |
| going alone, no partner | 11% (1/9) |
| milonguero style | 11% (1/9) |
| tango nuevo | 11% (1/9) |
| private lessons | 0% (0/10) |
| advanced dancers | 0% (0/10) |
| cheap or free classes | 0% (0/10) |
Overlap = shared venues ÷ distinct venues across both top-5 lists (Jaccard). Top-5 read from ChatGPT’s map carousel, the series’ original instrument — present again in 95 of 96 captures this run after vanishing in Helsinki (section 9).
Read the table top to bottom and the mechanism is plain. Prompts anchored to world-famous venues or walkable places agree: the milonga list is identical in both languages (La Viruta, La Catedral, El Beso, Club Gricel, Parakultural — all five shared), Palermo is at 100%, San Telmo and the tourist persona at 67%. The control prompt shares three of seven distinct venues: Escuela Mundial de Tango, La Viruta, and a San Telmo studio Google lists as “Clases de Tango”.
Divergence lives where the canon runs out. Private lessons, advanced training and cheap-or-free classes overlap at exactly zero — the English answers go to tourist-bookable teachers, the Spanish ones to neighborhood programs and city cultural centers. “Where do porteños actually learn” overlaps 11%, which is arguably the honest answer to the question.
The density law gets its middle point
Since Seoul, this series has tracked a single predictor for ChatGPT’s own-website citations: how much of the vertical owns a domain at all. Four measured cities sat at the extremes. Buenos Aires finally sits in the middle.
Of the 257 real schools, milongas and instructors in the registry, 162 (63%) list some web presence and 97 (37.7%) an own-domain site — the rest live on Instagram, Facebook, WhatsApp links, or a page inside the city government’s portal. That mid-range supply produced a mid-range share: ChatGPT put 25.3% of its 368 citations on venue-owned domains, 93 citations in all.
Five cities have now been measured directly, and the two columns rank identically in every one. Nothing else the series tracks — vertical type, language, continent, tourist volume — predicts the own-website share this well.
Every engine dances at La Viruta
The basement of the Armenian Cultural Association in Palermo — part school, part milonga, open till dawn — is named in 240 of 392 answers (61.2%), by all five engines, in both languages, from both proxies.
Who recommends whom — the per-engine matrix
Share of each engine’s captures whose answer names the venue. The engines agree on the podium and then each keeps a private favorite: Copilot’s is Escuela Mundial (68.8%), Gemini leans La Viruta and DNI, ChatGPT spreads widest.
| # | Venue | AI Mode96 captures | ChatGPT96 captures | Perplexity8 captures | Gemini96 captures | Copilot96 captures |
|---|---|---|---|---|---|---|
| 1 | La Viruta Tango Club | 61.5%(59) | 59.4%(57) | 8 | 66.7%(64) | 54.2%(52) |
| 2 | Escuela Mundial de Tango Gabriela Elias | 46.9%(45) | 35.4%(34) | 1 | 14.6%(14) | 68.8%(66) |
| 3 | La Catedral Club | 24%(23) | 4.2%(4) | 5 | 35.4%(34) | 34.4%(33) |
| 4 | El Beso | 16.7%(16) | 24%(23) | 2 | 28.1%(27) | 10.4%(10) |
| 5 | RECOLETA TANGO STUDIO | 7.3%(7) | 25%(24) | 0 | 5.2%(5) | 42.7%(41) |
| 6 | "Club Gricel" Tango Argentina | 19.8%(19) | 14.6%(14) | 0 | 27.1%(26) | 14.6%(14) |
| 7 | Palermo Tango Studio | 2.1%(2) | 17.7%(17) | 0 | 0(0) | 46.9%(45) |
| 8 | La Escuela Del Tango | 6.2%(6) | 11.5%(11) | 1 | 8.3%(8) | 28.1%(27) |
| 9 | DNI Tango Store | 10.4%(10) | 12.5%(12) | 0 | 29.2%(28) | 2.1%(2) |
| 10 | Mariposita de San Telmo | 10.4%(10) | 12.5%(12) | 0 | 15.6%(15) | 7.3%(7) |
| 11 | Salón Canning | 9.4%(9) | 3.1%(3) | 1 | 21.9%(21) | 7.3%(7) |
| 12 | CETBA Universidad del tango | 15.6%(15) | 8.3%(8) | 0 | 17.7%(17) | 0(0) |
Perplexity column shows raw counts on its 8 landed captures — a denominator too small for percentages (section 9).
Top 12 venues by answer mentions — click a marker for per-engine counts
Count mentions and count citations, and you crown two different schools
This series scores every city two ways on purpose. By answer mentions — how often an engine actually names a venue — La Viruta wins with 240 and a citation score of just 40. By citation counting — whose web pages the engines attach as sources — the winner is Escuela Mundial de Tango Gabriela Elías: score 131, its own domain drawing 63 citations from three engines, while sitting second on mentions (160).
The gap is the story. La Viruta’s fame lives in prose, guidebooks and city pages; Escuela Mundial’s SEO-tuned multilingual site is what the retrieval step actually reads. A citation-counted dashboard would rank the school above the landmark — and would be measuring the wrong thing, because the recommendation a user hears is the mention.
The raw citation-counted board also produced this run’s artifact: an Instagram-only studio (“Pa’ Bailar”) briefly “won” with a score of 177 because it lists instagram.com as its website, inheriting all 70 of Copilot’s Instagram citations. The identity guard added after Marseille’s equivalent case caught it; it appears on no board on this page.
Ask Gemini for a class, get a ticket counter
Each engine keeps its structural personality, and Buenos Aires adds a wrinkle the series hadn’t seen: a dinner-show commerce layer. 140 of Gemini’s 579 citations (24.2%) point at show-ticketing and tango tour agencies — showdetango.com alone takes 81, making a ticket reseller Gemini’s most-cited domain in a study about learning to dance.
Perplexity is not charted: its 8 landed captures produced 10 citations, a sample that would render as noise dressed up as a distribution.
352 of 526 citations on venue-owned sites. Copilot leads the other engines here, as it has everywhere I have recomputed this, but how much it gets follows the market: the share has run from 25.7% in Seoul coffee to 97% in Amsterdam bike shops, median 83%. Buenos Aires lands under that median, beside Istanbul’s 67% and Helsinki’s 71%. Social takes 16% of its column, and Instagram is the part of that I can name: 70 citations, 13.3% of everything Copilot cited, and no other engine cited Instagram once.
From a 17–20% Reddit share in four consecutive studies to 0.4% in Helsinki to literally zero here. Study-wide, reddit.com received 7 citations — 5 from Gemini, 2 from Perplexity, none from anyone else.
All 542 of its citations are google.com searchviewer links — the domain dimension died as an instrument in September. Section 9 and the svid study cover what happened.
ChatGPT trusts city hall more than the schools
Add up ChatGPT’s citations to the Buenos Aires city government — turismo.buenosaires.gob.ar plus buenosaires.gob.ar — and you get 94. Its citations to every venue-owned website combined: 93. The single most-cited non-Google domain in the whole study is the city’s own tourism guide at 85 citations across three engines.
Helsinki showed the first version of this — myhelsinki.fi topping the table — and Buenos Aires makes it two cities in a row, with a twist: here the government pages aren’t just guides. The city runs milonga listings, free classes at cultural centers, and program pages that dozens of venues use as their only web presence. When a porteño-persona prompt asks where locals learn, the engines answer from the city’s own cultural program.
For the engines’ purposes, Buenos Aires documents its dance better than the dance documents itself.
The twelve domains that answer tango questions
google.com sits out of this table: its 542 citations are all AI Mode pointing at itself. Among real destinations, the pattern is one government guide, one ticket seller, one social platform — and then schools’ own websites, which is where the citation-counted leaderboard is decided.
| Domain | Engines | Citations | What it is |
|---|---|---|---|
| turismo.buenosaires.gob.ar | 3/5 | 85 | The city government’s tourism guide — the study’s most-cited non-Google domain |
| showdetango.com | 1/5 | 81 | Dinner-show ticketing agency; every one of these citations is Gemini’s |
| instagram.com | 1/5 | 70 | Every one of these citations is Copilot’s |
| pixidia.com | 2/5 | 67 | Generic places directory |
| escuelamundialdetango.com | 3/5 | 63 | Escuela Mundial de Tango — the citation-counted #1 |
| adrianlunacoach.com | 3/5 | 59 | A private-lesson coach’s own site |
| lavirutatangoar.com | 3/5 | 48 | La Viruta — the consensus #1 by answer mentions |
| tangovida.org | 3/5 | 46 | Tango portal / blog |
| buenosaires.gob.ar | 2/5 | 46 | The city government proper — cultural-center program pages |
| mariposita.com.ar | 3/5 | 43 | Mariposita de San Telmo’s own site |
| elizabethmatiasarte.com | 2/5 | 37 | Recoleta Tango Studio’s own site |
| getyourguide.com | 4/5 | 29 | The only domain in the study cited by four engines |
Engine counts out of the 5 fired; a domain cited only by Gemini reads 1/5. getyourguide.com’s 4/5 makes the experiences marketplace the study’s most widely shared source — every engine except AI Mode reaches for it.
Where the machines put tango on the map
The registry spans the whole city — Almagro and Villa Crespo hold the milonga belt, San Nicolás and Monserrat the downtown academies, Palermo the tourist-facing studios. The engines’ district accuracy follows the canon: asked for Palermo classes, ChatGPT keeps 70% of its recommendations in Palermo; asked for San Telmo, 50%; asked for “downtown”, 0% — because the registry stores downtown as its real barrio names (San Nicolás, Monserrat), a seed-granularity artifact the series has hit before in Amsterdam and Berlin, disclosed rather than repaired.
All 257 mapped registry venues across Buenos Aires
What the instruments did this run
Perplexity failed, twice, and is reported as counts only
Both Perplexity batches (48 prompts × US, × AR) returned as unmatched error items; 5 captures survived the first pass and a targeted refire of only the missing prompt×proxy pairs recovered 3 more, all US-proxy. That leaves 8 of 96 (10 citations, 37 mentions). Its counts appear in study totals and the matrix; no Perplexity percentage appears anywhere on this page, and the 88 missing captures are disclosed in the methodology and the Dataset schema rather than filled in.
AI Mode’s source trail is now a hall of mirrors
100% of AI Mode’s 542 citations are google.com/searchviewer links — the September infrastructure change decoded in the searchviewer svid study: publisher URLs are wrapped in Google’s own viewer. The series’ “AI Mode google.com %” metric, 52–75% for ten studies, retires at exactly 100.
ChatGPT’s map carousel came back
A week earlier in Helsinki the carousel appeared in 1 of 96 captures and the overlap metric had to be rebuilt from prose mentions. Here it rendered in 95 of 96 captures, delivering 1,369 map entities — so section 1 is measured with the series’ original instrument again. Same model family, seven days apart. Treat any single-run widget behavior as weather.
Three answers recommended nothing
3 of 392 captures name no venue at all — all three are AI Mode captures that answered in generalities. Every other answer in the study recommended at least one resolvable place.
If you run a tango school in Buenos Aires
An own domain is the single measurable lever. 37.7% of the registry has one; those venues collect ChatGPT’s 93 own-site citations and dominate Copilot’s 66.9%. An Instagram page collects Copilot’s Instagram bucket — attributed to the platform, never to you.
Get into the city’s official listings. The government portal outdraws every venue site in this study (85 + 46 citations). A line in the municipal milonga agenda or a cultural center program is machine-visible in a way a flyer will never be.
Publish in both languages, but know what it buys. The famous-venue canon already crosses languages by itself; the 0%-overlap prompts — private lessons, advanced training, cheap classes — are where an English page still decides whether tourists can find you at all.
Watch where Gemini sends your students. A quarter of its citations are ticket resellers. If your school also runs a show, the resellers are describing you to the machines — on their margin, in their words.
Conclusion
Eleven studies ago this series began with a simple question — which businesses do the machines recommend, and from what evidence. Buenos Aires answers with its most structured picture yet: a mention canon led by a basement milonga every engine knows, a citation economy led by whichever school built the best website, a government portal quietly outranking both, and one engine funneling students toward ticket counters.
The result worth carrying forward is the language one. The tourist-city split that looked structural after Istanbul and Mexico City turns out to be conditional: it happens where fame is local-language-only. Where the venues are famous in every language, the engines converge on one canon — and tango’s canon has been global for a century. The corpus was multilingual before the models were.
Next in the series: a vertical where the density law predicts a share we haven’t observed yet, and a continent it hasn’t visited.
Methodology
The grid
24 prompt templates (control, styles, formats, levels, personas, districts, price, plus milonga and dinner-show entity-bleed controls) × 2 languages (English, Rioplatense Spanish) × 5 engines (ChatGPT, Perplexity, Gemini, Copilot, Google AI Mode) × 2 proxy countries (US, AR), fired 2026-09-02 via Bright Data through the pipeline’s in-process fallback runner (Modal is unreachable from the run sandbox; identical payloads and storage). 392 of 480 captures landed: ChatGPT, Gemini, Copilot and AI Mode each 96/96; Perplexity 8/96 — both its batches returned error items, one targeted refire recovered 3 captures, and the 88 absent captures are disclosed, never imputed.
The registry
432 Apify Google Maps rows from tango-specific searches (“tango school”, “escuela de tango”, “clases de tango”, “tango lessons”, “milonga”), classified by a name-verified category map: schools, milongas and instructor-led studios kept real (257 rows) — Google files famous milongas under Dance hall, Night club, Ballroom, Playground and Tourist attraction, all kept — while dinner-show houses, generic cultural centers, Spanish schools and tango-shoe retail are classified out. Three name-targeted recovery passes added 9 venues the engines recommend that the broad pull missed (dinner-show houses seeded as non-venues for the bleed probe). One judgment call is flagged for review in the PR: DNI Tango’s only Google listing is its street-level shop, reclassified here as the school it fronts.
Extraction and resolution
Venues named in each prose answer were extracted by an LLM NER pass — a Claude extractor this run, standing in for the usual Gemini one under the same rule set, with every one of the 2,281 extracted names mechanically verified to appear verbatim in its source answer. Names were resolved to the registry in strict precedence: exact normalized match, fuzzy (≥0.86), Spanish-diacritic folding, a hand-checked alias table, then unique token containment — 83.8% of mention rows resolved. Two same-address duplicate Google listings (La Viruta’s two pages, La Catedral’s two pages) are merged by street address, documented in the run’s resolution script. Unresolved names stay flagged as outside-registry recommendations; the largest are a dance-agenda app (BailaBA) and four instructor brands with no Google Maps presence.
Counting rules
Citations are per-capture deduplicated URLs; source buckets are keyword-and-domain based with venue identity decided only by registry website domains (a keyword heuristic was tested and rejected this run — in a vertical where every portal contains “tango”, it misfiled ~200 citations). Social platforms, city-government pages and the hoy-milonga agenda never count as a venue’s identity domain. The leaderboard counts each venue once per capture; the citation score counts ChatGPT map presence plus citation-domain matches. Percentages in the source-mix chart are largest-remainder rounded per engine to sum to exactly 100. All headline stats are published as summary.csv.
Frequently Asked Questions
Related studies
The density law's top point, and the first 100% language overlap
The other Latin American city — where the two languages shared nothing
Zero English–Turkish overlap — the series floor, matched by Mexico City
Why AI Mode's citation column reads 100% google.com this run