The state of AI search in travel

Travellers now plan trips with AI but still book with brands they trust, and AI answers lean on OTAs, review sites and blogs rather than brands' own pages. Citations fade within weeks, most hotel websites we audited are hard for AI to read, and the first personal agents have started booking. The work for travel brands now is visibility: be readable, cited and accurate, and run bookability as a separate track.

Published: 3 October 2026
Data checked to: 1 October 2026
Type: Evidence review of 2026 studies, with an AuraScope audit and survey
Written for: Airlines, accommodation and travel operators
Also as: Markdown JSON data

22.6%: of ChatGPT travel answers cite a web source. The rest come from the model's memory.; Source: Similarweb, May 2026, US desktop. 2
68%: of travellers prefer to book with a trusted travel brand over an AI chatbot.; Source: Expedia Group and YouGov, 5,700 adults, March 2026. 3
<1%: of Booking Holdings room nights come from LLM referrals, with "no material change".; Source: Booking Holdings Q2 2026 call, 4 August 2026. 4
3.4 weeks: is the half-life of a ChatGPT citation.; Source: Scrunch and Stacker, 3.5 million citation events, September 2025 to March 2026. 6

Contents

Key findings

AI is now a normal place to plan a trip but not to book one: 53% of travellers are comfortable with AI suggestions, and 8% with AI booking.
At NZ Tech Expo (n=89, a tech-heavy room, not a general traveller sample), 88.8% had used AI to plan a trip and 1 of 79 users had let it book.
The average brand appears in 16.30% of AI answers about its category, and 84% of answers cite someone else's site.
Which sources AI trusts depends on the engine and the question. On hotel questions, OTAs take 55.3% of citations; on open travel questions, independent blogs dominate.
Citations fade fast. A ChatGPT citation has a half-life of 3.4 weeks.
79.4% of the 107 New Zealand hotel websites AuraScope evaluated fully scored Weak or Critical on the signals AI uses to read a site, and airlines go unlinked in 74.6% of ChatGPT flight mentions.
The answer engines held back from checkout, while personal agents from Meta, OpenAI and xAI have started booking. Muse shops for hotels by browsing OTA sites.

Each numbered box links to the source list. Figures tagged Vendor datasetcome from AthenaHQ's pooled dataset, which nobody else has corroborated. Secondary coveragemarks trade coverage or a company's own unaudited statement, and AuraScope researchmarks AuraScope's own measurement. Everything else was checked against its primary source.

Part 1: Where trips start

A trip now starts with a question, not a search. This part shows how far AI planning has spread, and where it stops: at the booking.

Answers are replacing clicks

Search engines and assistants now answer the question themselves. Search referrals to publishers fell 22% to 60% over two years, and in travel the AI referrals that replace them convert 28% worse.

Traditional search against AI search
Traditional search	AI search	Source
People clicked a link and found the answer on a publisher's site.	The answer sits on the results page. 68% of Google searches now end without a click.	21
Search sent sites a steady flow of visitors.	Search referrals fell 22% for large publishers and 60% for small ones over two years. Google Search referrals to publishers fell 34% in 2025.	20
Visitors arrived early in their research.	In May 2026, AI referrals to US travel sites stayed 70% longer and bounced 41% less than other traffic, and converted 28% worse.	19

Source: Chartbeat via Axios, 17 March 2026 (December 2023 to December 2025); Growth Memo H1 2026 report; Adobe Digital Insights, 17 June 2026.

The traveller asks, reads the answer and often clicks nothing. That breaks two assumptions behind search optimisation: that a ranking produces a visit, and that a visit produces a conversion.

That changes what to measure: whether your brand is named in the answer, and whether your site is one of the sources behind it.

Planning a trip with AI is now normal

ChatGPT alone has more than a billion weekly users, and AI-sourced traffic to US travel sites nearly tripled in the year to May 2026.

1 billion+: weekly ChatGPT users; Source: OpenAI, 31 August 2026. 28
62%: of audiences in Tourism New Zealand's key markets are likely to use AI tools to plan their next trip. 53% expect to use AI more.; Source: Tourism New Zealand active-considerer research, 27 July 2026. 47
13% to 30%: of travellers used ChatGPT extensively to plan trips, a year apart: up from 13% to 30%; Source: Skift Research and McKinsey, reported 20 March 2026. 25
44.5 million: AI referral visits a month to travel sites worldwide, up 115.6% year on year. These are visits, not bookings.; Source: Similarweb, June 2025 to May 2026. 2
40%: of travellers use AI to help build itineraries. 59% still use search engines, and 8% of US and UK travellers rely on AI chatbots and agents.; Source: Expedia Group and YouGov, March 2026. 3
194%: growth in AI-sourced traffic to US travel sites in the year to May 2026. Since October 2024 it has grown 2,215%.; Source: Adobe Digital Insights, 17 June 2026. 19

It is already live in New Zealand

Google launched AI Mode in New Zealand in August 2025, and Expedia's ChatGPT app is available to New Zealand users. Tourism New Zealand ran AI discoverability workshops in five centres in July 2026, and MBIE has put $800,000 of levy funding into making Tourism New Zealand's operator database and AI travel assistant machine-readable.

Sources: 33 44 47 48

Booking is where travellers still hold back

53% of travellers are happy for AI to suggest a trip, but only 8% are comfortable booking through it. The booking platforms see the same: AI referrals are growing from a small base.

53%: are comfortable letting AI suggest travel options; Source: Expedia Group and YouGov, 5,700 adults in the US, UK and India, 10 to 25 March 2026. 3
68%: prefer to book with a trusted travel brand over AI chatbots and agents; Source: Expedia Group and YouGov, March 2026. 3
8%: are comfortable booking through an AI platform. 66% would not trust an AI assistant to buy or book anything for them.; Source: Expedia Group and YouGov, March 2026. 3
2%: would give an AI tool full autonomy to make and change bookings; Source: McKinsey and Skift Research, 1,002 travellers, September 2025. 24

What the booking platforms say about AI referrals
Company	Statement	Date	Source
Booking Holdings	LLM referrals are "less than 1% of total room nights", with "no material change over the last few months or quarters". Direct mix is flat in the mid-60s. Marketing spend is up 11%.	Q2 2026 call, 4 August 2026	4
Expedia Group	Answer-engine optimisation is "still a small channel but one of the fastest growing". Agentic traffic shows "higher purchase consideration".	Q2 2026 call, 5 August 2026	5
Airbnb	Will launch its own agent in 2027. Stays out of ChatGPT apps and treats assistants as a source of traffic.	Skift Global Forum, 23 September 2026	42

Source: PhocusWire earnings coverage, 4 and 5 August 2026; Skift, 23 September 2026.

These surveys record what people say, not what they do, and none of them samples New Zealand or Australian travellers. AuraScope's own survey, in the next section, is a small New Zealand sample from a tech-heavy room. No platform or airline has published a count of bookings completed by an AI agent.

So the planning stage is where to work now. That is where AI decides which brands to mention and which sources to trust.

At NZ Tech Expo, AI planned the trip and booking sites took the booking

NZ Tech Expo attendees (n=89), not a general traveller sample. 88.8% had used an AI assistant to plan a trip, but only 1 of 79 users had let one book. Most would let AI book and pay if they can approve the choice and see the full price first.

88.8%: have used ChatGPT, Gemini or Copilot to plan a trip. 16.9% say it is now their starting point.; Source: NZ Tech Expo attendees (n=89), all respondents. 14 AuraScope research
61.6%: used an AI assistant while planning their last stay. 17.4% opened one first, ahead of Google or a booking site.; Source: NZ Tech Expo attendees (n=89), the 86 who had taken a trip. 14 AuraScope research
19.8%: booked their last stay direct with the property. At the same price, 45.3% would rather book direct.; Source: NZ Tech Expo attendees (n=89), the 86 who had taken a trip. 14 AuraScope research
86.5%: would let AI book and pay on some terms: 50.6% only if they approve first, 9.0% for stays under $500 and 27.0% without hesitation.; Source: NZ Tech Expo attendees (n=89), all respondents. 14 AuraScope research

Where they booked their last stay
Booked through	Share of respondents with a trip
A booking site	54.7% (47)
Direct with the property	19.8% (17)
Airbnb or similar	17.4% (15)
Someone else booked it	5.8% (5)
Don't remember	2.3% (2)

Source: NZ Tech Expo attendees (n=89), the 86 who had taken a trip. 14 AuraScope research

What would matter most before letting AI book
Condition	Share of respondents
See the exact total price before it pays	40.4% (36)
Adjust or approve its choice before it books	38.2% (34)
A confirmation from the property itself	19.1% (17)
Nothing would, I wouldn't let it	2.2% (2)

Source: NZ Tech Expo attendees (n=89), all respondents. 14 AuraScope research

This room used AI to plan and booking sites to buy. Of the 79 who had used an AI assistant for a trip, 65 mainly used it for ideas or to build an itinerary, and one had let it book. Those who booked through a booking site most often said it had everything in one place (37.1%) or was cheaper (32.3%).

The direct-booking gap shows up even here. 45.3% would rather book direct when the price is the same, but only 19.8% did, and of the 39 who prefer direct, 10 booked that way.

Asked about the future, most would let AI book, on their own terms. That is a much warmer answer than Expedia's 8% or McKinsey's 2%, but the questions differ, and 70.8% of this room works in technology or AI, counting the 7.9% who work in both travel and tech. Read it as where tech-literate people are heading, not as the market. Part 4 comes back to this: an agent that shows the full price and waits for approval is what this room asked for.

Sources: 14 3 24

Part 2: Inside the answer

When AI answers a travel question, it names a few brands and leans on other people's sources. This part shows who gets named, which sources get cited and what kind of content they are.

Most brands are missing from most answers

Category leaders appear in AI answers three and a half times as often as the average brand. Even when a brand appears, the answer usually cites someone else's site.

16.30%: average brand mention rate, against 56.48% for the leading brands; Source: AthenaHQ, State of AI Search 2026. 1 Vendor dataset
84%: of responses cite a third party rather than the brand's own domain. 16.05% cite the brand's domain.; Source: AthenaHQ, State of AI Search 2026. 1 Vendor dataset
6 to 27: distinct domains cited per answer, depending on the engine. The average across engines is 12.; Source: AthenaHQ, State of AI Search 2026. 1 Vendor dataset

Share of voice by brand rank, all segments
Brand rank in category	Share of all mentions
Rank 1	33.64%
Rank 2	19.46%
Rank 3	13.29%
Everyone else	33.61%

The top three brands hold 66.39% of mentions. The top brand is mentioned 3.5 times as often as the average brand in its category.

Source: AthenaHQ, State of AI Search 2026. 1 Vendor dataset

Distinct domains cited per answer, by engine
Engine	Mean domains per answer
Grok	26.99
ChatGPT	18.86
Google AI Mode	11.16
Google AI Overviews	8.74
Perplexity	8.23
Gemini	6.14
Copilot	5.77

AthenaHQ's chart names six engines. The seventh value, 5.77, carries only an icon, so it is assigned to Copilot as the one engine left. Independent 2026 counts differ: about 26 citations per Google AI Mode travel answer (Search Agency, July 2026) and 15.22 sources per AI Overview (SE Ranking, February 2026).

Source: AthenaHQ, State of AI Search 2026. 1 17 18 Vendor dataset

The engines also read very differently. The widest-reading engine cites more than four times as many domains per answer as the narrowest, so a source plan built for one engine will not carry over to the next.

The sources AI trusts change by engine

No single list of sources holds across engines. Reddit is the clearest case: it leads on Google's AI surfaces, and its weight on ChatGPT swung sharply in August 2026.

21.85%: of citations across every category come from reddit.com, pooled across seven engines; Source: AthenaHQ, State of AI Search 2026. 1 Vendor dataset
0: of the ten most cited sources is a brand's own site; Source: AthenaHQ, State of AI Search 2026. 1 Vendor dataset
73%: fall in ChatGPT's daily Reddit citations over two days in August 2026. Other engines held flat.; Source: Otterly, 27 August 2026. 8
76%: of Gemini answers carried a citation in March 2026, down from 99% in February; Source: Seer Interactive, 82,000 responses. 10

Ten most cited domains, all segments, seven engines pooled
Domain	Share of citations
reddit.com	21.85%
youtube.com	10.32%
en.wikipedia.org	7.34%
linkedin.com	6.51%
forbes.com	5.26%
facebook.com	3.99%
medium.com	3.54%
nerdwallet.com	3.44%
pmc.ncbi.nlm.nih.gov	2.55%
amazon.com	2.53%

Community platforms, encyclopaedias and a few publishers. None is a brand's own site.

Source: AthenaHQ, State of AI Search 2026. 1 Vendor dataset

Independent 2026 measurements, engine by engine
Engine	Finding	Study	Source
ChatGPT	Across all US queries, not only travel: reddit.com 16.8% of mentions, en.wikipedia.org 7.0%, youtube.com 2.5%. tripadvisor.com is 21st at 1.5%. Booking.com and Expedia are outside the top 50.	Ahrefs Brand Radar, US, updated 2 September 2026	9
Google AI Mode, travel prompts	Independent travel blogs 56.8%, OTAs and booking platforms 11.1%, video 9.7%, forums 7.3%, social 5.1%, Google properties 4.5%, editorial 4.2%, official tourism 1.2%. About 26 citations per answer.	Search Agency, 100 unbranded travel prompts, 9 July 2026	17
Gemini	Answers with any citation fell from 99% to 76% in a fortnight. Within cited answers, Reddit 44% and Wikipedia 33%. YouTube fell from 18% to 3%. Best-of listicles lost 40%.	Seer Interactive, 16 February to 2 March 2026	10
Four platforms, travel	Tripadvisor is the most cited travel source and Reddit ranks fifth or sixth. Reddit is over-represented on AI Overviews and AI Mode compared with ChatGPT and Gemini.	Scrunch, December 2025 prompts, published 5 February 2026	7

Source: Studies named per row.

Reddit tops the pooled table because it turns up in every industry and because Google's AI Overviews and AI Mode lean on it. On ChatGPT its share has swung twice in a year, so a single Reddit figure is a weak basis for a content plan.

Gemini cites less often, and when it does it prefers reference sources to social posts and listicles. It rewards pages on your own site that explain things properly.

These rankings move within months, which is why a one-off audit dates quickly.

In travel, OTAs and review sites hold 11% to 55% of citations

How much the OTAs and review sites hold depends on what the traveller asks. Hotel questions send most citations to OTAs. Open travel questions send most to independent blogs.

32.30%: of pooled travel citations come from Tripadvisor (tripadvisor.com and tripadvisor.co.uk), Booking.com, Expedia and Kayak; Source: AthenaHQ, State of AI Search 2026, travel segment. 1 Vendor dataset
55.3%: of citations for hotel prompts across ChatGPT, Perplexity and Gemini went to OTAs. Hotel websites took 13.6%.; Source: Cloudbeds, 810 prompts, six destinations, 145 properties. 12 Vendor dataset
11% to 55%: of citations go to OTAs, metasearch and review sites across 2026 studies, depending on engine and question; Source: Search Agency (AI Mode travel guides), Sitter (AI Mode hotels) and Cloudbeds (hotel prompts on three engines). 17 11 12
21.2%: of Google AI Mode hotel citations went to hotels' own websites; Source: Sitter, 4,000 queries across 8 cities, February 2026. 11

Ten most cited domains, Travel and Hospitality, seven engines pooled
Domain	Share of citations
reddit.com	21.97%
facebook.com	11.58%
tripadvisor.com	11.38%
en.wikipedia.org	9.43%
youtube.com	8.39%
booking.com	7.13%
expedia.com	6.58%
instagram.com	4.01%
kayak.com	3.65%
tripadvisor.co.uk	3.56%

Six of the market's ten most cited domains are replaced in travel. LinkedIn and Forbes drop out. Tripadvisor, Booking.com, Expedia and Kayak come in.

Source: AthenaHQ, State of AI Search 2026, travel segment. 1 Vendor dataset

Where AI enters a travel brand's own site
Path	Share of AI entries
/products	46.08%
/blog	15.29%
/home	14.50%
Other paths	24.13%

Source: AthenaHQ, State of AI Search 2026, travel segment. 1 Vendor dataset

Intermediary share of citations, by study
Study	Intermediaries	Suppliers and others	Source
ChatGPT, Perplexity and Gemini, hotel prompts (Cloudbeds)	OTAs 55.3%.	Hotel websites 13.6%. 72.4% of recommended properties were affiliated with a hotel brand or group.	12 Vendor dataset
Google AI Mode, hotel queries, February 2026 (Sitter)	Metasearch and reviews 29.2%, OTAs 17.3%. tripadvisor.com 16.5%, booking.com 3.9%, expedia.com 2.7%.	Hotel websites 21.2%, Google properties 19.9%, editorial 10.9%, user-generated 1.4%.	11
Google AI Mode, travel guides, July 2026 (Search Agency)	OTAs and booking platforms 11.1%.	Independent blogs 56.8%, video 9.7%, forums 7.3%, social 5.1%, official tourism 1.2%.	17
ChatGPT, all US queries, not only travel, September 2026 (Ahrefs)	tripadvisor.com 1.5%, ranked 21st. Booking.com and Expedia outside the top 50.	reddit.com 16.8%, Wikipedia 7.0%, YouTube 2.5%.	9

Source: Studies named per row.

The question changes the answer as much as the engine does. When travellers ask about hotels, OTAs and review sites take close to half of citations or more: 55.3% across three engines in Cloudbeds' study, 46.5% on Google AI Mode in Sitter's. When they ask for open travel guidance, independent blogs dominate and OTAs take 11.1%. Ahrefs puts Tripadvisor 21st on ChatGPT, but that ranking covers every US query, most of them unrelated to travel, so it says little about travel answers.

Social is unconfirmed. Facebook at 11.58% and Instagram at 4.01% stand out only in AthenaHQ's pooled data, while two AI Mode studies put social at 5.1% and user-generated content at 1.4%. Test it in your own prompts before planning around it.

AI cites explanations first and comparisons second

AI cites pages that explain before pages that compare, and pages that compare before pages that sell.

59.39%: of citations are informational or comparative content; Source: AthenaHQ, State of AI Search 2026. 1 Vendor dataset
56.8%: of citations in Google AI Mode travel answers go to independent guides and blogs; Source: Search Agency, July 2026. 17
19.42%: of Gemini's citations go to learning and education content, the most of the three engines shown; Source: AthenaHQ, State of AI Search 2026. 1 Vendor dataset

Share of citations by content intent and engine, all segments
Intent	ChatGPT	Perplexity	Gemini	Market average
Informational	35.54%	36.74%	32.41%	36.23%
Comparative, selection	23.82%	21.88%	21.80%	23.16%
Acquisition, obtaining	15.54%	15.23%	12.80%	15.05%
Learning, education	13.10%	13.73%	19.42%	13.43%
Navigation, institutional	3.99%	4.22%	4.72%	3.70%
Consumption, experience	3.24%	3.89%	3.08%	3.48%
Updates, news	1.94%	1.47%	2.23%	1.84%
Investigation, research	1.67%	1.68%	2.11%	1.81%
Optimisation, improvement	1.17%	1.17%	1.43%	1.30%

Source: AthenaHQ, State of AI Search 2026. Columns show the three engines AuraScope measures. 1 Vendor dataset

The order holds on all three engines: explainers and guides first, comparisons and best-of pages second, pricing and product pages third. Only the weighting changes.

Gemini leans furthest towards teaching. It gives learning and education content six points more weight than the market, and gives acquisition content the least of the three engines. A brand that appears in ChatGPT but not in Gemini is often thin on explanatory content.

For a travel brand that means three kinds of page. Guides that explain a destination, route, room or fare. Comparisons of one product against another, and against competitors. And dated updates for schedules, openings and loyalty changes.

In travel, AI reads product pages first

AI enters a travel brand's site through product pages 46.08% of the time, against 13.79% across all markets, so room, route and fare pages are read first. Travel answers also lean on content that explains rather than compares.

Travel and Hospitality against all segments
Measure	All segments	Travel	Difference
Average brand mention rate	16.30%	17.75%	+1.45 pts
Top brand mention rate	56.48%	56.63%	+0.15 pts
Average citations of your own domain	16.05%	18.48%	+2.43 pts
Share of voice, rank 1	33.64%	38.59%	+4.95 pts
Top three combined share of voice	66.39%	71.44%	+5.05 pts
AI entries via /blog	37.53%	15.29%	-22.24 pts
AI entries via /products	13.79%	46.08%	+32.29 pts
Informational content	36.23%	51.13%	+14.90 pts
Comparative, selection	23.16%	14.56%	-8.60 pts
Acquisition, obtaining	15.05%	7.68%	-7.37 pts
Updates, news	1.84%	9.23%	+7.39 pts

Source: AthenaHQ, State of AI Search 2026, travel segment (67 websites) against all segments (1,854 websites). The travel sample leans towards hotels and OTAs, so treat it as directional. 1 Vendor dataset

Travel content explains more than it compares

Informational content rises 15 points to 51.13% and comparative content falls 9. AI describes travel options more than it ranks them.

Dates matter more

Updates and news carry five times the market's weight in travel, so an undated page looks stale.

Mentions are more concentrated

The top brand's mention rate barely differs from the market, 56.63% against 56.48%. But travel's top three brands take 71.44% of mentions, five points more than the market, which leaves less for everyone below them.

Part 3: Where brands lose ground

Appearing in an answer once is not enough. Citations fade within weeks, most hotel websites we audited are hard for AI to read, and airlines get named without the link.

Citations fade within weeks

A ChatGPT citation has a half-life of 3.4 weeks, Perplexity 5.8 weeks and Google's surfaces 4.3 to 4.8 weeks. Visibility has to be measured continuously, not once.

Half-life of a citation, by engine
Engine	Half-life
ChatGPT	3.4 weeks
Google surfaces	4.3 to 4.8 weeks
Perplexity	5.8 weeks

The Google bar is drawn at the midpoint of the published range.

Source: Scrunch and Stacker, 3.5 million citation events, September 2025 to March 2026. 6

What changed the numbers in 2026
Date	Event	Effect on measurement	Source
February 2026	Gemini 3 arrived in AI Overviews. 42.4% of previously cited domains were replaced, and sources per Overview rose from 11.55 to 15.22.	Long-tail domains dropped out as the source count rose.	18 Secondary coverage
16 February to 2 March 2026	Gemini's citation rate fell from 99% to 76%.	Fewer answers can be traced to any source.	10
8 and 14 August 2026	ChatGPT's Reddit citations stepped down twice, 73% in total. Other engines stayed flat.	A Reddit-led plan lost most of its ChatGPT value in a week.	8
31 August 2026	Google AI Mode began wrapping citations in google.com/searchviewer links.	Third-party tools can no longer see which domains AI Mode cites.	11
15 September 2026	Cloudflare began blocking mixed-use AI crawlers by default on ad-supported pages for new and free-tier sites, and replaced pay-per-crawl with pay-per-use.	A site can drop out of AI retrieval without changing a page.	43

Source: Studies and coverage named per row.

Engines change how they retrieve sources without notice, and a CDN or crawler setting can take a site out of an engine's reach overnight.

Two things follow. Read trends across repeated runs and rolling averages, never a single answer. And treat crawler access to your own domain as a marketing risk that someone reviews every month: an agent that cannot reach a page cannot cite it.

Most New Zealand hotel websites are hard for AI to read

In May 2026 AuraScope audited the websites of 147 New Zealand luxury and boutique hotels. Of the 107 we could evaluate fully, 79.4% scored Weak or Critical on the on-page signals AI engines use to read and date a site.

79.4%: of fully evaluated hotel websites scored Weak or Critical, below 50 out of 100. Three, or 2.8%, scored Strong.; Source: AuraScope audit, 107 of 147 hotels, 10 May 2026. 15 AuraScope research
52.3%: publish no machine-readable dates, so an engine cannot tell how current a page is.; Source: AuraScope audit, 10 May 2026. 15 AuraScope research
23.4%: have no structured data at all: no JSON-LD, Microdata or RDFa.; Source: AuraScope audit, 10 May 2026. 15 AuraScope research

107 New Zealand hotel websites by on-page score band
Score band (out of 100)	Share of hotels
Strong, 70 or more	2.8% (3)
Moderate, 50 to 69	17.8% (19)
Weak, 30 to 49	39.3% (42)
Critical, under 30	40.2% (43)

The average score was 35.8.

Source: AuraScope audit, 10 May 2026. 15 AuraScope research

These are the basics the companion agent-readiness paper calls retrieval hygiene: whether an engine or agent can read a hotel's own facts and tell how current they are. The audit does not show that low scores cost citations. It shows that most of these sites give an engine less to work with than an OTA listing, which Cloudbeds describes as structured, real-time content.

Sources: 16 12

Airlines are named in answers but rarely linked

In ChatGPT flight conversations, airlines were named without a link 74.6% of the time. OTAs and metasearch sites went unlinked less than 9% of the time.

74.6%: of airline mentions in ChatGPT flight conversations carried no link, against 8.8% for OTAs and 8.6% for metasearch; Source: PROS, reported by Travel Weekly, 4 November 2025. 23
2.3%: of airline website referral traffic came from ChatGPT in September 2025, up from 1.8% in August. ChatGPT sent 95% of airlines' AI referrals.; Source: PROS, reported by Travel Weekly, 4 November 2025. 23
500+: airlines bookable through Meta's Muse in the US since 9 September 2026. Google AI Mode tracks flight prices but does not book, and the Gemini app "can't book a flight or hotel for you".; Source: Duffel and Travel Daily News, September 2026; Google, 27 August 2026; Gemini help centre. 38 35 29 32

Where an airline can appear before the booking
Step	What the traveller asks	Who answers today	Where the airline can act	Source
Destination	"Two weeks in November, good food and plenty of nature. Where should we go?"	Guides and blogs (56.8% of AI Mode travel citations), Reddit, Tripadvisor and Wikipedia. Official tourism sites 1.2%.	Destination content made with tourism bodies and local operators, with dated updates.	17
Route and airline	"What's the best way to fly from London to Tokyo?"	Metasearch, OTAs and forums. On ChatGPT, airlines are named without a link 74.6% of the time.	Route pages that answer the whole question: schedule, aircraft, cabin, connections and baggage.	23
Cabin and product	"Is premium economy worth it on an overnight flight with kids?"	Reviews, YouTube and forums.	Comparison pages and explainers on the airline's own site, with dates and specifics.
Booking	"Book it."	Expedia's ChatGPT app searches and hands off to Expedia, and is live in New Zealand. Virgin Australia and Webjet apps search in ChatGPT and complete bookings on their own sites. Google AI Mode books hotels only, in the US only. Meta's Muse books US flights through Duffel.	The airline's own site stays the booking surface. The work is making it the place the assistant sends the traveller.	44 45 46 29 38 Secondary coverage

Source: Travel Weekly, PROS data, 4 November 2025; TechCrunch, 27 August 2026; Expedia, February 2026; Virgin Australia newsroom, 2026; Travel Weekly Australia, 11 May 2026.

In the journey above, the destination is settled before any airline comes up. For an airline flying into that destination, the tourism body's content and local operators' content become part of how travellers find it. At the route question, OTAs and metasearch hold the links while airlines get the mentions.

The same work applies at every step: publish the page that answers the question in full, in HTML an agent can read without running JavaScript, and make sure AI crawlers can reach it.

The PROS figures are the only published airline citation data we found, and they are almost a year old.

Part 4: From answer to booking

The answer engines have held back from taking the booking. Personal agents have not, and the way they book raises the stakes on everything above.

The answer engines held back from checkout

OpenAI retired in-chat checkout in March 2026 and built a US$1 billion advertising run rate instead. Google AI Mode books hotels, in the US only. Neither books flights.

Checkout and booking moves by the big platforms, status at 1 October 2026
Date	Move	Status	Source
5 March 2026	OpenAI retires Instant Checkout inside ChatGPT, "prioritising making ChatGPT search and product discovery great". Transactions move to apps and merchant sites.	Withdrawn	26
9 February to 31 August 2026	ChatGPT ads go from a US pilot to Europe, then to a US$1 billion annualised run rate in under 200 days across more than 40 countries. Targeting uses the context of the conversation.	Live, revenue data	28
6 May 2026	Mindtrip completes flight checkout inside chat using Sabre and PayPal. Hotels to follow.	Live, no usage data	34
20 May 2026	Google's Universal Commerce Protocol checkout goes live in select US retail. Australia is named for "the coming months". New Zealand is not named.	Limited	31 Secondary coverage
9 August 2026	OpenAI shuts the ChatGPT Atlas browser and retires the consumer ChatGPT agent. Browser agent work moves to ChatGPT Work on paid plans and a Chrome extension.	Withdrawn, replaced	27
27 August 2026	Google AI Mode books hotels in the US, in English, with Booking.com, Expedia, Hilton, Marriott, IHG, Wyndham, Priceline, Choice, Hotels.com and Trip.com. Google Pay handles checkout and the hotel or OTA stays merchant of record. Flights are not bookable, but prices can be tracked in 180+ countries. Other hotels cannot join yet: Google's UCP for Lodging onboarding and specs are "coming soon".	Live, US only, no usage data	29 30 13
23 September 2026	Airbnb commits to launching its own agent in 2027 and stays out of ChatGPT apps.	Announced	42

Source: Announcements and coverage named per row.

No platform, card network or airline has published a count of bookings completed by an AI agent. Our view: visibility is not bookability. Work on visibility now, whether AI can find, read and cite your own facts, and run bookability, an agent completing the reservation, as a separate programme. The next section explains why that programme cannot wait until 2027.

Personal agents have started booking travel

Meta, OpenAI and xAI launched personal agents between August and September 2026, and Instinct's founder says more than half of its transactions are travel. Muse books flights through a direct connection and hotels by browsing consumer sites.

50%+: of transactions on Instinct, an invite-only personal agent, are travel, according to its founder. He says the platform is approaching US$1 billion in annual transactions but gave no basis for the figure.; Source: TechCrunch, 29 September 2026, from a podcast interview. 36 Secondary coverage
2 routes: for a personal agent to book: a direct connection to supply, as Muse has with Duffel for flights, or a browser operated like a person, as Muse uses for hotels.; Source: Skift, 9 September 2026. 37

Personal agents and how they book travel, at 1 October 2026
Agent	Launched	How it books travel	Availability	Source
Grok Bot (xAI)	11 August 2026, beta	A general agent on its own cloud computer, signed in to the user's accounts. It navigates sites and fills in forms. No travel partnership has been announced.	SuperGrok Heavy, Cursor Ultra and Cursor Teams Premium	41 Secondary coverage
Muse (Meta)	9 September 2026 for travel	Flights through a direct Duffel integration with live inventory from 500+ airlines, paid by Stripe virtual card. Hotels by browsing consumer sites such as Expedia and Hotels.com.	Travel booking in the US only	38 37 35
Instinct	Invite-only	More than half of transactions are travel. Its founder describes requests such as "I need to be in New York tonight".	Invite-only	36 Secondary coverage
dots (OpenAI)	29 September 2026	Takes actions for the user, including booking flights, according to launch coverage. OpenAI has not published how bookings are completed.	ChatGPT Pro and Business plans	39 40 Secondary coverage

Source: Announcements and coverage named per row.

The browsing route favours whoever is easiest to read

Only the direct route gives a travel company a defined data connection and a commercial relationship. On the browsing route, Skift's reading is that "the agent controls where it searches and books, leaving suppliers as passive websites", while suppliers carry the cost of extra look-to-book traffic.

That is where this report meets our companion paper on agent readiness. An agent that books by browsing can only use a site it can reach and read. Muse's hotel shopping runs on OTA sites built for exactly that. A hotel whose rates sit behind JavaScript or a bot wall gives the agent nowhere else to go.

Personal agents also narrow the gap between visibility and bookability. When the agent that reads the page is the one that books, the same readable facts, rates and availability serve both.

Sources: 37 16

Part 5: What to do now

What the evidence asks of a travel brand today.

What travel brands can do now

Work on visibility now: be readable, cited and accurate at each question in the trip, and measure it as a trend. Run bookability as a separate track, because personal agents have started to book.

One trip, four questions

A family wants two weeks away in November, with good food and nature. They ask an assistant where to go, how to get there and which room or seat is worth paying for, then they book. Each question draws on a different set of sources, and each is a chance to appear before the booking page.

On your own website

Being cited starts with being readable. An agent has to reach a page and read the answer in it before it can ground a response on it, so facts that appear only after JavaScript runs, or pages behind a bot wall that returns HTTP 403, drop out. AuraScope's companion paper sets out what that asks of a hotel website.

Sources: 16

One experiment, start to finish

Traveller question
Evidence gap
Change
Owner
Success measure
Review date

Putting it together

The trip now starts in an AI answer. That answer names a few brands, leans on other people's sources and forgets them within weeks, and the agents that are starting to book read the same pages. The travel brands that are readable, cited and accurate there are best placed for both.

Appendix: New Zealand context

Context for New Zealand readers: arrivals are back, budgets are not.

New Zealand arrivals passed 2019 levels in July 2026

New Zealand had 256,600 overseas visitor arrivals in July 2026, 100.4% of July 2019 and the first month above pre-COVID levels. Gross travel bookings in Australia and New Zealand are forecast to fall 5% this year.

256,600: overseas visitor arrivals in July 2026, 100.4% of July 2019 and the first month above pre-COVID levels; Source: Stats NZ, 14 September 2026. 49
3.69 million: arrivals in the year to July 2026, up 305,000. Australians made up 53% of July arrivals.; Source: Stats NZ, 14 September 2026. 49
-5%: forecast change in Australia and New Zealand gross travel bookings in 2026, to US$37.4 billion. Hotels -9%, airlines -1%.; Source: Phocuswright, August 2026. 51

Visitor arrivals in the year to June 2026 as a share of 2019, by market
Market	Share of 2019 arrivals
United States	105.1%
Australia	104.7%
United Kingdom	84.0%
China	74.9%

Source: Stats NZ, year ended June 2026, read through a release mirror. 50 Secondary coverage

Australia and the United States carry the recovery. The United Kingdom and China are still well below 2019. Fuel and living costs are pulling forecast bookings down even as arrivals rise.

MBIE committed $800,000 of International Visitor Levy funding on 30 July 2026 to make Tourism New Zealand's systems and operator database ready for AI, and to add regional content to its AI travel assistant. In June, IATA halved its 2026 airline industry profit forecast to US$23 billion, citing fuel prices and disruption in the Middle East.

Sources: 48 52

How to read the numbers

AthenaHQ's State of AI Search 2026 is a vendor dataset from a company that sells generative-engine-optimisation software. It pools seven engines (ChatGPT, Gemini, Perplexity, Grok, Copilot, Google AI Overviews and Google AI Mode), is weighted towards the US and covers December 2025 to March 2026. It tracks 1,854 active websites across all segments and 67 in Travel and Hospitality, a sample that leans towards hotels and OTAs. Prompt volume and site selection are not published. This report uses it for structure and relative comparisons, not for absolute claims.
Independent 2026 measurements are named beside each figure with their sample, engine and month. None is peer reviewed, and most come from measurement vendors. Where two vendors disagree, as they do on Reddit's weight in ChatGPT, both are shown.
This is an evidence review, not a meta-analysis. It gathers 2026 studies, surveys and company statements that measure different things, and reads them side by side rather than pooling them statistically.
AuraScope's own survey asked NZ Tech Expo attendees in Auckland how they planned and booked their last stay (n=89, 18 to 24 September 2026). It is not a general traveller sample: 70.8% of respondents work in technology or AI, including 7.9% who also work in travel.
AuraScope's other contribution is one small audit: the on-page signals of 147 New Zealand luxury and boutique hotel websites, measured on 10 May 2026, with 107 evaluated fully. It measures readability, not presence in AI answers.
Surveys record stated preference, not observed behaviour. None of the external surveys samples New Zealand travellers, and AuraScope's own survey samples tech expo attendees, not travellers in general.
AuraScope measures ChatGPT, Perplexity and Gemini. Google AI Overviews and AI Mode appear here as third-party context only.
Every 2026 figure is dated. Anything older than November 2025 is labelled.

Key terms

Answer engine: A service that answers a question in its own words instead of listing links: ChatGPT, Gemini, Perplexity, Google AI Overviews and Google AI Mode.
Personal agent: Software that acts for one person over time, such as Meta's Muse or OpenAI's dots. It can browse sites, fill in forms and, in some cases, book.
Citation: A link or named source an answer engine shows as the basis for its answer.
Grounding: Basing an answer on a page the engine actually read, rather than on what the model already remembers.
Half-life: How long it takes for half of the citations a source holds today to disappear from answers.
Mention rate: The share of answers about a category that name a given brand.
Share of voice: A brand's share of all brand mentions across a set of answers.
OTA: Online travel agency, such as Booking.com or Expedia.
Metasearch: A site that compares prices from many sellers and sends the traveller elsewhere to book, such as Kayak or Google Flights.
LLM referral: A visit or booking that arrives from a link inside an AI answer. LLM stands for large language model.
Visibility: Whether AI can find, read, cite and recommend a brand and its own facts.
Bookability: Whether an agent can complete a reservation, through a booking engine, a direct connection or a checkout protocol.
Look-to-book: The number of searches or page loads for each booking made. Agents that browse can push it up.
Merchant of record: The business that takes the payment and owns the transaction with the traveller.
UCP: Universal Commerce Protocol: Google's standard for agents to complete purchases. A version for hotels has been announced but is not open yet.
HTTP 403 and bot walls: A 403 is the code a server returns when it refuses a request. Bot-protection settings, often in a CDN such as Cloudflare, can return it to AI crawlers and agents.
Pooled dataset: Figures combined across several engines into one number, which can hide large differences between them.

Sources

1AthenaHQ, State of AI Search 2026, 2026. Vendor dataset. Seven engines pooled, US-weighted, December 2025 to March 2026.
2Similarweb, AI referral traffic by industry, June 2025 to May 2026.
3Expedia Group, The AI Trust Gap, 14 April 2026.
4PhocusWire, Booking Holdings Q2 2026 earnings, 4 August 2026.
5PhocusWire, Expedia Group Q2 2026 earnings, 5 August 2026.
6Scrunch, The half-life of AI citations, 2026.
7Scrunch, The Reddit paradox, 5 February 2026.
8Otterly, ChatGPT Reddit citations, 27 August 2026, updated 14 September 2026.
9Ahrefs, Most cited domains in ChatGPT, Updated 2 September 2026.
10Seer Interactive, Gemini citations decreased 23pp, 13 April 2026.
11Nicolas Sitter, Google AI Mode hotel study 2026, February 2026, updated September 2026.
12Cloudbeds, How AI recommends hotels, research data, Accessed 30 September 2026. Publication date not shown. Vendor research: 810 prompts across ChatGPT, Perplexity and Gemini, six destinations, 145 properties.
13Google Developers, UCP for Lodging, Accessed 9 September 2026.
14AuraScope, NZ Tech Expo attendee survey, Auckland, 18 to 24 September 2026. AuraScope's own survey. NZ Tech Expo attendees (n=89). Not a general traveller sample.
15AuraScope, On-page audit of 147 New Zealand hotel websites, 10 May 2026. AuraScope's own measurement. Raw scores held by AuraScope; method summarised in the report.
16AuraScope Research, A hotel's take on agent readiness, September 2026. Companion paper on whether agents can reach and read a hotel's own website.
17Search Agency, AI Mode travel guides still perform, 9 July 2026.
18SE Ranking, Gemini 3 impact on AI Overviews, 26 February 2026.
19Adobe Digital Insights, AI traffic to travel sites, 17 June 2026.
20Axios, Chartbeat search traffic data, 17 March 2026.
21Growth Memo, AI halftime report H1 2026, 27 July 2026.
22Google, AI search driving more queries and higher quality clicks, 6 August 2025.
23Travel Weekly, OTAs have edge over airlines in ChatGPT (PROS data), 4 November 2025.
24McKinsey and Skift Research, Remapping travel with agentic AI, 16 September 2025.
25Skift, OTAs, AI discovery and transactions, 20 March 2026.
26Skift, OpenAI ChatGPT checkout walkback, 5 March 2026.
27OpenAI, Evolving Atlas into ChatGPT for browser-based agentic work, July 2026.
28OpenAI, Expanding access to AI with ChatGPT ads, 31 August 2026.
29TechCrunch, Google's AI Mode can now track flight prices, help book hotels and more, 27 August 2026.
30Skift, Google confirms hotel agentic booking is now in testing, 7 August 2026.
31Google, Shopping updates at Google Marketing Live, 20 May 2026.
32Google, Gemini Apps help: flights and hotels, Accessed September 2026.
33Google New Zealand, Introducing AI Mode in Aotearoa New Zealand, 22 August 2025.
34Skift, Sabre, Mindtrip and PayPal launch agentic AI travel booking, 6 May 2026.
35Travel Daily News, Meta's Muse adds Duffel for AI-powered travel bookings, 10 September 2026.
36TechCrunch, Instinct founder said more than 50% of transactions on the platform are travel-related, 29 September 2026.
37Skift, Meta says its Muse agent books travel. Here's what that actually means, 9 September 2026.
38Duffel, Millions of users can now use Duffel to search, book and manage holidays on Muse, September 2026.
39The San Francisco Standard, OpenAI launches cute, 'capable' AI agents, 29 September 2026.
40CBS News, Sam Altman unveils dots, OpenAI's new AI personal agent, 29 September 2026.
41DataCamp, Grok Bot explained: SpaceXAI's always-on AI agents, Accessed 1 October 2026. Secondary explainer. xAI's own announcement page could not be retrieved.
42Skift, Brian Chesky commits to launching an AI agent in 2027, 23 September 2026.
43TechCrunch, Cloudflare's new policy pushes AI companies to pay for publishers' content, 1 July 2026.
44Expedia, Expedia in ChatGPT, Updated 20 February 2026.
45Virgin Australia, Virgin Australia flight search lands in ChatGPT, 2026.
46Travel Weekly Australia, Webjet launches ChatGPT app for flight and hotel searches, 11 May 2026.
47Tourism New Zealand, TNZ hosts AI discoverability workshops for industry, 27 July 2026.
48MBIE, Using AI to improve New Zealand tourism visibility, 30 July 2026.
49Stats NZ, International travel: July 2026, 14 September 2026.
50Stats NZ, Annual visitor arrivals up 9 percent, year ended June 2026, 14 August 2026. Read through a release mirror.
51Phocuswright, Why the ANZ travel market is contracting in 2026 and what recovers by 2029, August 2026.
52IATA, Middle East disruptions and high fuel prices halve airline industry profitability, 7 June 2026.