Which Sources Keep Appearing in AI Travel Discovery?

A hotel appears in an AI travel answer. Its website is cited. Someone takes a screenshot and shares it with the commercial team.

That is a useful observation. It is not yet a result the team knows how to repeat.

Should the hotel update its website, correct a third-party listing, or publish better destination information? A source list can suggest where to look. It cannot settle which change deserves the work.

We analyzed 29,320 retained source-domain records across 51 travel destinations and four AI systems. A small group of domains appeared repeatedly, surrounded by a much larger intermittent set. Closely spaced scans shared a median of 31% to 66% of their combined domain sets, depending on the engine. [1]

The opportunity is to identify what keeps returning, inspect its relevance, and use that evidence to design better experiments.

Measurement is not the end of the marketing workflow. It is what makes the next experiment worth running.

Why this matters: AI discovery is becoming a booking channel

Hotels do not need another visibility score for its own sake. They need to enter the traveler's consideration set, convert that interest, and understand how the reservation reaches them. Appearing against a known competitive set is not the same as being discovered by someone who has not chosen a property yet.

Two booking paths are taking shape. In August 2026, Hilton described a ChatGPT integration through which travelers can explore properties before continuing to Hilton.com to book. That is an AI-assisted route into the hotel's own channel. [2]

Google is bringing the transaction into the conversation. Its August 27 announcement introduced hotel booking within AI Mode, initially rolling out in the U.S. in English. Travelers can choose an integrated partner, select a room and pay within Google's interface. Announced partners include Hilton and Marriott, alongside Booking.com and Expedia. [3]

At September's Destination AI conference, Hilton CIO Michael Leidinger emphasized making property information accessible to AI systems and anticipated travelers delegating more decisions and transactions to agents, according to HOTELS' coverage. [4]

This is becoming a distribution question, not just a content question.

An AI-assisted booking is not automatically a direct booking. In Google's flow, either the hotel or a booking platform can be the merchant of record. [3] The hotel still needs to understand acquisition terms, who handles the reservation, and how the guest relationship is managed.

Nor will every discovery produce a website visit. Pew Research Center's study of 900 US adults' March 2025 browsing found traditional-result clicks on 8% of Google visits associated with an AI summary, compared with 15% without one. This is general-search evidence, not a hotel-conversion estimate. [5]

Hotels should prepare for both paths: AI that sends travelers to their own channels and AI that helps complete a booking through an integrated partner. Source Intelligence addresses the information side. Booking outcomes must be measured alongside visibility, not inferred from mentions.

What did we measure?

Anana's public Source Tracker runs defined travel-discovery questions against ChatGPT, Claude, Gemini and Perplexity through their APIs. A retained record means a domain appeared in at least two successful responses within one destination-engine scan.

The primary cohort contains 342 scans per engine across 51 destinations, collected over seven selected periods between July 27 and September 14, 2026. Its 29,320 records contain 2,898 distinct domains. [1]

Each scan combines questions naming the destination with open-ended travel questions. A record's destination label identifies its collection context, not necessarily its webpage's subject.

The engines expose different source evidence, including answer citations, search results and grounding sources. We analyze each separately. The tables describe what each collection mechanism retained, not a ranking of engine quality. Full collection rules and exclusions appear in the methodology.

A small group of domains appears repeatedly

Each engine's ten most frequently observed domains accounted for between 27% and 29% of its retained records.

Source: Anana Source Tracker, primary cohort. Each engine has its own top-ten list and record denominator. [1]

The recurring names differed. tripadvisor.com appeared in 339 of 342 Claude scans and the same number of Perplexity scans. expedia.com appeared in 295 Gemini scans. ChatGPT's leading domain was travel.state.gov, retained in 277 scans. [1]

These are frequencies of observation, not estimates of a website's influence on an answer. Their practical value is prioritization: a commercial team has a starting list to investigate rather than a reason to examine every website equally.

The concentration pattern survived a collection-quality check. Restricting the analysis to 237 higher-success snapshots produced top-ten shares between 26.1% and 30.3%. [1]

A 2025 preprint by Kai-Cheng Yang also found concentrated news citations in AI search. It concerns a different subject and measurement approach, so it provides context, not replication of these findings. [6]

A recurring core sits alongside an intermittent tail

A domain can be common across the dataset without appearing consistently in any one destination's scans. We examined the 37 destinations represented in all seven selected periods, defining a persistent destination-domain pairing as one retained in six or seven scans.

Source: Anana Source Tracker, balanced 37-destination panel. The seven selected periods are not seven consecutive calendar weeks. [1]

The averages describe a manageable recurring set, not a fixed list size for every destination. Some members were recognizable accommodation or tourism sources: atlantisbahamas.com appeared in all seven ChatGPT scans associated with Nassau, and grecotel.com did so for Corfu. visitjamaica.com appeared in all seven Kingston scans for both Claude and Perplexity. [1]

In nine destinations with higher collection success throughout, average persistent counts ranged from 6.3 to 18.0. That subset has a different geographic mix; it is a sensitivity check, not a corrected estimate for the wider panel. [1]

The distinction is useful without being absolute. Persistent sources deserve systematic inspection. Intermittent sources can still reveal relevant information, but a single appearance is a weaker basis for prioritizing work.

A source associated with a destination is not necessarily about it

How widely a domain appears and what its pages discuss are separate questions.

In Perplexity's primary cohort, 39 domains appeared across at least 26 of 51 destinations and accounted for 53.1% of its records. Broadly recurring domains accounted for 47.2% of Claude's records, 35.9% of Gemini's and 21.4% of ChatGPT's, each calculated within its own sample. [1]

A global booking platform can cover many destinations. A hotel-group domain can contain individual property pages and general travel articles. The domain alone cannot tell the reader which was returned.

The page examples are revealing. In ChatGPT's records, the US State Department's Mexico travel-advisory page appeared in 147 scans across 46 destinations, 42 outside Mexico. Pages under Four Seasons' Bora Bora section appeared in 55 scans associated with 30 destinations. [1]

The mixed prompts matter: open-ended questions can legitimately return sources about somewhere other than the named destination. The historical export combines both prompt types and cannot assign each retained domain back to its originating question. Equally, a narrowly observed domain is not automatically a local source.

Before treating a recurring domain as an opportunity, inspect the captured page. Does it describe the destination, a relevant traveler need, or a competing property? One stored URL is an example, not the domain's complete source history.

Liu, Zhang and Liang's 2023 evaluation found imperfect citation support in earlier generative-search systems, reinforcing the value of inspection. Those historical findings do not measure the systems studied here. [7]

Recurrence helps prioritize the review. Relevance determines whether the source belongs in the decision.

How much can source lists change between scans?

Scheduling produced opportunities to compare the same destination and engine again within 2.5 days. These incidental repeats provide a check on short-interval variation, not a controlled experiment.

Overlap is Jaccard similarity: shared domains divided by all distinct domains in either scan. Pairs cover July 26 to October 6; Claude pairs end in September. Higher-success pairs had at least 72 of 80 successful responses at both collection points. [1]

For three engines, median overlap was below half the combined domain set. Retrieval variation, incomplete collection and the retention threshold can all contribute.

That threshold matters. A domain can move from two supporting responses to one and disappear from the export even though the engine still returned it. Conversely, domains supported by more responses in the first scan were more likely to be retained again within each engine. [1]

When a hotel edits a page and observes a different answer, this variability is part of the context for interpreting the result. It does not isolate the effect of the edit, but it helps explain why one AI search cannot measure a hotel's visibility.

One favorable answer is useful. Repeated favorable answers under comparable conditions provide a stronger basis for deciding what to test further.

Turn source evidence into better experiments

AI Source Intelligence examines the public sources surfaced in or alongside AI travel answers. It complements Hotel AI Visibility: one asks which sources appear around the response; the other asks how the hotel itself appears.

Anana's AI Presence workflow in Visibility Lab supports repeated observations around defined traveler questions and competitive sets. Source Tracker adds an external, destination-associated view. Together they help teams formulate hypotheses, not assume that every difference is a proven opportunity for visibility uplift.

Start with a commercial objective. Guest conversations can help identify which information deserves attention: what people ask, what remains unclear, and what prevents them from deciding.

Consider an illustrative resort that wants to appear for family stays with connecting rooms. The team defines relevant traveler questions, records how the resort and competitors appear, and inspects the accompanying sources.

If its own information leaves room arrangements unclear, the intervention becomes specific: publish accurate details about room combinations, occupancy limits and whether connections are guaranteed or requested. That is more testable than "we need better AI SEO."

After the change, repeat the questions under comparable conditions. Does the property appear more accurately? Does it appear more consistently? Where feasible, compare against unchanged questions or properties rather than attributing every movement to the edit.

Publishing the page is an action. Improved representation and more frequent recommendation are outcomes to measure separately.

Academic work supports testing rather than assuming. Aggarwal and colleagues' GEO: Generative Engine Optimization, accepted at KDD 2024, found that content changes could improve visibility under their experimental measures, with effectiveness varying by domain. That supports experimentation, not a universal hotel-optimization formula. [8]

Our source research addresses an earlier step: finding recurring evidence, checking relevance and understanding variation before attributing a result to a change.

Where tracking allows, follow the experiment through to referral visits, qualified inquiries and bookings. Keep hotel-direct and intermediary routes distinct. A higher visibility score is not the commercial outcome.

A better question than "did we appear?"

The research reveals recurring sources, intermittent observations and meaningful short-interval variability. It gives hotels a starting point for investigation, not a shortcut around experimentation.

Identify what keeps returning. Check what the relevant pages say. Choose a change, define the expected result, and measure what follows.

The goal is not another list of where the hotel appeared. It is to earn consideration for demand the hotel can serve, turn that interest into a booking, and build evidence about which work helped.

Methodology and limitations

Collection and units. Source Tracker sends 20 English-language prompts per destination to each engine: ten traveler personas, each with a destination-named and an identical-across-destinations open question. Calls use APIs and structured collection, not consumer browsing sessions. A retained record is one domain observed in at least two successful responses within a destination-engine scan. The collector applies a fixed host-exclusion list; subdomains and country-domain variants can remain separate. Historical storage contains aggregates and one representative URL per domain, not raw response histories. [1]

Cohorts. The full export contains 45,922 records from 596 destination-week snapshots collected July 26 to October 9, 2026. The primary cohort contains 342 four-engine snapshots, equivalent to 342 scans per engine, and 29,320 records across 51 destinations over seven selected periods from July 27 to September 14. Exclusions cover rejected and identified mis-geocoded entries, the initial partial panel, a period with substantial missing coverage, and the October 9 batch collected after pipeline changes. Persistence uses 37 destinations with all seven periods; repeat-pair analysis separately includes eligible observations up to 2.5 days apart. Destinations were not sampled to represent global tourism. [1]

Extraction and completeness. ChatGPT records inline answer citations; Claude records retrieved search results; Perplexity combines search results and citations; Gemini records grounding domains. These mechanisms prevent like-for-like cross-engine citation comparisons. Concentration uses each engine's retained-record total; persistence counts destination-domain pairings across seven periods. Repeat similarity uses the union of domains across two scans. Findings were checked against higher-success subsets, including 237 primary snapshots with at least 72 of 80 successful responses. This measures total collection success, not equal success per engine. Historical per-engine denominators are unavailable; the export does not establish whether additional incomplete snapshots were absent or filtered out. [1]

Interpretation. The two prompt types are blended. Domain absence can reflect the retention threshold or incomplete collection. Actual collection dates determine elapsed time; observations are unevenly spaced and repeat pairs were not controlled replicates. Findings describe retained source presence, not source influence, traveler exposure, bookings or an optimization intervention's effect. [1]

Frequently asked questions

Which sources do AI travel engines return most often?

Recurring examples included Tripadvisor in Claude and Perplexity scans, Expedia in Gemini scans, and the US State Department travel site in ChatGPT scans. These are protocol-specific observations, not a universal ranking. [1]

Does appearing on a frequently observed source improve hotel AI visibility?

This study does not measure that effect. A recurring, relevant source can inform an experiment; changes in hotel representation and recommendation frequency require separate measurement.

Can travelers book a hotel inside an AI search experience?

Google announced hotel booking within AI Mode in August 2026, initially in the U.S. in English with integrated partners. Other paths, including Hilton's ChatGPT integration, send travelers to a booking website. [2, 3]

How does Source Intelligence complement hotel AI Presence?

Source Intelligence examines observable sources. AI Presence examines how a hotel appears for defined traveler questions and comparison sets. Together they inform what to investigate, change and measure.

References

1. Anana proprietary research. Source Tracker July-October 2026 export (source-tracker-all-time.csv; MD5 b5e9c6d827694cda586198f95c0d7573), with integrity, exploratory and cross-review analyses. Cohort definitions and exclusions appear in Methodology and limitations. The public Source Tracker is the collection tool, not the frozen historical analysis cohort. Claim-level calculation locations are retained in the editorial evidence register.
2. Hilton (2026, August 28). New Extensions of Hilton's Industry-Leading Technology Platform are Connecting the Journey from Search to Stay. Official announcement, cited for the booking path rather than performance evidence.
3. Byers, J. (2026, August 27). 3 new ways to plan and book travel in Search. Google. Protocol note: Google's separate UCP for Lodging documentation describes compatibility with MCP and planned adoption for AI Mode. Joining the booking integration does not improve property ranking. This protocol initiative is distinct from the announced booking rollout.
4. Eisen, D. (2026, September 29). How AI is rewriting hotel distribution, one booking at a time. HOTELS. Conference reporting; executive views are attributed as such.
5. Chapekis, A., and Lieb, A. (2025, July 22). Google users are less likely to click on links when an AI summary appears in the results. Pew Research Center.
6. Yang, K.-C. (2025). News Source Citing Patterns in AI Search Systems. arXiv:2507.05301. Preprint.
7. Liu, N. F., Zhang, T., and Liang, P. (2023). Evaluating Verifiability in Generative Search Engines. Findings of EMNLP 2023, 7001-7025.
8. Aggarwal, P., Murahari, V., Rajpurohit, T., Kalyan, A., Narasimhan, K., and Deshpande, A. (2024). GEO: Generative Engine Optimization. Accepted at KDD 2024.

Industry developments checked October 12, 2026. Linked Anana resources provide definitions and product context, not independent validation of these findings.