AGR Research | Luxury Traveler AI Survey | November 2026
AI is entering the luxury hotel decision process, but it does not hold the same position at every stage. Among 2,768 survey participants who answered whether they had booked a hotel first heard about through AI, 1,530 (55.3%) said yes. When asked where they check an AI-recommended hotel, 2,130 of the 3,562 survey submissions selected the hotel’s own website. Yet only 285 of the 3,562 selected an AI assistant as one of their two most trusted sources for a final luxury hotel decision. These are separate aggregate measures, not a traced customer journey.
Americas Great Resorts conducted a voluntary ten-question survey of its frequent luxury-traveler audience. On November 2, 2026, we emailed invitations to 525,000 selected records from our proprietary high-net-worth frequent-traveler segment. Responses closed on November 19, 2026, at 5 p.m. Eastern. We received 3,562 survey submissions, each containing at least one answer; none of the ten individual questions was mandatory. The questionnaire did not measure respondents’ income or net worth. The results reflect aggregate selections and self-reported behaviors, not verified bookings or estimates representative of all luxury travelers.
Basis: AGR voluntary email survey, November 2–19, 2026. The four statistics use their expressly stated question-specific or full-submission denominators. These are separate aggregate findings, not linked individual journeys.
Quick Answers
Have surveyed travelers booked a hotel they first heard about from AI?
Among 2,768 respondents who selected a Q5 answer, 1,530 (55.3%) said they had done so in the past 12 months. Another 884 (31.9%) said no, and 354 (12.8%) were not sure. The 794 submissions that left Q5 blank are excluded from those Q5-answerer percentages. The finding reports remembered discovery and booking, not independently verified referrals or incremental bookings.
Which AI assistant is most frequently selected for travel planning?
ChatGPT received 1,909 selections, equivalent to 53.6% of all submitted surveys. Google AI Mode or AI Overviews received 1,633 (45.8%) and Gemini 898 (25.2%). Q3 allowed multiple selections, so these are not exclusive market shares and cannot be summed into a unique-user count.
Where do respondents say they check an AI hotel recommendation?
The hotel’s own website received 2,130 selections (59.8% of all submissions), followed by Google reviews with 1,833 (51.5%) and Tripadvisor with 1,057 (29.7%). Q6 allowed multiple selections and was optional. These totals describe selected verification sources, not actual observed visits or the sequence of checks.
What do respondents trust for the final luxury hotel decision?
In Q8, the hotel’s own website received 1,224 selections (34.4% of all submissions), Google reviews 1,112 (31.2%) and friends or family 997 (28.0%). An AI assistant received 285 (8.0%). Respondents could select up to two sources. The data do not establish a single top source per individual or whether any specific respondent used AI earlier in the same booking.
Would respondents allow AI to book and pay on their behalf?
Of the 2,719 who answered Q9, 1,631 (60.0%) chose ‘No.’ Another 680 (25.0%) would let AI hold a room but would pay themselves. The remaining 408 (15.0%) selected some willingness to let AI book and pay, either for any hotel or only one they had previously stayed at. These are stated preferences, not experiments in autonomous booking.
Scope: These findings describe voluntary submissions from one selected AGR email audience. They are not population-wide estimates, and aggregate results cannot link individuals across questions.
Survey Overview: Who Was Asked and How It Was Conducted
| Survey and distribution measure | Recorded result | Interpretation |
|---|---|---|
| Invitations sent | 525,000 | Selected portion of AGR’s frequent luxury-traveler email audience |
| Deliverability | 96.2% | Email campaign deliverability metric |
| Estimated delivered invitations | 505,050 | 525,000 × 96.2%; derived, not an independently supplied delivered count |
| Unique open rate | 24.7% | Unique recorded opens; sent-versus-delivered denominator not supplied. Opens do not prove human readership |
| Survey submissions | 3,562 | Submitted forms, each answering at least one question |
| Submissions / invitations sent | 0.68% | 3,562 ÷ 525,000, an email-to-survey completion yield |
| Submissions / estimated delivered invitations | 0.71% | 3,562 ÷ 505,050; estimated email-to-survey completion yield |
| Survey launch | November 2, 2026 | Start of collection |
| Survey closure | November 19, 2026, 5 p.m. | Eastern time |
| Instrument | 10 questions | At least one response required to submit; no individual question required; four select-all, one select-up-to-two, five select-one |
We sent invitations to 525,000 selected records from our proprietary high-net-worth frequent luxury-traveler audience. Our published data and targeting description explains the wider audience-selection approach. The questionnaire did not independently measure respondents’ income, assets, or net worth. Recruitment was voluntary, not a random national sample. People who did not open the email or chose not to submit the survey may differ materially from participants.
Response terminology: The 0.71% figure is a calculated completion yield against estimated delivered invitations, not a probability-sample survey response rate or a click-through rate. The 24.7% figure is the unique email open rate from the campaign; its sent-versus-delivered denominator is not specified in the retained research figures. Email tracking does not independently establish human readership. Survey-link click counts were not supplied.
Key limitations: Voluntary, nonprobability recruitment; income and net worth not verified among respondents; all individual questions optional; no respondent-level records; reported behavior and errors not independently observed. Results describe this participating audience only.
At a glance
- Reported discovery can precede booking. Of those answering Q5, 55.3% said they had booked a hotel they first heard about from an AI assistant. This is respondent recall, not transaction-level attribution.
- AI is selected more often for discovery than for making the booking. Q4 recorded 1,959 selections for finding hotels to consider and 178 for making the booking. This does not establish a conversion funnel because the selections cannot be matched to individuals.
- Hotel websites remain prominent in checking and deciding. The hotel’s own website was selected by 2,130 respondents in Q6 and by 1,224 in Q8, two differently worded questions with different response limits.
- Respondents report AI errors. Wrong prices received 1,072 Q7 selections, and wrong amenities or restaurants received 908. These are unverified reports; overlapping selection is possible.
- Payment authority remains a separate issue. Among Q9 answerers, 60.0% would not let AI book and pay, while 25.0% would permit a hold but pay personally.
What this study measures
The study records what participating travelers selected about their leisure hotel experience, AI assistants used, planning activities, self-reported AI-originated hotel bookings, verification behavior, inaccurate information, sources trusted for the final decision, willingness to delegate booking/payment, and the channel used for their most recent luxury hotel reservation.
A selection is not necessarily a respondent-level trajectory. The final spreadsheet preserves only aggregate counts by answer. It does not preserve which respondent selected which answers. Cross-question statements cannot claim that travelers using AI for hotel discovery were the same individuals who consulted hotel websites, trusted one source, or booked through any particular channel.
None of the ten individual questions was mandatory; however, at least one response was required to submit the survey. For five single-answer questions (Q1, Q2, Q5, Q9, Q10), we can calculate the number that supplied no response from the reported total of 3,562. For multiple-answer questions we cannot reconstruct the number of unique respondents who answered, because a person could select more than one option. In those questions, percentages are consistently calculated as selections divided by all 3,562 submissions. These are shares of the survey-submission cohort selecting an option, not question-answerer shares.

The respondents: leisure hotel nights and nightly rates
Exhibit 1Leisure hotel nights during the past 12 months (Q1; select one)
| Response | Selections | % of those answering |
|---|---|---|
| None | 31 | 1.0% |
| 1 to 5 | 458 | 15.0% |
| 6 to 14 | 1,069 | 35.0% |
| 15 to 30 | 1,008 | 33.0% |
| More than 30 | 489 | 16.0% |
Answered Q1: 3,055 of 3,562; unanswered: 507. Percentages in Exhibit 1 use Q1 answerers.
Exhibit 2Typical nightly leisure hotel rate (Q2; select one)
| Response | Selections | % of those answering |
|---|---|---|
| Under $300 | 133 | 5.0% |
| $300 to $599 | 399 | 15.0% |
| $600 to $999 | 799 | 30.0% |
| $1,000 to $1,999 | 932 | 35.0% |
| $2,000 or more | 399 | 15.0% |
Answered Q2: 2,662 of 3,562; unanswered: 900. Self-reported typical rates are not independently validated.
Of Q2 answerers, 2,130 of 2,662 (80.0%) selected $600 or more. This shows a concentration in premium room-rate categories among those answering, but the survey did not measure income or net worth.
AI assistants and their role in trip planning
Exhibit 3AI assistants used for travel planning in the past 12 months (Q3; select all that apply)
| Response | Selections | % of all submissions |
|---|---|---|
| ChatGPT | 1,909 | 53.6% |
| Gemini | 898 | 25.2% |
| Google AI Mode or AI Overviews | 1,633 | 45.8% |
| Claude | 171 | 4.8% |
| Perplexity | 208 | 5.8% |
| Copilot | 214 | 6.0% |
| Another AI assistant | 116 | 3.3% |
| None | 212 | 6.0% |
Exhibit basis: optional question; multiple selections permitted; each displayed percentage is selections divided by all 3,562 submissions, not the unknown number who answered this question.
Exhibit 4Where AI was used during trip planning (Q4; select all that apply)
| Response | Selections | % of all submissions |
|---|---|---|
| Choosing a destination | 1,598 | 44.9% |
| Finding hotels to consider | 1,959 | 55.0% |
| Comparing a shortlist | 1,077 | 30.2% |
| Checking hotel details | 1,060 | 29.8% |
| Checking prices | 1,241 | 34.8% |
| Making the booking | 178 | 5.0% |
| Did not use AI | 521 | 14.6% |
Exhibit basis: optional question; multiple selections permitted; each displayed percentage is selections divided by all 3,562 submissions, not the unknown number who answered this question.
Finding hotels to consider received about 11 times as many selections as making the booking: 1,959 versus 178. This compares answer-option selection totals, not individual travelers or conversion rates. The count of 521 selecting “Did not use AI” in Q4 should not be forced to match the 212 selecting “None” in Q3. Those options answer different questions, and the aggregate dataset cannot explain the difference.
Booking a hotel first heard about through AI
Exhibit 5In the past 12 months, did you book a hotel you first heard about from an AI assistant? (Q5; select one)
| Response | Selections | % of those answering |
|---|---|---|
| Yes | 1,530 | 55.3% |
| No | 884 | 31.9% |
| Not sure | 354 | 12.8% |
Answered Q5: 2,768; unanswered: 794. The “Yes” share is 55.3% among Q5 answerers or 43.0% across all submitted surveys.
The 1,530 “Yes” selections are a central self-reported finding. In the past 12 months, these respondents said they booked a hotel they first heard about through an AI assistant. The survey does not establish that AI caused the booking decision. They do not reveal whether the booking occurred on a hotel website, with an OTA, by telephone, through a travel advisor, or within an AI assistant. The survey cannot attribute incremental demand or measure how often an AI introduction was decisive.
Checking a hotel after an AI recommendation
Exhibit 6Where travelers check before booking when AI recommends a hotel (Q6; select all that apply)
| Response | Selections | % of all submissions |
|---|---|---|
| Google reviews | 1,833 | 51.5% |
| Booking.com | 702 | 19.7% |
| Expedia or Hotels.com | 435 | 12.2% |
| Tripadvisor | 1,057 | 29.7% |
| Hotel’s own website | 2,130 | 59.8% |
| Travel advisor | 358 | 10.1% |
| Forbes Travel Guide or Michelin | 289 | 8.1% |
| Don’t check | 178 | 5.0% |
Exhibit basis: optional question; multiple selections permitted; each displayed percentage is selections divided by all 3,562 submissions, not the unknown number who answered this question.
The property’s own website ranked first by selections in this question. That matters because the survey distinguishes an AI system suggesting a hotel from a source a respondent reports checking. Nevertheless, respondents were not observed visiting hotel websites, and no individual-level linkage can show whether a website check followed a particular AI recommendation. The 178 “Don’t check” selections are retained as recorded; contradictory individual selections, if any, cannot be audited from totals.
Incorrect hotel information reported by respondents
Exhibit 7Has an AI assistant ever given you wrong information about a hotel? (Q7; select all that apply)
| Response | Selections | % of all submissions |
|---|---|---|
| Closed or under renovation | 359 | 10.1% |
| Wrong price | 1,072 | 30.1% |
| Wrong amenities or restaurants | 908 | 25.5% |
| Hotel did not exist | 97 | 2.7% |
| No, never | 1,008 | 28.3% |
| Not sure | 859 | 24.1% |
Exhibit basis: optional question; multiple selections permitted; each displayed percentage is selections divided by all 3,562 submissions, not the unknown number who answered this question.
Wrong prices and wrong amenities or restaurants were the most frequently selected specific problems. The survey did not test model accuracy against hotel records, ask respondents to identify the AI engine responsible, or capture screenshots or dates of the disputed answers. It also cannot determine how many unique people experienced any inaccuracy, because several problem types could have been selected by one person. “No, never” and “Not sure” were options in a multi-select question; no individual-level validation was preserved to test for contradictory selections.
Which sources matter most for the final luxury hotel decision?
Exhibit 8Sources trusted most for the final decision (Q8; select up to two)
| Response | Selections | % of all submissions |
|---|---|---|
| An AI assistant | 285 | 8.0% |
| Google reviews | 1,112 | 31.2% |
| Reviews on Booking.com | 435 | 12.2% |
| Reviews on Expedia or Hotels.com | 272 | 7.6% |
| Tripadvisor | 400 | 11.2% |
| Hotel’s own website | 1,224 | 34.4% |
| Travel advisor | 530 | 14.9% |
| Forbes Travel Guide or Michelin | 427 | 12.0% |
| Condé Nast Traveler or Travel + Leisure | 288 | 8.1% |
| Friends or family | 997 | 28.0% |
| Social media | 140 | 3.9% |
Exhibit basis: optional question; up to two selections allowed; each displayed percentage is selections divided by all 3,562 submissions, not the unknown number who answered Q8.
The hotel’s own website led by selections, followed by Google reviews and friends or family. AI assistants drew 285 selections, compared with 1,224 for the hotel website. The question asked for up to two trusted sources, not for a complete ranking, so the results do not show which one was each respondent’s first choice. Q8 recorded 6,110 selections. The form allowed up to two choices, giving a theoretical maximum of 7,124 selections across 3,562 submissions. The aggregate total falls below that ceiling but alone cannot verify each individual selection count.
Would travelers let AI book and pay?
Exhibit 9Permission to let an AI assistant book and pay for a hotel (Q9; select one)
| Response | Selections | % of those answering |
|---|---|---|
| Yes, any hotel | 136 | 5.0% |
| Only a hotel I’ve stayed at before | 272 | 10.0% |
| Hold a room, but I pay | 680 | 25.0% |
| No | 1,631 | 60.0% |
Answered Q9: 2,719; unanswered: 843. Percentages use Q9 answerers.
The distinction between holding a room and paying for it is important. A quarter of Q9 answerers chose “Hold a room, but I pay,” indicating willingness to delegate a limited reservation step while retaining payment authority. The question describes hypothetical willingness; the data do not establish actual adoption, merchant availability, fraud concerns, or the reasons for reluctance.
Where respondents reported booking their most recent luxury hotel stay
Exhibit 10Reported channel for most recent luxury hotel booking (Q10; select one)
| Response | Selections | % of those answering |
|---|---|---|
| Hotel website or phone | 1,253 | 52.6% |
| Travel advisor | 451 | 18.9% |
| Booking.com | 301 | 12.6% |
| Expedia or Hotels.com | 176 | 7.4% |
| Other | 150 | 6.3% |
| Inside an AI assistant | 50 | 2.1% |
Answered Q10: 2,381 of 3,562; unanswered: 1,181. Percentages use Q10 answerers and may total 99.9% due to rounding. Self-reported channel, not verified transaction data.
Of those answering Q10, 1,253 of 2,381 (52.6%) selected “Hotel website or phone.” This combines website and telephone reservations; it cannot be interpreted as exclusively online direct bookings. Fifty respondents (2.1%) selected “Inside an AI assistant.” Q10 asks about the most recent luxury hotel stay, whereas Q5 asks whether respondents booked a hotel they first heard about through an AI assistant in the past 12 months. They measure different things, and the aggregate totals cannot be combined into a booking funnel.
What this means for luxury hotels
This section is interpretation, not another measured result. The survey suggests AI can play a meaningful role in introducing hotels, while hotel-owned information and established review sources continue to receive substantial selections for verification and the final decision. At the same time, respondents reported wrong AI information, and many are unwilling to delegate payment. None of these patterns establishes causality or which channel ultimately captured demand from an AI-originated discovery.
Check the information a traveler can validate. A hotel can examine whether its own website clearly communicates room categories, rates, amenities, restaurant availability, and reservation details. The survey identifies those categories as relevant to respondents; it does not test whether changing website content improves AI recommendations.
Distinguish discovery from booking attribution. An AI introduction is not proof of a direct booking, and a direct booking is not proof of AI influence. Hotels seeking to measure these relationships need consented, respondent- or session-level attribution methods not available in this survey.
Do not mistake willingness for autonomous purchase adoption. The Q9 preference responses and Q10 reported booking channels measure different constructs. Product adoption cannot be calculated by subtracting them.
Monitor answer quality independently of traffic. Wrong prices, amenities, and operational status were reported by participants; hotels should verify their public information and test representative AI responses, without assuming this survey measured actual error rates.
Methodology
Study design and instrument. We designed a ten-question, self-administered multiple-choice survey. The respondent-facing title was “How you choose a luxury hotel now.” The instrument described the survey as taking under three minutes. Questions covered past-year leisure nights, typical leisure hotel nightly rate, AI assistants used for travel planning, AI planning activities, whether a hotel first heard about through AI was subsequently booked, verification sources, reported incorrect AI hotel information, up to two trusted final-decision sources, willingness to let AI book/pay, and channel of the most recent luxury hotel booking. The survey instrument and aggregate response spreadsheet form the supporting research materials.
Respondent-facing survey interface

Recruitment and fieldwork. On November 2, 2026, we emailed the survey to 525,000 selected records from our proprietary high-net-worth frequent luxury-traveler audience. Our data strategy describes the travel-interest, demographic, and other criteria used in audience selection. The survey itself did not verify individual respondents’ income or net worth. We accepted responses through November 19, 2026, at 5 p.m. Eastern and compiled the results that day.
Delivery and participation. The email deployment achieved 96.2% deliverability and a 24.7% unique open rate. Applying the deliverability rate to 525,000 invitations yields an estimated 505,050 delivered messages. We received 3,562 survey submissions; each contained at least one answer because the form did not permit a completely blank submission. This equates to 0.68% of invitations sent or 0.71% of estimated delivered invitations. The retained campaign figures do not specify whether the unique open rate was calculated against sent or delivered messages, and survey-link click totals are unavailable. An email open does not independently establish that a person read the invitation. These rates describe this email deployment, not a probability-sample survey response rate.
Answer design and denominator rules. No individual question required a response, but at least one answer was needed to submit the form. Q1, Q2, Q5, Q9, and Q10 allowed one answer each. Q3, Q4, Q6, and Q7 allowed multiple selections. Q8 permitted up to two. For single-answer questions, exhibits use the number of recorded answers to that question as the primary denominator, and state the number missing. For multi-select questions, exhibits report each answer count divided by all 3,562 survey submissions, because unique question-answerer counts are not available. Multiple-answer percentages can exceed 100% when added. A blank response is not recoded as “No,” “None,” or “Not sure.” All percentages are recalculated from raw counts and rounded to one decimal place.
Missing responses in single-answer questions. Q1: 3,055 answered, 507 missing. Q2: 2,662 answered, 900 missing. Q5: 2,768 answered, 794 missing. Q9: 2,719 answered, 843 missing. Q10: 2,381 answered, 1,181 missing. These missing totals are calculated by subtracting answer selections from 3,562, assuming the platform allowed at most one selection per question as shown by the instrument.
Data retention and reproducibility. We retained aggregate answer counts rather than individual response records or combinations of answers. We reviewed the tabulated counts before publication. The exhibits are reproducible from the final aggregate counts and the denominators stated alongside them. Because respondent-level records were not retained, we cannot reconstruct individual answers or perform cross-tabulations, conditional analyses, or respondent-level attribution. We did not calculate survey weights, confidence intervals, or significance tests. The aggregate workbook alone cannot independently verify the original entries held by the survey platform.
Sampling limitations. This is a voluntary, nonprobability email survey of selected AGR records. The recruitment file is commercially relevant to luxury hospitality but is not a representative probability sample of all affluent travelers, all U.S. travelers, or luxury-hotel guests. Unknown nonresponse and coverage differences prevent national generalization or conventional sampling margins of error. Demographic selection attributes were not independently verified in the survey responses supplied. Self-reported past use, bookings, inaccuracies, and future willingness are subject to recall, interpretation, and self-selection effects.
Question wording and construct limits. “Used AI for making the booking” (Q4) does not specify whether AI executed payment. “Booked a hotel first heard about from an AI assistant” (Q5) does not identify the ultimate booking channel. “Where do you check” (Q6) is a stated behavior, not tracked visits. Q7 errors are respondents’ reports rather than verified inaccuracies. Q8 permits up to two choices and does not rank them. Q9 is hypothetical willingness. Q10 “Hotel website or phone” combines two direct booking routes, and Q10 concerns the most recent booking rather than Q5’s past-12-month discovery question. Questions concern different periods or conditions and cannot be pooled as a funnel.
Potential instrument ambiguity. “None,” “Did not use AI,” “No, never,” and “Don’t check” were offered alongside other responses within select-all questions, without individual-level data to verify logical exclusivity. Q3’s “None” and Q4’s “Did not use AI” are not equivalent constructs. These limitations do not invalidate the aggregate counts, but constrain interpretation.
Data availability. The research materials consist of the ten-question instrument, the aggregate response counts, and methodology notes. Respondent-level submissions were not retained in the research files; individual-level cross-question analysis and independent reconstruction of the original submissions are not possible.
Prior and related work
This survey adds travelers’ self-reported use of AI to AGR’s earlier measurements of which luxury hotels AI systems recommend, how those recommendations change, and where the accompanying links lead. The studies examine different units of observation and were conducted under different protocols. Their counts are not pooled, and none establishes that an AI recommendation caused an individual booking.
The AGR Luxury Hotel AI Visibility Index (July 29, 2026). Our six-market study recorded 824 ranked hotel recommendation slots in 180 AI answers. It examined how recommendations were distributed among luxury hotels, before any traveler decision or booking could be observed. The present survey addresses a different question: how participating travelers say they use AI when considering hotels.
Which Hotels Do AI Engines Recommend? A Study of 10 Destinations (September 27, 2026). This study assessed 1,177 substantive answers across ten destinations and four AI engines, recording 6,110 featured hotel appearances. It documented variation by engine and travel question. The traveler survey complements those measured recommendations with reported discovery, verification and booking behavior.
Does AI Recommend the Same Hotels Five Weeks Later? (November 2, 2026). Repeating the September question set showed that seven of ten destinations retained the same leading hotel while individual hotel lists changed. That is evidence about AI answer variation, not a measure of how travelers reacted to those recommendations.
The AGR AI Hotel Link Study: Where 900 AI Answers Send Travelers (November 11, 2026). The link study recorded where citations and links in 900 AI hotel answers pointed. Among 899 answers naming a hotel, 536 contained no hotel or hotel-brand website link. The new survey separately records where participants say they check recommendations and where they report booking. No individual traveler path connects the two datasets.
Ask AI the Same Question in Three Languages. Get Three Different Answers. (November 17, 2026). This Miami study examined recommendations across English, Spanish and Portuguese questions and different country origins. It found variation even when questions were repeated under the same conditions. The survey does not independently measure how language or location affected respondents’ use of AI.
AGR’s AI Visibility, KFO & Hospitality AI Resource Index organizes additional original research by subject. Our data and targeting strategy describes the travel-data audience used for survey recruitment. That audience description is sampling background, not evidence that voluntary respondents represent all luxury travelers.
How to cite this study
Suggested citation: Andrew Paul, Americas Great Resorts. How Travelers Use AI to Discover, Evaluate and Book Luxury Hotels: AGR 2026 Survey. Fieldwork November 2–19, 2026. Aggregate survey of 3,562 voluntary survey submissions. Published November 2026.
Suggested short citation: “In AGR’s voluntary November 2026 traveler survey, 1,530 of 2,768 respondents answering Q5 (55.3%) reported that, in the past 12 months, they booked a hotel they first heard about from an AI assistant.”
Alternative short citation: “AGR’s November 2026 survey recorded 2,130 selections of the hotel’s own website as a place to check an AI hotel recommendation, out of 3,562 survey submissions.”
Research materials: Ten-question survey instrument, aggregate response workbook, denominator definitions and calculation notes. Individual-level response records were not retained.
Research disclosure. The study draws on a selected proprietary email audience and voluntary submissions. Results should be cited with fieldwork dates, base counts and specific questions. No claim of national representativeness is made.

