Ask AI the Same Question in Three Languages. Get Three Different Answers.

AGR Research  |  AI Hotel Recommendations by Language  |  Miami 2026

We asked ChatGPT, Gemini and Google’s AI for Miami’s best luxury hotels in English, Spanish and Portuguese from the US, Mexico and Brazil. The lists didn’t match.

A traveler in São Paulo, one in Mexico City and one in the United States can ask AI the same question about Miami’s best luxury hotels and get different lists back. Part of that is language. But the lists also changed when the same question was asked again in the same language from the same country. A hotel missing from that answer is missing for that traveler. And AI recommended one hotel 23 times that was imploded seven months ago.

Published: November 17, 2026 Fieldwork: November 17, 2026 Market: Miami Cells: English US, Spanish US, Spanish Mexico, Portuguese Brazil
56%
of hotels were shared when the same question was asked again in the same language and country
51%
were shared when the same question was asked in a different language or country
23
recommendations of Mandarin Oriental, Miami, a hotel demolished on April 12, 2026
219
days after the implosion, the hotel was still being recommended

Basis: 224 AI requests on November 17, 2026. Ten Miami luxury hotel questions, four AI surfaces (ChatGPT, Gemini, Google AI Mode, Google AI Overview), four language-and-origin cells. The 56% and 51% figures come from two questions asked three times in every cell. 195 requests returned an answer. The capture dataset and scoring ledger are available upon request.

The lists changed with the language. They also changed when nothing changed at all. Asked the same question twice, in the same language from the same country, the engines shared 56% of their hotels. Asked in a different language or from a different country, they shared 51%. Most of the movement was ordinary answer-to-answer variation. Changing language or country moved the lists somewhat further.

Quick Answers

Does AI recommend different Miami luxury hotels in English, Spanish and Portuguese?

The lists differ. In AGR’s November 17, 2026 study, on the two repeated questions, answers on the same AI surface shared 56% of their hotels within the same language and origin and 51% across language-and-origin cells. Holding the US origin constant, English and Spanish answers shared 49%, against 61% for repeated answers within either US cell. Most of the difference was already present when the question was simply asked again.

Which Miami hotels did AI recommend in every language?

Four Seasons Hotel Miami, Faena Hotel Miami Beach, The Setai, Four Seasons Hotel at The Surf Club and The St. Regis Bal Harbour Resort were among the most frequently named properties in all four cells. The four cells named 45 distinct properties combined, and 32 to 41 each. Seven properties appeared in only one cell, each in a single answer.

Is AI still recommending the demolished Mandarin Oriental, Miami?

Yes. On November 17, 2026, Mandarin Oriental, Miami was recommended in 23 answers: 15 from Gemini, 7 from ChatGPT and 1 from Google AI Overview. The hotel closed on May 31, 2025 and was demolished by controlled implosion on April 12, 2026, 219 days before the fieldwork. Ten of the 23 were in Portuguese. Two more Gemini answers recommended a different hotel and wrongly described it as the former Mandarin Oriental.

Did the AI engines say the hotel was closed?

Sometimes. ChatGPT answered the same Brickell question twelve times. Seven answers recommended the demolished hotel and four said it had closed. In Spanish from the US, ChatGPT reported the closure on the first run and recommended the hotel on the next two. Gemini never mentioned the closure. Google AI Mode did not recommend the hotel.

What does the language of the question change?

The cited sources and whether Google answered at all. Spanish- and Portuguese-edition websites made up under 1% of cited domains in English, 5% in Spanish from the US and 9% from both Mexico and Brazil. Google showed an AI Overview for 10 of 14 English searches and 5 or 6 of 14 in Spanish and Portuguese.

Scope: These findings describe one day of fieldwork in one market, ten traveler questions, four AI surfaces and four language-and-origin cells. The repeat comparison rests on two questions asked three times. The study does not establish permanent platform behavior, booking outcomes, causal source effects or a complete factual-error rate.

At a glance


  1. One AI answer is one draw. Asked again in the same language and country, two answers shared only 56% of the hotels named across the pair.
  2. Changing the cell moved the lists further. Different-cell answers shared 51% of their hotels, compared with 56% within a cell. Holding the US origin constant produced a larger language gap: 49% versus 61%.
  3. A demolished hotel kept appearing. Mandarin Oriental, Miami was recommended 23 times, 111 days after AGR’s Index documented the same error.
  4. The same engine gave both answers. ChatGPT reported the closure in some answers and recommended the hotel in others, for the same question in the same language.
  5. Language changed what AI cited. Spanish and Portuguese answers carried a regional layer of sources that English answers almost never cited.

What the study measures


Miami is a major US gateway for Latin American travelers, and many of them plan in Spanish or Portuguese. A traveler in Mexico City or São Paulo who asks an AI assistant for Miami’s best luxury hotels may never see the English answer. This study asks whether those travelers receive the same recommendations as an English speaker in the United States.

AGR wrote ten luxury traveler questions about Miami and asked each one in four cells: English from the United States, Spanish from the United States, Spanish from Mexico and Portuguese from Brazil. Each question went to ChatGPT, Gemini, Google AI Mode and Google AI Overview. Two questions, on Brickell and on high-end shopping, were asked three times in every cell. That repeat test measures how much the answers change when nothing changes, the baseline any language comparison needs.

Overlap, the main measure, is the share of hotels named as places to stay in either of two answers that appear in both. The study also records recommendation frequency by cell, cited sources and specific factual errors. It does not measure bookings, hotel quality or the probability that a traveler will choose a property.

Same question, different answers


Exhibit 1 shows what changed across the four cells. It describes the differences; it does not establish that language caused them. Exhibit 2 compares those differences with ordinary run-to-run variation; the US-only comparison later in this section provides the study’s cleanest test of language.

Exhibit 1

Answers naming each property, by language and origin. The same hotels led in every cell; the counts moved.

PropertyEnglish, USSpanish, USSpanish, MexicoPortuguese, Brazil
Four Seasons Hotel Miami28212224
Faena Hotel Miami Beach20222021
The Setai14171724
Four Seasons Hotel at The Surf Club16151921
The St. Regis Bal Harbour Resort12191519
EAST Miami18171717
Acqualina Resort & Residences128811
Kimpton EPIC Hotel121197
The Miami Beach EDITION91067
1 Hotel South Beach96107
Mandarin Oriental, Miami (demolished)44510
Answers returned52484847
Distinct properties named37323541

Source: AGR’s November 17, 2026 capture dataset, 224 requests, 195 answers. A property is counted once per answer that names it as a place to stay. Answers that mentioned Mandarin Oriental only to report its closure, or that misidentified another hotel as the former Mandarin Oriental, are not counted as Mandarin Oriental recommendations. Questions 3 and 8 were asked three times in every cell, so the cells remain comparable.

Read across any row and the counts move. The Setai appeared in 24 Portuguese answers and 14 English answers. The St. Regis Bal Harbour Resort appeared in 19 Spanish answers from the US and 12 English answers. Read down the columns and the leaders hold: the same handful of properties led every cell.

Across all ten questions on a single run, two answers to the same question on the same AI surface in two different cells shared 48% of their hotels on average. On that evidence alone, language and country would look decisive. The repeat test changes the reading. On the Brickell and shopping questions, two answers in the same cell shared 56% of their hotels. Two answers in different cells shared 51%.

Exhibit 2

Hotels shared between two answers: asked again in the same cell versus asked in a different cell

Legend: Same question, same cell, asked again Same question, different cell ChatGPT Gemini Google AI Mode All surfaces ChatGPT, same cell: 55% ChatGPT, different cell: 56% Gemini, same cell: 64% Gemini, different cell: 56% Google AI Mode, same cell: 45% Google AI Mode, different cell: 41% All surfaces, same cell: 56% All surfaces, different cell: 51% 55%56% 64%56% 45%41% 56%51% Average share of hotels two answers have in common. Questions 3 and 8. All surfaces includes Google AI Overview. n = 75 same-cell pairs, 323 different-cell pairs.

On smaller screens, swipe or scroll horizontally to view the full chart.

Source: AGR’s November 17, 2026 capture dataset. Same-cell pairs compare all three runs with each other within one cell and AI surface. Different-cell pairs compare every run in one cell with every run in another, same question and surface. Results are averaged across pairs. Pair counts describe comparisons, not independent observations; the same answer contributes to several pairs. One answer that named no hotel was excluded. Google AI Overview appeared too rarely on these two questions to report separately.

For ChatGPT, a second answer in the same language and country was no closer to the first than an answer from another cell. Gemini showed a gap of about eight percentage points and Google AI Mode about four. For a traveler, which run they receive mattered more than the measured cell difference.

Holding country constant. The two US cells held the requested origin constant while changing the question’s language and, for Google surfaces, the interface language. On the repeated questions, two answers within one US cell shared 61% of their hotels (38 pairs); an English and a Spanish answer from the US shared 49% (53 pairs). That is the cleanest language comparison in the study, and it shows a larger gap than the pooled figure. It rests on two questions and should be read as directional.

The result matches AGR’s earlier repeat work. The five-week follow-up to AGR’s ten-destination study found that leading hotels were usually stable while individual recommendation lists frequently differed. This study adds language to that picture: the leaders held across English, Spanish and Portuguese, and the rest of the list moved in every cell.

The four cells


English from the United States produced the most answers and the most AI Overviews. Google showed an AI Overview for 10 of 14 English searches. Four Seasons Hotel Miami led with 28 answers, more than in any other cell. Spanish- and Portuguese-edition sources were nearly absent: 2 of 291 domain occurrences, counting each domain once per answer.

Spanish from the United States added a Spanish citation layer. The St. Regis Bal Harbour Resort rose to 19 answers from 12 in English. Spanish editions of travel and tourism sites made up 14 of 282 domain occurrences. Google showed an AI Overview for 6 of 14 searches.

Spanish from Mexico cited more regional sources. 21 of 235 domain occurrences were Spanish-edition or Mexican sites, including the Spanish-language edition of Greater Miami’s official tourism site and Mexican editions of Expedia and Tripadvisor. Four Seasons Hotel at The Surf Club and The Setai each appeared in 17 or more answers.

Portuguese from Brazil had the strongest Setai showing and the most Mandarin Oriental recommendations. The Setai appeared in 24 answers, tying Four Seasons Hotel Miami for the lead, the only cell where it did. 25 of 270 domain occurrences were Portuguese-edition or Brazilian sites, including Brazilian travel blogs and the Portuguese edition of Greater Miami’s tourism site. Brazil also received 10 of the 23 recommendations for the demolished Mandarin Oriental. Google showed an AI Overview for 5 of 14 searches, the fewest of any cell.

The demolished hotel, 111 days later


AGR’s Luxury Hotel AI Visibility Index documented on July 29, 2026 that ChatGPT and Google AI Mode recommended Mandarin Oriental, Miami five times, 108 days after the building was demolished. The hotel closed permanently on May 31, 2025. Local 10 reported the controlled implosion on April 12, 2026.

On November 17, 2026, 111 days after the Index fieldwork, the hotel was recommended in 23 answers. The error had not gone away, and it appeared on different surfaces. Google AI Mode, which recommended the hotel twice in July, did not recommend it in November. Gemini, which did not recommend it in the July Index answers, recommended it 15 times and never mentioned the closure. ChatGPT recommended it seven times, all in answers to the Brickell question.

Mistaken identity. Two Gemini answers in Spanish from the US recommended Brickell Arch, a Luxury Collection Hotel, and described it as the former Mandarin Oriental. Brickell Arch is a different property at a different address. Those two answers are counted as identity errors, not as Mandarin Oriental recommendations.

Exhibit 3

Mandarin Oriental, Miami was recommended 5 times in July and 23 times in November, under different study designs

May 31, 2025 April 12, 2026 July 29, 2026 November 17, 2026 Hotel closes permanently Building demolished by implosion Index: recommended 5 times by 2 platforms This study: recommended 23 times by 3 AI surfaces 108 days 111 more days

On smaller screens, swipe or scroll horizontally to view the full chart.

Timeline is schematic, not to scale. The July and November captures used different protocols, surfaces and question sets. They are separate sightings of the same error, not a measured trend. AI recommendations are recorded in AGR’s capture datasets.

Correct in one answer, wrong in the next. The repeat test exposed something a single capture cannot. ChatGPT answered the Brickell question twelve times across the four cells. Seven answers recommended Mandarin Oriental, Miami, some with its former address and phone number. Four reported that it had closed and left it out. In Spanish from the US, ChatGPT reported the closure on the first run and recommended the hotel on the second and third. In Portuguese from Brazil, it recommended the hotel on the first run and reported the closure on the next two. The same engine gave the correct status in some answers and recommended the demolished hotel in others, for the same question in the same language.

Renovation closure. The Ritz-Carlton Bal Harbour, Miami appeared in 12 of 41 answers to the luxury shopping question, across all four AI surfaces and all four cells. None mentioned that the resort is closed. The hotel’s official website states that it is undergoing a transformation and plans to welcome guests again starting February 2027. One day earlier, in AGR’s study of ten luxury hotel openings, ChatGPT, Gemini and Google AI Overview all described the same resort as closed when asked about it directly. Asked about the hotel, the answers were correct. Asked for a recommendation, 12 answers listed it without the closure.

Exhibit 4

Closed-hotel recommendations by AI surface, November 17, 2026

AI surfaceMandarin Oriental, Miami recommendedAnswers reporting its closureRitz-Carlton Bal Harbour recommended (shopping question)
Gemini1504 of 12
ChatGPT744 of 12
Google AI Overview101 of 5
Google AI Mode003 of 12
Total23412 of 41

Source: AGR’s November 17, 2026 capture dataset. Gemini also produced two answers that misidentified Brickell Arch as the former Mandarin Oriental; they are excluded from the Mandarin Oriental count. No answer that recommended The Ritz-Carlton Bal Harbour, Miami mentioned its closure. Google AI Overview appeared on 5 of the 12 shopping searches.

What language changed


Differences between cells were clearer in the visible citation mix and AI Overview availability than in the identity of the leading hotels.

Exhibit 5

Spanish- and Portuguese-edition sources and AI Overview appearance, by cell

CellSpanish- or Portuguese-edition domain occurrencesGoogle AI Overview shownExamples from the cell’s overall citation mix
English, US2 of 291 (0.7%)10 of 14Greater Miami’s English tourism site, Forbes Travel Guide, Condé Nast Traveler, hotel brand sites
Spanish, US14 of 282 (5.0%)6 of 14Spanish edition of Greater Miami’s tourism site, Spanish-language luxury hotel lists, Spanish and Latin American Tripadvisor editions
Spanish, Mexico21 of 235 (8.9%)6 of 14Spanish edition of Greater Miami’s tourism site, Mexican editions of Expedia and Tripadvisor
Portuguese, Brazil25 of 270 (9.3%)5 of 14Portuguese edition of Greater Miami’s tourism site, Tripadvisor Brazil, Brazilian travel blogs

Source: cited-source logs in AGR’s November 17, 2026 capture dataset. Counts are domain occurrences: each cited domain counted once per answer, summed by cell. A domain counts as a Spanish- or Portuguese-edition source when it carries a Spanish-speaking country or Brazilian domain, a Spanish or Portuguese language subdomain, or is the Spanish or Portuguese edition of Greater Miami’s tourism site. Spanish or Portuguese pages on international domains are not captured. Citations do not reveal every source or influence behind an answer.

English answers cited almost no Spanish- or Portuguese-edition sources. Spanish and Portuguese answers continued to cite many of the same international editorial and brand domains, while adding a more visible regional-edition layer of 4 to 9 percentage points. The leading hotels largely stayed the same while the cited sources shifted. For a property, that means Brazilian and Mexican travelers may be shown answers citing sources that an English-only audit would never see.

The AI Overview gap affects visibility before any hotel is named. A Spanish or Portuguese search was less likely to return a Google AI answer at all. When no AI Overview appears, the traveler sees conventional results instead.

What this means for a luxury hotel


A hotel that checks its AI visibility once, in English, has seen one draw from a moving list. The same question asked again, or asked in Spanish or Portuguese, can return a different set of hotels. A single answer shows whether a hotel appeared once. Repeated checks show how often it appears across the tested questions, languages and origins.

For properties that sell to Latin American travelers, the regional citation layer is part of what those travelers are shown. Spanish and Portuguese answers cited regional editions of destination, booking and review sites and local travel blogs that English answers almost never cited. An English-only audit can miss that regional citation layer.

The closed-hotel results show a second problem. Engines that describe a hotel’s status correctly when asked about it directly still list it in recommendations. Correcting a property record is necessary, but it does not guarantee that every answer will reflect the correction. This observational study does not show that any content change will improve recommendations or generate bookings.

AGR resources. AGR’s playbook on how to get a hotel recommended by AI sets out the sequence for that work. The AI Visibility Audit examines a property’s presence and representation in AI answers, and the real-time AI visibility report shows how the platforms describe a property today. Knowledge Formation Optimization (KFO), AGR’s discipline for work on the public information environment, is defined in the KFO canonical reference.

Methodology


Design. One market, Miami. Four cells: English from the United States, Spanish from the United States, Spanish from Mexico and Portuguese from Brazil. Four AI surfaces: ChatGPT, Gemini, Google AI Mode and Google AI Overview. Ten traveler questions: best luxury hotels, five-star beachfront in Miami Beach, Brickell or downtown, family resort, romantic stay, spa, oceanfront suites, high-end shopping, dining and where to stay for a luxury trip. Each question was asked once in every cell on every surface, producing 160 requests.

Repeat test. Question 3 (Brickell or downtown) and Question 8 (high-end shopping) were asked two more times in every cell on every surface, adding 64 requests for a total of 224. Each of those two questions therefore has three runs per cell and surface.

Prompt format. Each question was asked on its own, with no instructions, formatting requests or follow-up questions. AGR wrote the Spanish and Portuguese questions for Mexican and Brazilian travelers rather than machine-translating the English. Both Spanish cells used the same Spanish wording.

Capture conditions. All requests were captured on November 17, 2026 through consumer AI surfaces without a signed-in account, routed through the stated country. Google AI Mode and Google AI Overview were also set to a city location (New York for both US cells, Mexico City and São Paulo) and an interface language matching the question. ChatGPT and Gemini were located at country level. The two US cells held the requested origin constant while changing the question’s language and, for Google surfaces, the interface language. The Mexico and Brazil cells changed both language and location.

Answer availability. ChatGPT, Gemini and Google AI Mode returned an answer to every request. Google AI Overview appeared on 27 of 56 searches. One Google AI Mode answer asked follow-up questions and named no hotel; it counts as a returned answer but is excluded from overlap calculations.

Scoring. A property counts once per answer that names it as a place to stay, regardless of position, including through its named spa or restaurant and through map or place cards shown with the answer. Names that appear only in cited-source titles or links are not counted. Overlap between two answers is the number of properties both name divided by the number named in either. First-run comparisons pair every two cells for each question and surface (197 pairs). In the repeat test, same-cell pairs compare all three runs within one cell and surface (75 pairs); different-cell pairs compare every run in one cell with every run in another for the same question and surface (323 pairs). Averages are taken across pairs. Pair counts describe comparisons, not independent observations.

Property normalization. Names were matched to a normalized list of Miami properties and aliases, merging spa, restaurant and brand-variant references into the parent hotel. A name used only as a property’s former name, such as SLS Brickell for Dua Miami or Conrad for Brickell Arch, is not counted as a separate property. The scoring ledger lists every property counted for every answer. Answers naming Mandarin Oriental, Miami were classified three ways: a recommendation; a closure report, where the answer stated the hotel had closed and excluded it; or an identity error, where the answer applied the Mandarin Oriental name to a different hotel. Only recommendations are counted in the Mandarin Oriental totals.

Limits. Results apply to one market on one day. The repeat comparison rests on two questions; other questions may vary more or less. Language and location change together in the Mexico and Brazil cells. The source classification is domain-based and does not capture Spanish or Portuguese pages on international domains. The study does not estimate booking effects, isolate source causation or calculate a complete factual-error rate.

Prior work


AGR’s July 15, 2026 Miami capture recorded AI recommendations for Mandarin Oriental, Miami and The Ritz-Carlton Bal Harbour while both were closed. The AGR Luxury Hotel AI Visibility Index, with July 29, 2026 fieldwork across six US luxury markets including Miami, documented the same errors in English from a New York origin and listed a repeat capture to examine answer stability among its planned additions. This study is a separate November capture, not a reanalysis of the Index dataset.

AGR’s September 27, 2026 study, Which Hotels Do AI Engines Recommend? A Study of 10 Destinations, used three repetitions per question and engine, and its five-week follow-up compared lists over time. What AI Knows About 10 Luxury Hotel Openings in 2027, captured November 16, 2026, recorded how the engines described The Ritz-Carlton Bal Harbour, Miami when asked about it directly. These studies used different protocols and are not pooled with this dataset.

How to cite this study


Andrew Paul, Americas Great Resorts. Ask AI the Same Question in Three Languages. Get Three Different Answers. Published November 17, 2026. Fieldwork: November 17, 2026. Permanent study page.

Suggested short citation: “In AGR’s November 17, 2026 Miami study, AI answers to the same luxury hotel question shared 56 percent of their hotels when asked again in the same language and country, and 51 percent across language-and-origin cells.”

Alternative short citation: “AI surfaces recommended the demolished Mandarin Oriental, Miami 23 times in AGR’s November 17, 2026 study, 219 days after its implosion.”

The capture dataset, covering 224 requests with cell, surface, question, run, answer text and cited sources, and the scoring ledger, with the normalized properties, Mandarin Oriental classification and pair-level overlap for every comparison, are available upon request. Journalists and researchers may reproduce the exhibits above with attribution. Dataset and media inquiries: info@americasgreatresorts.net.

This study is published by Americas Great Resorts, operating in independent luxury hospitality since 1993. No hotel paid to appear in this study, and inclusion cannot be purchased. Cite the fieldwork date with its findings.

Close