AI Search for Tacos in Mexico City (2026):Michelin beats every taquería on the internet
TL;DR: We asked five AI engines 25 taco questions in English and Spanish, through US and Mexican proxies — the series’ first complete five-engine grid, 500 of 500 captures. ChatGPT cited a taquería-owned website exactly zero times in 356 citations, the first absolute zero in nine studies. What replaces the missing websites is the guidebook shelf: guide.michelin.com is the most-cited non-Google domain (200 citations, reaching all five engines), with Mexican critic Marco Beteta’s guide at 147. The second zero: English and Spanish answers to “best taquerias in Mexico City” share zero of their top-5 venues. And the consensus winner is El Vilsito — an auto repair shop that becomes a taquería at night, named in 174 of 500 answers.
Executive Summary
Mexico City is the series’ first stop in the Americas, its first Spanish-language market, and the hardest possible test of a question Seoul raised: when a food scene has no websites, what do AI engines actually read?
A taquería is street infrastructure. Of the 722 real venues in our registry, 143 (19.8%) have anything at all in Google’s website field, 53 of those are just a Facebook, Instagram or link-page URL, and 90 rows (12.5%) point at an own-domain site — several of them shared chain domains. In Seoul the same scarcity pushed each engine onto a different social platform. Here the answer came out differently: the vacuum belongs to professional restaurant guides. guide.michelin.com collected 200 citations — more than Instagram (163), more than Reddit (150) — and Marco Beteta’s mbmarcobeteta.com, a guide most tourists have never heard of, took 147.
Two zeros anchor the findings. ChatGPT’s own-website share, which fell from a 32–42% service-vertical band to 8–10% for bookstores and coffee, then to 0.9% for Paris bistros, lands at 0 of 356 citations — the floor, reached. And on the control prompt, the English top-5 and Spanish top-5 share no venues at all: English CDMX eats at Roma Norte counters with websites and Michelin paragraphs; Spanish CDMX eats at El Vilsito and the canasta stands.
The grid itself is a first, too: all five engines × both languages × both proxies landed — 500/500 captures, 5,867 citations — so every cross-engine comparison in this study runs at full n with nothing imputed. One Copilot batch needed a refire, disclosed in the methodology; 11 captures recommend nothing, all of them Perplexity’s.
Two words carry precise meanings on this page: citations are the URLs an engine attaches under its answer, mentions are the venue names inside the answer itself. The source analysis runs on the former, the leaderboard on the latter — and keeping them apart is the whole plot of section 5.
Five engines, five source diets
All 5,867 cited URLs, bucketed per engine. Reading the chart top to bottom: AI Mode cites Google, Gemini cites guidebooks and travel writers, Perplexity spreads across everything including OTAs, ChatGPT reads the food press and Reddit, and Copilot — alone — still tries to cite the taquerías themselves.
Integer percentages via largest-remainder rounding — each column sums to exactly 100. Raw counts per engine: AI Mode 3,071 citations, Gemini 1,187, Perplexity 915, ChatGPT 356, Copilot 338. Small buckets (delivery apps, food tours, hotel blogs, reference) are folded into “Other”.
US food media reads CDMX to you
A quarter of ChatGPT’s citations (25%) are global food press — eater.com and theinfatuation.com tie at 32 citations each — plus 26.7% local editorial and 17.1% Reddit. Its taquería-website count: zero.
The guidebook engine
Gemini’s two favourite domains in the whole study are Beteta’s guide (102 citations) and Michelin (98) — 16.8% of its pool sits in the restaurant-guide bucket, five times any other engine’s share. Add 15% travel blogs and Gemini reads like a concierge with a bookshelf.
Still hunting for entity pages
131 of Copilot’s 338 citations (38.8%) hit venue-owned domains — five times ChatGPT’s share of the same city. When the domain is missing it falls back to Facebook pages (33 citations, its single most-cited domain). Section 4 puts this in series context.
The Michelin effect, measured
In 2024 the Michelin Guide starred El Califa de León, a five-square-metre pastor stand — the first taco stand ever in the guide. Two years later, this study can quantify what that era did to machine-readable Mexico City: the guide’s domain is the single most-cited non-Google source for taco questions, on every engine we fired.
The second name on the guide shelf matters as much as the first. Marco Beteta is a Mexican dining critic whose site, mbmarcobeteta.com, took 147 citations — fourth among all non-Google domains, ahead of Time Out Mexico and every local newspaper. Gemini in particular treats Beteta and Michelin as its ground truth: 102 and 98 citations respectively, 16.8% of everything it cited. An engine picking a local critic over the global brand as its #1 source is the most locally-literate retrieval decision any engine made in this study.
Note what did not fill the gap. In Seoul, the missing website layer routed citations into social platforms — Instagram, Naver, Reddit — and we framed the result as engines each picking a social surface. Mexico City breaks that framing: the social bucket here stays moderate (3–19% per engine), while curated editorial — guides, city press, food press — takes the weight. Street food’s missing web went to critics, not to influencers.
0 of 356: ChatGPT’s own-website share hits the floor
Nine studies in, the own-website chart has a bottom, and Mexico City found it. Every prior market left at least a sliver — Paris bistros managed 3 URLs (0.9%), Seoul 7 (2.3%). ChatGPT’s CDMX run produced not one citation to any taquería-owned domain.
The zero is over-determined. Supply barely exists (90 own-domain rows in a 722-venue registry) — but supply alone doesn’t explain it, because Copilot found 131 venue-domain citations in the same city. ChatGPT’s retrieval simply prefers a well-written third-party roundup over a thin first-party page, and CDMX taco coverage is the richest third-party corpus this series has touched: Eater’s guides, The Infatuation’s lists, Michelin’s paragraphs, a decade of Reddit threads. Given that menu, it never once needed elvilsito.com.
The Reddit line, meanwhile, barely moved between continents: 61 of 356 citations (17.1%) against Seoul’s 17.4%. Two food markets, nine time zones apart, and ChatGPT allocates Reddit a near-identical sixth of its pool — while its Instagram count stays at zero in both. Whatever weighting sits behind that pairing, it is engine configuration, and no amount of local social-media effort appears to move it.
Seoul’s n=27 question, answered at n=100
Seoul reported Copilot’s entity-website share collapsing from its 74–97% series band to 25.7% — on 27 surviving captures, published with a caveat on every figure. CDMX reran the situation with a complete batch: 100 captures, 338 citations, and a share of 38.8%. The collapse is real; the caveat can retire.
Two things are now separable that Seoul’s partial batch couldn’t separate. First: the entity habit survives — 38.8% is still the highest own-website share of any engine here by a factor of nearly five, and Copilot dug up genuinely obscure taquería domains (elvilsito.com, taqueriaelborregoviudo.mx, taqueriaorinoco.com) that other engines ignored. Second: the ceiling is set by the market, and where the domain supply runs out, Copilot swaps in the venue’s Facebook page. Its identity as the entity engine was never a fixed percentage; it’s a resolution strategy whose output tracks how much ownable web the vertical actually has.
An auto shop wins Mexico City
By day, El Vilsito fixes cars in Narvarte. After dark it serves what many locals rank among the city’s best al pastor — and all five AI engines agree: named in 174 of 500 captured answers, more than any other venue, with every engine recommending it in at least a fifth of its captures. Los Cocuyos, the Centro beef-snout counter Anthony Bourdain filmed, follows at 154. Every one of the twelve venues below appears on all five engines.
Per-engine mention matrix
| # | Taquería | AI Mode100 captures | ChatGPT100 captures | Perplexity100 captures | Gemini100 captures | Copilot100 captures |
|---|---|---|---|---|---|---|
| 1 | El Vilsito | 40%(40) | 34%(34) | 21%(21) | 42%(42) | 37%(37) |
| 2 | Los Cocuyos | 33%(33) | 32%(32) | 19%(19) | 32%(32) | 38%(38) |
| 3 | Taquería Orinoco | 21%(21) | 21%(21) | 25%(25) | 20%(20) | 25%(25) |
| 4 | Los Especiales | 27%(27) | 16%(16) | 4%(4) | 24%(24) | 15%(15) |
| 5 | Tacos Hola El Güero | 20%(20) | 20%(20) | 1%(1) | 20%(20) | 5%(5) |
| 6 | El Califa de León | 15%(15) | 19%(19) | 6%(6) | 14%(14) | 10%(10) |
| 7 | Cariñito Tacos | 16%(16) | 11%(11) | 3%(3) | 12%(12) | 21%(21) |
| 8 | Tacos El Huequito | 13%(13) | 10%(10) | 5%(5) | 24%(24) | 7%(7) |
| 9 | Taquería El Gran Abanico | 10%(10) | 15%(15) | 1%(1) | 9%(9) | 12%(12) |
| 10 | El Borrego Viudo | 14%(14) | 8%(8) | 1%(1) | 2%(2) | 19%(19) |
| 11 | El Pescadito | 9%(9) | 9%(9) | 8%(8) | 8%(8) | 8%(8) |
| 12 | Tacos Charly | 8%(8) | 6%(6) | 3%(3) | 13%(13) | 11%(11) |
| # | Taquería | Answer mentions | Citation score | Engines (of 5) |
|---|---|---|---|---|
| 1 | El Vilsito | 174 | 26 | 5 |
| 2 | Los Cocuyos | 154 | 3 | 5 |
| 3 | Taquería Orinoco | 112 | 51 | 5 |
| 4 | Los Especiales | 86 | 10 | 5 |
| 5 | Tacos Hola El Güero | 66 | 1 | 5 |
| 6 | El Califa de León | 64 | 0 | 5 |
| 7 | Cariñito Tacos | 63 | 4 | 5 |
| 8 | Tacos El Huequito | 59 | 1 | 5 |
| 9 | Taquería El Gran Abanico | 47 | 1 | 5 |
| 10 | El Borrego Viudo | 44 | 2 | 5 |
| 11 | El Pescadito | 42 | 6 | 5 |
| 12 | Tacos Charly | 41 | 11 | 5 |
Now read the citation-score column against the mention column, because the two metrics crown different winners and both stories are instructive. Taquería Orinoco — the Monterrey-born chain with locations in Roma and Polanco and an actual website — tops citations at 51 while sitting third on mentions (112). El Vilsito, the mention champion, scores 26. Los Cocuyos (154 mentions) scores 3. El Califa de León, the Michelin-starred stand this study’s guide layer orbits around, scores 0 — no domain, nothing for a citation counter to match.
The raw cite-ranked CSV also carries a warning exhibit at #2: “BEST TACOS IN MÉXICO CITY”, score 29 — a real Google Maps listing whose registered name is a keyword string. It rides ChatGPT’s map widget on name-match alone, appears in no engine’s prose recommendations at that rank, and the mapdata generator’s artifact guard keeps it off the board above. Anyone selling “AI visibility scores” for local businesses is one keyword-stuffed listing away from ranking it first.
The domains that feed the answers
Cross-engine domain table, normalized. 9 domains reach all five engines — Michelin leads that club at 200 citations, and the rest of it is small local editorial (foodandtravel.mx, excelsior.com.mx) plus one aggregator (tourme.app). ChatGPT’s social presence is Reddit-shaped, Copilot and AI Mode carry the Instagram/Facebook weight, and Gemini touches neither.
| Domain | Engines | Cites | What it is |
|---|---|---|---|
| guide.michelin.com | 5/5 | 200 | The Michelin Guide — most-cited non-Google domain, reaching every engine |
| instagram.com | 3/5 | 163 | Taquería profiles; zero of these come from ChatGPT or Gemini |
| reddit.com | 4/5 | 150 | Community threads; ChatGPT’s biggest single source (61) |
| mbmarcobeteta.com | 4/5 | 147 | Marco Beteta’s Mexican dining guide — Gemini’s #1 domain (102) |
| cdmxsecreta.com | 4/5 | 111 | Secret CDMX — local listicle press |
| youtube.com | 4/5 | 105 | Street-food tour videos |
| facebook.com | 3/5 | 101 | Taquería pages; Copilot’s #1 domain (33) |
| timeoutmexico.mx | 4/5 | 99 | Time Out Mexico — local editorial |
| chilango.com | 4/5 | 84 | Chilango — CDMX city magazine |
| mymexicotrip.com | 4/5 | 61 | Travel blog |
| tourme.app | 5/5 | 59 | Tour/review aggregator — cited by all five engines |
| foodandtravel.mx | 5/5 | 48 | Food & Travel México — all five engines |
google.com is excluded from the table: its 2,136 citations all come from AI Mode referencing Google’s own surfaces and would bury every other row (see section 10).
The local press holds its ground
cdmxsecreta.com (111), timeoutmexico.mx (99) and chilango.com (84) put Mexican city media on four of five engines each — a healthier local-editorial showing than Seoul’s, where tourism portals dominated. Spanish-language answers in particular draw on this layer heavily.
Instagram splits the engines cleanly
All 163 Instagram citations come from AI Mode (145), Copilot and Perplexity. ChatGPT: zero, for the third study running. Gemini: zero as well — its social bucket is the smallest of the five engines at 3%, crowded out by its guidebook diet.
Ask in Spanish, eat somewhere else
“Best taquerias in Mexico City” and “las mejores taquerías de la Ciudad de México” are the same question. ChatGPT’s top-5 answers to them share zero venues — 0 of 10 distinct names. The series’ previous low was Marseille’s 11%; most markets sit near 25%. Across all 22 comparable templates, 17 overlap at 11% or less and none reaches 43%.
| Prompt template | EN vs ES top-5 overlap |
|---|---|
| district: Polanco | 43% (3/7) |
| tacos de canasta | 43% (3/7) |
| district: Coyoacán | 29% (2/7) |
| district: Centro Histórico | 25% (2/8) |
| fish tacos (pescado) | 25% (2/8) |
| vegan tacos | 11% (1/9) |
| local persona | 11% (1/9) |
| tourist persona | 11% (1/9) |
| cheap eats | 11% (1/9) |
| safe late-night areas | 11% (1/9) |
| birria | 11% (1/9) |
| carnitas | 11% (1/9) |
| breakfast tacos | 11% (1/9) |
| control (best taquerias) | 0% (0/10) |
| street food stands | 0% (0/10) |
| district: Roma Norte | 0% (0/10) |
| district: Narvarte | 0% (0/10) |
| gourmet tacos | 0% (0/10) |
| Michelin-recognised tacos | 0% (0/8) |
| barbacoa | 0% (0/10) |
| suadero | 0% (0/10) |
| late-night tacos | 0% (0/10) |
Overlap = shared venues ÷ distinct venues across both top-5 lists (Jaccard).
Nine of the 22 templates overlap at exactly zero, and the two languages fail in different places: even district prompts split (Polanco converges at 43%, Roma Norte and Narvarte at 0%), and the two questions you’d most expect to converge — the control and the Michelin prompt — both come back disjoint. The joint leaders, Polanco and tacos de canasta, share a mechanism: a small universe of famous venues that both webs know. Everywhere else, English answers assemble the food-press canon (Orinoco, Cariñito, the Roma Norte circuit) while Spanish answers surface neighbourhood institutions the English-language web has never written up.
Two translation accidents made the miss vivid. Perplexity answered one Spanish street-food prompt (“los mejores puestos de comida callejera”) with resources about homelessness — it parsed de calle as “situación de calle” and returned social-services links. Copilot, asked for the city’s best barbacoa (slow-pit lamb), recommended Pinche Gringo BBQ — an American-style smokehouse. Both answers read fluently; both are the wrong cuisine, and only a bilingual reviewer would ever notice.
The companion metric — whether ES prompts pull more .mx domains — is unmeasurable here for the same reason it was in Seoul: with ChatGPT citing zero venue domains, the entity-TLD sample has no numerator. That the measurement keeps failing because the entity layer keeps not existing is itself a series finding.
Where the recommended tacos actually are
The first map holds the full registry (828 rows, 722 real taquerías); the second isolates the twelve consensus venues. Their geography is narrower than the city’s: the AI favourites string along the Roma–Condesa–Centro–Polanco corridor, with El Vilsito’s Narvarte block as the southern outpost — even though the seed crawl itself skewed hard to the periphery (Iztapalapa alone contributed 91 rows before the district-targeted second pass rebalanced it).
All 722 mapped registry taquerías across Mexico City
Top 12 taquerías by answer mentions — click a marker for per-engine counts
District prompts: mostly obeyed, except Coyoacán
For the six colonia-targeted templates, we checked whether ChatGPT’s resolved venues actually sit in the asked-for neighbourhood:
| District prompt | EN accuracy | ES accuracy |
|---|---|---|
| Narvarte | 100% (14/14) | 100% (9/9) |
| Condesa | 100% (2/2) | 100% (2/2) |
| Roma Norte | 100% (14/14) | 88% (7/8) |
| Centro Histórico | 90% (9/10) | 94% (15/16) |
| Polanco | 87% (13/15) | 77% (10/13) |
| Coyoacán | 33% (3/9) | 29% (2/7) |
This is the best district discipline the series has recorded — five of six colonias at 77–100% in both languages (Condesa’s 100% rests on tiny denominators: 2 resolved venues per language). The exception is systematic: Coyoacán prompts scored 33% (EN) and 29% (ES), because the engine repeatedly recommends Centro institutions for Coyoacán questions — plausibly an echo of a hundred “day trip to Coyoacán” itineraries that mix both neighbourhoods into one listicle.
Nine studies, one scoreboard
CDMX against the running series metrics. Prior values re-read from the data arrays inside Seoul coffee, Paris bistros, Berlin tattoo and Marseille coffee as published on this site.
| Metric | Prior studies | CDMX tacos | Reading |
|---|---|---|---|
| ChatGPT own-website % | Services 32–42 (yoga ×2, bikes, tattoo) · food/retail: Tokyo 8 · Marseille 10 · Seoul 2.3 · bistros 0.9 | 0.0% | The floor. First absolute zero in the series |
| Copilot entity-website % | Tattoo 95 · Tokyo 89 · Marseille 83 · bistros 74 · Seoul 25.7 (n=27) | 38.8% (full batch) | Seoul’s collapse confirmed at full n — market-limited, engine habit intact |
| AI Mode google.com % | Range 53–80 (tattoo 53 · Seoul 57 · Berlin yoga 59 · Tokyo 61 · bistros 70 · Marseille 80) | 69.6% | Inside range — the one constant across every market |
| ChatGPT Reddit share | Seoul 17.4% (and 0 Instagram) | 17.1% (and 0 Instagram) | Nearly identical across continents — the ratio travels with the engine rather than the market |
| Top non-Google domain | Reddit (Tokyo/bistros/tattoo) · Instagram (Marseille, Seoul) | guide.michelin.com (200) | First study where a professional guide outcites every social platform |
| Language→TLD coupling | Tokyo .jp 5.0× · Berlin yoga .de 1.5× · Seoul unmeasurable | Unmeasurable | Zero entity citations → no numerator, second study running |
| EN vs local top-5 overlap (control) | Tattoo 67 · Tokyo/bistros/Seoul 25 · Marseille 11 | 0% | Series low — English CDMX and Spanish CDMX are disjoint |
| Mentions #1 vs cites #1 | Marseille: Deep vs an IG artifact · Seoul: Fritz (0 cites) vs artifact | El Vilsito 174 vs Orinoco 51 · keyword-name listing at cite-#2 | Citation counting inverts the ranking; artifact guard caught a new failure shape |
Engine-level field notes
Perplexity is the only engine that comes back empty
All 11 zero-recommendation captures in the study are Perplexity’s (8 Spanish, 3 English) — every other engine named venues in all 100 of its answers. Seoul saw the same signature at a higher rate in Korean. Across two non-English markets now, Perplexity is the engine most likely to hand back generic advice instead of names.
AI Mode, 69.6% self-referential
2,136 of AI Mode’s 3,071 citations point at google.com surfaces — near the top of its 53–80% series range and indistinguishable from its behaviour in Paris or Tokyo. Both MX-proxy batches landed without the trigger rejections that plagued the FR proxy in earlier studies.
ChatGPT searched the live web 88 times in 100
88 of ChatGPT’s 100 captures triggered live web search, and the run yielded 872 map-widget entities across engines. The no-search answers leaned on memorised canon — which in this city means the same Michelin-era names the search results would have returned anyway.
The refire that saved the grid
Copilot’s US batch initially returned a ready-but-empty snapshot; one refire recovered all 50 items, completing the 500/500 grid. Berlin tattoo’s 520/520 was the last complete grid — and this is the first since the pipeline moved to its in-process fallback runner. Full batch details live in the methodology.
If you serve al pastor for a living
- Guide and press coverage is the visibility lever. Michelin (200 cites), Beteta (147), Eater and The Infatuation (32 each from ChatGPT alone) carry venues into AI answers that own no web presence whatsoever. One well-placed review outperforms a website you don’t have.
- Your Google Maps name is machine-read literally. A keyword-string listing name reached cite-#2 in our raw ranking on name-matching alone. The flip side: an accurate, distinctive name is what lets NER-style systems (including ours) resolve your mentions correctly at all.
- English and Spanish visibility are separate campaigns. Zero control-prompt overlap means Eater coverage moves your English answers and does nothing for “las mejores taquerías” — that answer set is built from chilango.com, cdmxsecreta and local press.
- A cheap own-domain site still buys Copilot. 38.8% of its citations went to the 90 venue domains that exist — Orinoco’s site made it the citation leader of the whole study. One engine of five isn’t much; it’s also the difference between cite score 51 and cite score 0.
- Facebook still counts here. 33 of Copilot’s citations and a share of AI Mode’s social slice go to Facebook pages — in Mexico the platform remains a live web-presence substitute that engines actually link.
Who speaks for a business that has no website?
Seoul suggested that when the owned web disappears, every engine improvises with a social platform. Mexico City corrects the record: given a scene with world-class critical coverage, the engines put the critics in charge. The most consequential writing about El Vilsito is no longer a review a diner might read — it is whatever Michelin, Beteta, Eater and a decade of Reddit threads said, blended per engine into an answer the taquero will never see and cannot edit.
For the series, CDMX closes two open loops. Copilot’s Seoul collapse replicated at full sample size, so the entity engine’s ceiling is now established as a property of the market’s domain supply. ChatGPT’s own-website slide bottomed out at a literal zero, which converts “food and retail scenes underperform the service baseline” from a trend into a boundary condition. What CDMX opens is the guide question: does the critic layer take over anywhere the food is famous enough — Oaxaca, Lyon, Bangkok — or does it require a Michelin event like the Califa star to crystallise? The next food city this series visits should have no stars at all, as a control.
Study design
Data collection
- 25 prompt templates × 2 languages (EN/ES) × 2 proxy countries (US/MX) × 5 AI engines (ChatGPT, Perplexity, Gemini, Copilot, Google AI Mode) = 500 theoretical captures; 500 landed. Every engine finished 100/100 (AI Mode’s MX batch logged 47 successful at the run level and completed to 100 captures on the same snapshot). Captured 2026-08-05 via Bright Data.
- One refire, disclosed: Copilot’s US snapshot first returned ready-but-empty; a single retry recovered all 50 items. Nothing else failed; nothing anywhere is imputed. This is the series’ first complete five-engine grid since Berlin tattoo (520/520), and the first ever on the in-process fallback runner (Modal is unreachable from the run sandbox; payloads, tables and parsing are identical).
- Totals: 500 captures · 5,867 cited URLs · 872 map-widget entities · 3,026 extracted venue mentions, 86.6% resolved to the registry — the series high (Seoul 71.9%, bistros 70%).
- Registry: 740 Apify Google Maps rows in two passes (a city-wide crawl that skewed to outer alcaldías — Iztapalapa alone gave 91 rows — plus a district-targeted pass over Roma Norte/Condesa/Centro/Polanco/Coyoacán/Narvarte), 721 distinct after upsert, 615 real after classifying out generic restaurants, torterías and retail. Google Places recovery added 107 venues the crawl missed — the canonical scene: Tacos Hola El Güero, El Huequito, El Califa de León, Los Especiales, El Borrego Viudo — with 10 candidates rejected by validation. Final: 828 rows, 722 real taquerías.
- Spend: Apify $2.96 (two passes) + Bright Data ≈$0.83 (550 items incl. the refire) + Google Places ≈$4.80 (150 lookups) ≈ $8.6 total.
What we measured
- Venues named per answer (brand-aggregated; ranked by mentions, citation score shown for contrast)
- Cited URLs bucketed into a fifteen-part source taxonomy (eleven shown; small buckets folded into “other”)
- The restaurant-guide layer (Michelin/Beteta) per engine — new to this study
- ChatGPT own-website share vs the eight prior cross-vertical cases
- EN vs ES top-5 overlap across 22 comparable templates
- District-targeting accuracy for six named colonias
- Zero-recommendation captures and cross-language misfires (barbacoa→BBQ, callejera→homelessness)
From prose answers to countable taquerías
Engines reply in running text, and the same stand may surface as “El Vilsito”, “Taquería El Vilsito” or “the auto-shop taqueria in Narvarte”. The pipeline’s NER pass reads each of the 500 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 by normalised exact match and fuzzy similarity. Frequently recommended names that failed to resolve were checked against Google Places, which is how 107 canonical venues entered the registry. Final resolution: 86.6% of 3,026 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 10 parallel chunks. As a QA gate, all 3,026 extracted names were mechanically verified to appear in their source answer text; zero came back suspect. The Spanish prompt set is written in Mexican register but was not native-reviewed — the same limitation the Tokyo and Seoul studies carried for Japanese and Korean.
Caveats
- Website-field counts reflect what Google Maps knows: Places-recovered rows carry no website field, so “90 own-domain rows of 722” is built on incomplete data. The layer is genuinely thin; the exact figure is soft.
- Chain domains group venues: the citation score aggregates by domain, so multi-branch brands (El Califa, El Huequito, Orinoco) pool their citations across locations.
- Condesa’s 100% district accuracy rests on 2 resolved venues per language — read it as “no misses observed” rather than a robust rate.
- NER resolution is precision-first; unresolved mentions are dropped, so brand counts are conservative lower bounds.
- Citation scores are structurally unreliable where venues own no domains (section 5); we publish them to document the failure mode, and flagged the keyword-name listing as an artifact rather than a ranking.
- Single-city, single-run snapshot (August 2026); engine behaviour moves with model updates.
- Disclosure: no personal affiliation with any CDMX taquería.
Frequently Asked Questions
Summarize with AI
Continue Reading
The Seoul study this one answers, and the rest of the series.