AGR Research | Luxury Hotel AI Visibility | 5-Week Follow-Up
AGR repeated 100 questions across 4 AI engines to see which hotels stayed visible and which recommendations changed.
7/10
destinations kept the same leading hotel across the 2 collections
400
November answers from 100 questions across 4 AI engines
83.6%
September same-day hotel sets matched neither reference answer: 317/379
86.8%
November hotel sets matched neither of the same reference answers: 329/379
Basis: September repetition 3 and November answers tested against September repetitions 1 and 2 in 379 complete question-and-engine groups. Across all 3 reference-pair checks, paired gaps ranged from 3.2 to 6.9 percentage points.
Quick Answers
Did AI recommend the same hotels about 5 weeks later?
7 of 10 destinations kept the same leading hotel. Individual hotel sets frequently differed, including among September answers collected on the same day. A hotel could retain frequent visibility while other properties in the list changed.
How did the 5-week comparison relate to same-day variation?
In 1 matched check, 83.6% of September lineups and 86.8% of November lineups matched neither of the same 2 reference answers. Rotating the reference pair produced paired gaps of 3.2 to 6.9 percentage points. These observations do not isolate elapsed time as the cause.
Which engine reproduced a September hotel set most often?
When November answers were compared with all 3 September lineups, Copilot matched at least 1 in 47 of 100 groups, or 47.0%. ChatGPT, Google AI Mode and Gemini together matched at least 1 in 14 of 279 groups, or 5.0%. This is a separate test from the 2-reference check above.
What counts as a different hotel lineup?
The test compares identified hotel membership. Adding or removing 1 hotel creates a different set, even when every other hotel remains. Changes in order or wording do not create a nonmatch.
Which AGR study was repeated?
The follow-up repeats the fixed questions from AGR’s September 27, 2026 10-destination study. The July 29 6-market AI Visibility Index is separate research with a different question set and counting method.
At a glance
- Most destination leaders persisted. 7 of 10 leaders remained first. Hotel Bel-Air’s November lead was 1 appearance; 5 leaders had margins of 2.
- Same-day variation was already high. The matched checks recorded paired same-day versus November nonmatch gaps of 3.2 to 6.9 percentage points.
- Engine patterns differed. Copilot reproduced a September hotel set more often than the other engines in the 3-reference comparison.
- Identified lists were longer overall. The mean was 5.19 hotels per September answer and 5.86 per November answer. The main matched check showed an association between longer identified lists and higher nonmatching for ChatGPT and Gemini.
- Rank and appearance rate can move differently. Tortuga Bay’s recorded rate was 47.9% in September and 52.5% in November, while its destination rank moved from 1st to 3rd.
What this follow-up compares
7 of 10 destinations kept the same leading hotel when Americas Great Resorts repeated its AI hotel recommendation study about 5 weeks later.
Individual hotel lineups frequently differed, including among answers collected on the same day. In 1 matched check, 83.6% of September lineups and 86.8% of November lineups matched neither of the same 2 reference answers. Across all 3 reference-pair checks, the difference ranged from 3.2 to 6.9 percentage points. High same-day variation was already present before the 5-week interval.
The distinction matters. A hotel can continue to appear frequently while the other properties recommended alongside it vary. Destination leadership and an identical recommendation list measure different forms of consistency.
AGR's September study of AI hotel recommendations across 10 destinations assessed 1,177 substantive answers collected on September 27, 2026. On November 2, AGR asked the same 100 questions across the same 4 engines: ChatGPT, Google AI Mode, Gemini and Copilot.
September used 3 repetitions per question on each engine. November used 1, producing 400 substantive answers. The follow-up examines the recorded results on those 2 dates.
7 destinations kept the same leading hotel
For each date separately, AGR identified the hotel with the most qualifying appearances across a destination's 10 questions and 4 engines. The same appearance-counting convention was used for both collections. September and November answers were not pooled together to determine either date's leader.
7 leaders retained first place. The November runner-up counts show that several leads were narrow.
Exhibit 1
Destination leaders and November runners-up
| Destination | September leader | November leader and appearances | November runner-up and appearances |
|---|---|---|---|
| New York City | Mandarin Oriental, New York | Mandarin Oriental, New York: 18 | The Peninsula New York: 16 |
| Miami area | The Setai, Miami Beach | Faena Miami Beach: 25 | Four Seasons Hotel at The Surf Club: 20 |
| Las Vegas | Wynn Las Vegas | Wynn Las Vegas: 26 | Four Seasons Hotel Las Vegas and The Venetian: 24 each |
| Los Angeles area | Hotel Bel-Air | Hotel Bel-Air: 22 | The Beverly Hills Hotel: 21 |
| Orlando area | Four Seasons Orlando | Four Seasons Orlando: 32 | Ritz-Carlton Orlando, Grande Lakes: 30 |
| Chicago | The Peninsula Chicago | The Peninsula Chicago: 32 | Four Seasons Hotel Chicago: 29 |
| Cancún / Costa Mujeres | Excellence Playa Mujeres | Excellence Playa Mujeres: 23 | Atelier Playa Mujeres: 21 |
| Punta Cana | Tortuga Bay | Secrets Cap Cana Resort & Spa: 26 | Eden Roc Cap Cana: 24 |
| Los Cabos | Waldorf Astoria Los Cabos Pedregal | One&Only Palmilla: 28 | Waldorf Astoria Los Cabos Pedregal: 24 |
| Honolulu / Waikīkī | Halekulani | Halekulani: 33 | The Kahala: 28 |
Source: AGR’s September 27 and November 2, 2026 reviewed appearance ledgers. Each November destination has 40 assessed answers. Leaders are calculated separately for each collection.
September destination denominators ranged from 116 to 120. Some names are shortened for readability. These destination totals combine engines; a destination leader need not lead each engine individually.
Hotel Bel-Air led by 1 appearance. 5 leaders led by 2: Mandarin Oriental, New York; Wynn Las Vegas; Four Seasons Orlando; Excellence Playa Mujeres; and Secrets Cap Cana. These margins describe the collected answers and leave limited separation between the leading properties.
The Peninsula Chicago remained frequently featured across both collections: 96 of 118 September answers, or 81.4%, compared with 32 of 40 November answers, or 80.0%. Halekulani appeared in 91 of 118 September answers, or 77.1%, and 33 of 40 November answers, or 82.5%.
Rank and appearance rate moved differently
The leader changed in Miami, Punta Cana and Los Cabos. The former leaders continued to appear frequently.
The Setai featured in 51 of 117 Miami-area answers in September, or 43.6%. Its November rate was similar: 17 of 40, or 42.5%. Faena led November with 25 of 40, or 62.5%.
Tortuga Bay shows why a ranking needs its underlying counts. It led Punta Cana in September with 56 appearances in 117 answers, or 47.9%. Its November rate was 21 of 40, or 52.5%, yet it placed third.
The competitors ahead of it recorded larger differences between their 2 observed rates. Secrets Cap Cana featured in 41.9% of September answers and 65.0% of November answers. Eden Roc Cap Cana featured in 45.3% and 60.0%, respectively. Both November rates exceeded Tortuga Bay's 52.5%.
That explains the ranking mechanically: the destination order depends on competing properties as well as the hotel's own appearance rate. It does not explain why an engine selected any particular hotel.
Repeating the same question on the same day produced different hotel lists
The September report found that only 16 of 379 question-and-engine groups produced the same hotel set in all 3 repetitions. Membership varied at least once in the other 363 groups, even though those answers were collected on the same date.
That provides essential context for the November comparison. A different lineup 5 weeks later cannot, by itself, establish an effect of the interval between collections.
AGR therefore calculated a comparison with the same number of reference answers on each side. Using September repetitions 1 and 2 as references, it tested:
- Whether September's repetition 3 matched either reference.
- Whether November's answer matched either of those same references.
The calculation uses the 379 groups with 3 substantive September answers. The September collection contained 9 recorded collection errors and 14 nonanswers. Those 23 unusable responses were concentrated in 21 of the original 400 question-and-engine groups, leaving those groups without complete September coverage. Those 21 groups were excluded from the lineup comparison. Their available substantive answers remain included in the destination counts.
Exhibit 2
Same-day and 5-week nonmatches against the same 2 reference answers
| Engine | Eligible question-and-engine groups | September repetition 3 differed from both references | November answer differed from both references |
|---|---|---|---|
| ChatGPT | 91 | 77/91: 84.6% | 90/91: 98.9% |
| Google AI Mode | 91 | 85/91: 93.4% | 84/91: 92.3% |
| Gemini | 97 | 86/97: 88.7% | 96/97: 99.0% |
| Copilot | 100 | 69/100: 69.0% | 59/100: 59.0% |
| Total | 379 | 317/379: 83.6% | 329/379: 86.8% |
Source: AGR’s matched question-and-engine comparison, n = 379 groups. September repetition 3 and the November answer are each compared with September repetitions 1 and 2. Order and wording are ignored.
Exhibit 3
Hotel-set nonmatches varied by engine on the same-day and 5-week checks
On smaller screens, swipe or scroll horizontally to view the full chart.
Source: the comparison in Exhibit 2. Grey bars show the September same-day check; orange bars show November. A different set can result from 1 added or removed hotel.
The overall difference between those 2 recorded rates was 3.2 percentage points. The engine results varied: November had more unmatched lineups for ChatGPT and Gemini, slightly fewer for Google AI Mode, and fewer for Copilot.
AGR repeated the check 3 ways, each time holding out a different September repetition. Each comparison used the same 2 reference answers for both the held-out September answer and the November answer. The paired gaps were 6.9, 6.6 and 3.2 percentage points, respectively. High same-day variation remained evident in every check.
Exhibit 4
Each held-out September repetition and November answer tested against the same reference pair
| September answer held out | Reference repetitions | September nonmatches | November nonmatches | Gap, percentage points |
|---|---|---|---|---|
| Repetition 1 | 2 and 3 | 303/379: 79.9% | 329/379: 86.8% | 6.9 |
| Repetition 2 | 1 and 3 | 307/379: 81.0% | 332/379: 87.6% | 6.6 |
| Repetition 3 | 1 and 2 | 317/379: 83.6% | 329/379: 86.8% | 3.2 |
Source: AGR’s 379 complete question-and-engine groups. Gaps use unrounded fractions before rounding to 1 decimal place; they compare the paired rows, not the extremes of unrelated ranges.
These comparisons use identified hotel sets. Order, wording and repeated mentions are ignored. Adding 1 hotel is enough to create a different set. For example, 1 September ChatGPT answer to “What are the best luxury hotels in Chicago, Illinois?” featured 6 hotels. The November answer retained all 6 and added Pendry Chicago.
Copilot reproduced earlier lineups more often
A separate comparison gave each November answer all 3 September lineups as possible matches. Across the 379 eligible groups, 61 November lineups matched at least 1 September lineup and 318, or 83.9%, matched none.
Copilot accounted for 47 of those 61 matches. It reproduced at least 1 September hotel set in 47 of 100 comparisons, or 47.0%. ChatGPT, Google AI Mode and Gemini together did so in 14 of 279 comparisons, or 5.0%.
The 2-reference table provides a further check on that difference. Copilot already had a lower same-day nonmatch rate than the other engines in September. Its November result should be read alongside that existing difference between engines.
November's identified lists were longer overall
September's reviewed records contained 6,110 qualifying hotel appearances across 1,177 answers, averaging 5.19 hotels per answer. November contained 2,342 appearances across 400 answers, averaging 5.86, about 13% more.
Exhibit 5
Mean identified hotels per assessed answer, by engine and collection date
| Engine | September mean hotels per answer | November mean hotels per answer |
|---|---|---|
| ChatGPT | 6.82 | 7.52 |
| Google AI Mode | 4.83 | 4.89 |
| Gemini | 4.87 | 6.80 |
| Copilot | 4.28 | 4.21 |
| All engines | 5.19 | 5.86 |
Source: AGR’s reviewed answer–property ledgers. September: 1,177 assessed answers. November: 400. A hotel is counted at most once per answer.
Exhibit 6
Identified list length increased for ChatGPT and Gemini
On smaller screens, swipe or scroll horizontally to view the full chart.
Source: the per-engine means in Exhibit 5. Bars begin at zero; the chart displays observed means, not a causal estimate.
Means use all assessed answers for each date. The same broad list-length pattern was present when the calculation was restricted to the 379 groups with complete September coverage.
In the main matched-reference comparison, the 2 engines with clearly longer identified lists, ChatGPT and Gemini, also had higher November nonmatch rates than in September's repetition 3. Google AI Mode's mean list length was similar across dates and its nonmatch rate changed little. Copilot's mean was slightly lower and its nonmatch rate was lower. This is an engine-level association; it does not establish that list length caused the differences. Each additional hotel can create another membership difference under an exact-set test.
These figures describe the identified entries in the reviewed records. The extraction workflows differed between editions, and an extra or missed entry can alter both list length and exact-match status. The averages therefore cannot establish how much of the difference came from answer generation or from measurement, nor do they explain Copilot's pattern on their own.
How the answers were assessed
The fixed matrix contains 10 questions for each of 10 destinations. 8 question types are shared: general discovery, luxury, couples, families, boutique hotels, spas, dining and service. 2 further questions address local travel needs. The wording is oriented toward upscale travel.
A qualifying appearance counts an identifiable hotel featured in an answer, at most once per answer. Clear name variants are consolidated. Separately named hotels remain distinct, and explicit joint recommendations credit each identifiable hotel once. A clearly identified hotel's spa or relevant on-site amenity contributes to its parent hotel's visibility. Names found only in citations, ancillary cards or incidental prose are excluded, and ambiguous identities are withheld. The follow-up reused September's publication conventions and name decisions, with AI-assisted extraction followed by review of flagged identities and supporting excerpts.
The comparison is observational. November's single pass cannot establish lasting gains or losses, and small ranking margins may change in another collection. Exact model versions, session conditions and internal retrieval processes were not uniformly controlled. The results do not isolate elapsed time as the cause of a difference, rank the quality of the engines' advice, verify hotel operations or amenities, or measure bookings. They describe the specified questions and saved answers under the stated counting rules.
What held up across the 2 collections
7 destination leaders remained first in their respective collections. Several hotels also retained substantial appearance rates, even where their positions changed.
Complete recommendation lists were much less reproducible. September's own repetitions showed that this variation was already common within a single collection date. The matched-reference check places the 5-week result alongside that existing variability, while the list-length analysis shows another reason to interpret exact matches carefully.
For hotel visibility research, repeated presence and exact-list reproduction answer different questions. A hotel can remain frequently recommended while the surrounding competitive set changes. Reporting its appearance count, denominator, engine and collection date makes that persistence visible without requiring every answer to contain the same hotels.
Check your own hotel. To see how ChatGPT, Gemini and Google AI Mode currently represent your property, run AGR’s real-time AI visibility report for your hotel.
Related AGR research
The September 10-destination study establishes the baseline used here. It assessed 1,177 substantive answers and recorded 6,110 featured hotel appearances.
The AGR Luxury Hotel AI Visibility Index examines a separate July 29, 2026 6-market collection. Its 824 recommendation slots should not be pooled with this follow-up or treated as the same measurement.
Publisher disclosure. AGR provides hotel marketing and AI visibility services.
How to cite this research
Andrew Paul, Americas Great Resorts. Does AI Recommend the Same Hotels Five Weeks Later? 2026. Fieldwork: September 27 and November 2, 2026.
Suggested short citation: “In AGR’s November 2, 2026 follow-up across 10 destinations and 4 AI engines, 7 destination leaders remained first in their respective collections.”
Alternative short citation: “Matched hotel-set comparisons recorded substantial variation within September as well as between September and November; paired nonmatch gaps ranged from 3.2 to 6.9 percentage points.”
AGR retains the capture records and comparison notes for research and media inquiries. Contact: info@americasgreatresorts.net.
