AI, Distribution, and the Future of Direct Versus Intermediated Lodging Booking

Executive summary

The question is not whether artificial intelligence will change how travelers discover accommodation. It already has. The harder question is whether a change in the interface through which demand begins will alter who controls the transaction, who owns the customer relationship, and who captures the economics of distribution.

The evidence available as of August 24, 2026 supports a more specific conclusion than either industry camp usually offers:

AI is weakening the historical link between search visibility and website traffic, but there is not yet credible evidence that this has shifted lodging transaction share toward small suppliers selling direct. The evidence so far cuts more strongly against the “AI levels the playing field” thesis than it proves the opposite thesis of inevitable OTA consolidation. For short-term rentals in particular, the structural advantages of intermediaries extend well beyond discovery, which makes a broad direct-booking shift difficult to justify from current evidence.

That distinction matters. AI can make it easier for an unknown property to be considered without making it easier for that property to become the contractual counterparty in a transaction.

This report follows a three-layer discipline: measured evidence is kept separate from causal interpretation and from the 24-to-36-month forecast.

Layer one: what is measured. The large travel platforms have not been economically weakened during the first years of generative AI. Booking Holdings' gross bookings rose from $76.6 billion in 2021 to $186.1 billion in 2025; Airbnb's gross booking value rose from $46.9 billion to $91.3 billion; Expedia Group's gross bookings rose from $72.4 billion to $119.6 billion. Their approximate revenue-to-gross-bookings ratios remained remarkably stable over the period. Booking's rose from roughly 14.3% to 14.5%, Airbnb's from 12.8% to 13.4%, and Expedia's from 11.9% to 12.3%. These are accounting ratios rather than contractual commission rates, but there is no visible evidence in them of collapsing intermediary economics.

Nor has customer acquisition ceased to matter. Booking spent $8.2 billion on marketing in 2025, about 30.4% of revenue. Airbnb reported $2.6 billion of sales and marketing expense, about 21.1% of revenue. Expedia's direct and indirect selling and marketing expense together totaled about $8.2 billion, approximately 55.6% of revenue, although Expedia's classification is broader and includes costs such as commissions to distribution partners.

The filings have, however, changed in a revealing way. Booking's 2025 risk language explicitly discusses competition for placement in AI-generated or AI-affected search results and the resulting risk to marketing efficiency. Airbnb began warning that consumers might become less reliant on conventional search and move toward AI-mediated channels. Expedia's 2025 filing went further, warning that generative and agentic AI could change search traffic, direct-booking rates, market share and marketing expense. Earlier filings discussed AI mainly as a technology, regulatory or operational issue.

This is not proof that management expects AI to hurt the platforms. It is evidence that their lawyers and management now consider AI-mediated distribution sufficiently material to describe explicitly to investors.

Search behavior has changed more decisively than booking economics. Pew Research Center's telemetry study of 900 U.S. adults and 68,879 Google searches in March 2025 found that users clicked a conventional search result on 8% of visits when an AI summary appeared, compared with 15% when one did not. Links embedded in the AI summaries themselves were clicked in only about 1% of visits. Searches with AI summaries were also more likely to terminate without another click. Long, natural-language queries were much more likely than very short queries to produce AI summaries.

Commercial SEO vendors find a similar directional effect, although not uniformly. Ahrefs' 300,000-keyword analysis estimated a 34.5% reduction in first-position click-through rate associated with AI Overviews in March 2025 and, using the same framework, a 58% reduction by December 2025. Semrush subsequently reported a more complicated pattern: AI Overviews were expanding into commercial and transactional queries, but its same-keyword analysis did not find a simple monotonic rise in zero-click behavior.

Meanwhile, traffic directly referred from standalone AI systems remains small compared with conventional search. Similarweb estimated 1.13 billion AI-platform referrals to the top 1,000 websites worldwide in June 2025 versus 191 billion Google referrals. Semrush's much broader 2025 study across more than 50,000 sites found AI systems accounted for less than 0.15% of total visits despite rapid growth. Both are proprietary vendor estimates, not audited internet-wide measurements.

The important result is therefore not “AI sends everyone new traffic.” At present, it is closer to the opposite: AI can eliminate the click altogether.

The lodging implementations that have actually shipped also lean toward structured intermediary inventory. Booking.com's AI Trip Planner combines language models with Booking.com's own accommodation, price and availability databases. ChatGPT introduced Booking.com and Expedia as launch partners for apps inside ChatGPT. Expedia has integrated with AI agents, is building an AI toolkit and composable distribution platform for third parties, and in July 2026 acquired AI-native travel planner Layla specifically to combine its conversational interface with Expedia's supply, booking technology and marketplace.

General-purpose agents are also becoming capable of operating websites. OpenAI's standalone Operator became ChatGPT agent in July 2025 rather than remaining a separate product. Google's Gemini Spark can use Chrome to research options and begin booking processes, but Google explicitly keeps the user involved in sensitive actions such as payments. Google's August 2026 connected-app announcement names partners for experiences, car rental, restaurants and event tickets, but not a general hotel-supply partner.

These developments demonstrate technical ability to route around an OTA website. They do not demonstrate an economic ability to route around the functions the OTA performs.

Layer two: the mechanism. Lodging intermediation bundles at least seven services that are too often treated as though they were one: discovery, inventory aggregation, real-time availability, reputation and identity signals, payment processing, fraud and chargeback management, and post-booking recourse. AI directly attacks the information costs involved in the first few. It does not, merely by becoming better at language or browsing, create a balance sheet willing to absorb a disputed stay, an insurance product, a re-accommodation operation, a verified payment counterparty or a worldwide inventory synchronization network.

That distinction is considerably more important for short-term rentals than for branded hotels. A traveler deciding whether to book a Marriott directly already has a brand, operating company, loyalty program and known party against which to seek recourse. A traveler booking an individual apartment from an unfamiliar host often does not. Airbnb's AirCover, Booking's partner protections and Expedia/Vrbo's marketplace controls exist precisely because information about a property is not equivalent to confidence that the transaction will be honored. Airbnb, for example, describes host damage protection of up to $3 million per stay, host liability insurance and guest protections for serious booking problems. Booking describes liability coverage for eligible accommodation partners, while Expedia explicitly identifies attempts to move Vrbo travelers into off-platform payment as both a revenue leakage and fraud problem.

The closest natural experiments also fail to show that removing an incumbent gatekeeper's search advantage automatically sends demand direct.

Google changed European travel-search presentation beginning in early 2024 as part of its Digital Markets Act compliance. Mirai, a hotel direct-booking technology vendor, initially measured 3,450 hotels and reported a 30% drop in Google Hotel Ads clicks and a 36% decline in bookings attributed to that channel in DMA markets relative to non-DMA markets. Its later analysis of more than 3,000 properties over a longer before-and-after period found something much less dramatic: hotels recaptured most of the lost Google Hotels contribution through organic search, paid search and other metasearch, leaving an estimated net decline of only 0.8% in direct reservations. Mirai argued that the remaining volume probably moved toward OTAs, but it did not observe OTA bookings directly.

This experiment is imperfect. Mirai sells direct-booking technology, attribution is observational, and Google's compliance implementation continued to change. Indeed, the European Commission concluded in July 2026 that Google was still favoring its own vertical services, including hotel results, and fined the company as part of an €890 million DMA enforcement action.

Still, the result is informative. Removing or degrading one gatekeeper interface did not produce an observable surge in supplier-direct volume. Demand rerouted through other parts of an already intermediated system.

Rate-parity reforms tell a similar story from another direction. Since November 14, 2024, Booking.com has been prohibited under the DMA from using parity clauses in the European Economic Area, leaving hotels free to offer better prices or conditions on their own channels. Academic natural experiments show that loosening parity can make direct prices cheaper. Ennis, Ivaldi and Lagos found direct sales became relatively cheaper for mid-level and luxury hotels after European interventions. A 2026 study using synthetic-control methods found restrictions on broad parity reduced prices by about 1.5% and raised occupancy by about one percentage point, while complete parity bans reduced prices by between zero and 4% but had no detectable occupancy effect. Neither result establishes a comparably large shift in direct-booking share.

The lesson is uncomfortable for the strongest version of the direct-booking thesis: giving a supplier the legal ability to undercut the intermediary does not ensure that consumers will bypass the intermediary. That is the same distinction that shows up when operators decide how to price a direct site against OTA totals: contractual freedom and channel choice are not the same thing.

Layer three: what follows. My highest-probability outcome through roughly August 2028 to August 2029 is not a dramatic channel reversal. I assign a 45% probability to broad continuity, in which AI becomes a major discovery and shopping interface but direct-versus-intermediated transaction shares move only modestly. I assign 35% to an OTA-infrastructure scenario, in which assistants increasingly become front ends to inventory, payment and servicing supplied by Booking, Expedia and other scaled intermediaries. A 15% probability goes to meaningful direct-channel gains, concentrated in branded hotels, professional vacation-rental managers and repeat guests rather than atomized STR supply. The remaining 5% goes to a genuinely new AI-native intermediary that contracts supply, owns checkout and assumes enough servicing liability to become a new bargaining center rather than merely an interface.

Those probabilities are judgments, not measurements.

The strongest near-term prediction is narrower and carries higher confidence: AI will separate discovery from transaction more than previous interfaces did. The company whose model recommends the property will increasingly not be the company that provides its inventory or bears the booking risk. That development favors companies capable of becoming invisible infrastructure.

For accommodation operators, the practical implication is consequently not “abandon direct” and not “AI will make OTAs unnecessary.” The rational strategy is to make the property technically legible to AI — the work described in how to get a vacation rental recommended by ChatGPT — while strengthening the things that make direct purchase rational after discovery: authoritative structured inventory, unique direct benefits, transparent cancellation terms, a high-performing checkout, first-party customer relationships, and credible post-booking support. For independent STR operators, direct-booking investment should be judged against repeat demand and trust acquisition costs, not against the assumption that AI will provide free top-of-funnel demand.

The central empirical gap remains large. There is no credible public dataset showing AI-originated lodging transactions divided among supplier-direct, OTA and other channels. Until one exists, claims that AI has already shifted bargaining power are forecasts being narrated as measurements.

Research question, assumptions, and method

The principal research question is:

As conversational and agentic AI becomes a material interface for discovering, comparing and booking lodging, which party gains bargaining power: accommodation suppliers selling direct, existing online travel agencies, or a new intermediary layer?

“Bargaining power” is used here more narrowly than general corporate power. It means the ability to capture or retain demand while influencing distribution cost, customer ownership, pricing discretion and the terms under which inventory reaches travelers. A supplier can gain visibility without gaining bargaining power. Likewise, an OTA could lose front-end brand prominence while gaining economic importance if it becomes the wholesale inventory and servicing layer behind an AI assistant.

The geographic focus is principally the United States and Europe because they offer the strongest public financial, regulatory and behavioral evidence. The accommodation scope includes hotels and short-term or vacation rentals, but they are treated separately where their economics differ. The forecast horizon is 24 to 36 months from the research cutoff of August 24, 2026.

Layer one: measured evidence. The baseline was built primarily from SEC filings by Booking Holdings, Airbnb and Expedia Group, using annual data for fiscal years 2021 through 2025. Gross bookings or gross booking value, revenue and marketing expense were taken from company disclosures. “Implied take rate” in this report means simply:

Implied revenue ratio = reported revenue ÷ reported gross bookings or GBV

It should not be mistaken for the contractual commission charged on a particular reservation. Revenue recognition, merchant versus agency transactions, taxes, incentives, ancillary revenue and timing differ across companies. The ratio is useful longitudinally within a company and only cautiously across companies.

Marketing measures are similarly non-identical. Booking Holdings reports “marketing expense”; Airbnb reports a broader “sales and marketing” line; Expedia separates direct and indirect selling and marketing, with categories that include consumer acquisition and certain partner commissions. The percentages below are therefore trend measures rather than proof that one company pays exactly twice another's customer-acquisition cost.

Risk factors were compared across successive 10-Ks to identify when AI moved from a generalized technology risk into distribution-specific language. That method has an important limitation: a risk factor is not an admission that management expects the event to happen. It establishes that management judged the risk sufficiently plausible and material to disclose.

The search analysis prioritizes Pew's opt-in browsing telemetry because the underlying observations are human browsing behavior rather than survey recollection. Vendor datasets from Ahrefs, Similarweb and Semrush are included because they cover far larger keyword or traffic sets, but their proprietary sampling and commercial interests are stated rather than suppressed.

Natural experiments were given greater weight than industry commentary. The two most useful are the European DMA changes to search and booking platforms, and European restrictions on hotel price-parity clauses. The parity literature includes peer-reviewed work using transaction data and quasi-experimental methods. The DMA traffic evidence is weaker because the most specific hotel-level study available publicly comes from Mirai, a company whose business is increasing hotels' direct sales.

What could not be measured reliably is itself important. There is no standardized, public, representative time series for vacation-rental direct-booking share. There is no audited split of AI-assisted lodging transactions by final channel. The large platforms do not disclose a common paid-versus-unpaid traffic metric. Commercial terms between major AI assistants and lodging distribution partners are generally not included in public launch announcements. Expedia and Booking also do not publish a clean standalone P&L for vacation rentals that can be compared directly with Airbnb.

Those absences materially constrain the forecast.

Distribution baseline before the AI shift

Layer one: measured. The starting point is a travel-intermediation sector whose largest firms expanded substantially from 2021 to 2025 and whose revenue capture relative to gross booking value remained fairly stable.

Company	FY	Gross bookings / GBV	Revenue	Revenue ÷ gross	Marketing / S&M	Marketing / revenue
Booking Holdings	2021	$76.6B	$11.0B	14.31%	$3.80B	34.7%
Booking Holdings	2022	$121.3B	$17.1B	14.09%	$5.99B	35.1%
Booking Holdings	2023	$150.6B	$21.4B	14.18%	$6.77B	31.7%
Booking Holdings	2024	$165.6B	$23.7B	14.34%	$7.28B	30.7%
Booking Holdings	2025	$186.1B	$26.9B	14.46%	$8.19B	30.4%
Airbnb	2021	$46.9B	$6.0B	12.78%	$1.19B	19.8%
Airbnb	2022	$63.2B	$8.4B	13.29%	$1.52B	18.1%
Airbnb	2023	$73.3B	$9.9B	13.54%	$1.76B	17.8%
Airbnb	2024	$81.8B	$11.1B	13.58%	$2.15B	19.4%
Airbnb	2025	$91.3B	$12.2B	13.41%	$2.59B	21.1%
Expedia Group	2021	$72.4B	$8.6B	11.87%	$4.22B	49.1%
Expedia Group	2022	$95.0B	$11.7B	12.27%	$6.10B	52.3%
Expedia Group	2023	$104.1B	$12.8B	12.34%	$6.86B	53.5%
Expedia Group	2024	$110.9B	$13.7B	12.34%	$7.63B	55.7%
Expedia Group	2025	$119.6B	$14.7B	12.32%	$8.19B	55.6%

Booking figures come from successive 10-K disclosures; Airbnb's 2021-2023 figures can be reconstructed from its 2022 and 2023 filings and 2024-2025 from the 2025 filing; Expedia's 2021-2022 series comes from its 2022 filing and 2023-2025 figures from the 2025 filing.

The most striking feature is not a rising take ratio. It is stability. None of the three series exhibits the compression one would expect if suppliers had already gained substantial pricing leverage over their principal platforms.

Booking is unusual in providing a quantitative signal about its own acquisition channel. In 2025 it said the proportion of room nights booked through its direct channel was in the mid-fifties percent range and increased year over year. Its marketing expense as a percentage of revenue has also fallen from the 2021-2022 level, which management has partly associated with increasing direct mix and marketing efficiency.

That “direct” metric is direct to Booking.com, not direct to the accommodation supplier. This distinction is easily missed. One of Booking's most important strategic accomplishments has been disintermediating Google and other paid acquisition sources without allowing the hotel to disintermediate Booking.

Airbnb and Expedia do not disclose a comparable standardized percentage. Expedia warns that reduced direct traffic can raise acquisition costs, and its 2025 filing continues to describe search engines and other marketing channels as important sources of consumer demand.

The evolution of AI language in the filings is more informative than any executive prediction.

Company	Earlier filings	Latest distribution-specific language	Interpretation
Booking Holdings	2023 focused substantially on risks from building, using and regulating AI systems.	2025 explicitly warns that competition for placement in AI-generated or AI-affected search results can raise cost per click and reduce marketing efficiency.	AI has become a customer-acquisition risk, not merely a technology risk.
Airbnb	2023 discussion centered more heavily on AI technology, legal and operational considerations.	By 2024-2025, Airbnb warns that consumers may become less reliant on search engines and use AI applications and other channels; it separately identifies hosts booking guests directly as a platform risk.	Airbnb now links AI-mediated discovery and supplier disintermediation to demand acquisition.
Expedia	2024 AI language was less specific to channel displacement.	2025 warns that rapid adoption of generative and agentic AI may shift consumers toward AI-driven platforms, reduce direct-booking rates, increase marketing cost and affect market share.	Among the three, Expedia states the possible channel consequence most plainly.

The progression is visible in the companies' own filings.

The short-term rental measurement problem

The market is substantially less measurable once the question moves from public-company economics to STR channel share.

Hostfully's 2022 industry survey, reported by Skift, covered 375 property managers and hosts, approximately three quarters of whom were in the United States. It reported direct bookings plus referrals at about 19%, versus 24% in 2020 and 21% in 2021. The construction matters: direct and referral were combined, the sample was a vendor-industry survey rather than a representative census, and operator geography was highly skewed toward the United States.

Lodgify later publicized a figure approaching 34% for bookings made through direct-booking sites in its data. That number is widely repeated, but Lodgify sells property-management and direct-booking software, and the publicly available presentation does not establish that its customer panel represents all vacation-rental inventory. It should be read as the behavior of a software-vendor ecosystem, not “the global direct-booking share.”

More recent operator surveys remain inconsistent. A 2026 PhocusWire summary of Hospitable research reported that 38% of surveyed hosts said they received no direct bookings in 2025 and another 48% received only 1% to 10% direct. The same article cited Hostaway research in which 70% of operators had a direct-booking website, yet 62% received less than one quarter of bookings direct and 18% reported none.

A Key Data panel reported through industry media points in the same general direction while producing different absolute values: direct represented roughly 23% of reservations and 35% of revenue in the first quarter of 2026, versus 26% and 39% respectively a year earlier. Because Key Data aggregates participating professional managers rather than all hosts, that too is a panel statistic rather than an industry census.

Skift Research, meanwhile, has estimated that Airbnb, Booking and Expedia/Vrbo together accounted for 71% of global STR bookings in 2024, up from 53% in 2019. The underlying commercial research is paywalled and its full sampling and estimation procedure is not public, so the figure is useful as a vendor estimate, not as an audited market fact.

The defensible finding is therefore not that STR direct share is 10%, 23% or 34%. It is that nobody has published a sufficiently transparent, representative and longitudinal dataset to establish the number with confidence.

That is a material weakness in much of the industry argument. Assertions that AI will double a channel whose current share is itself uncertain carry false precision.

Layer two: mechanism. The hotel analogy understates this problem because hotel direct distribution rests on accumulated brand capital. Major chains can use an OTA for discovery and subsequently persuade a traveler to transact at brand.com using loyalty points, member rates, existing accounts and confidence in post-booking support. Expedia itself identifies hotel chains' use of loyalty programs and lower direct rates as an important competitive force.

The economics of an individual vacation rental differ. A property may have excellent photographs, reviews and a well-structured website and still lack the independent reputation necessary for a traveler to send thousands of dollars directly. That is the conversion problem described in what builds trust on a new direct booking website. The OTA's commission therefore buys something other than traffic. It buys a recognizable counterparty.

Layer three: what follows. A meaningful AI-induced rise in supplier-direct bookings should therefore appear first in categories where the supplier already has an identity independent of the intermediary: chain hotels, destination resorts, professional multi-property STR managers, branded apartment operators, and repeat-guest businesses. A surge among one-property hosts should require a second development in addition to AI discovery, namely a scalable way to provide payment protection, verification and recourse without the OTA.

Confidence: high on the ordering, low on the eventual magnitude.

A falsifier by August 2027 would be credible transaction-panel data showing that atomized one-to-few-property STR hosts are gaining direct share at least as quickly as branded and professional lodging suppliers, without an accompanying third-party trust or guarantee layer.

What AI has actually changed in search and booking

Layer one: measured. The best independent behavioral evidence is Pew's July 2025 analysis of browser telemetry. Its sample consisted of 900 U.S. adults who agreed to have browsing activity measured during March 2025. Across 68,879 distinct Google searches, 12,593 produced an AI-generated summary. Conventional result links were clicked in 8% of visits with an AI summary and 15% without one. Only 1% of visits containing an AI summary produced a click on a source link inside the summary. Users also ended their browsing session after 26% of AI-summary searches compared with 16% of searches without a summary.

Pew also found strong query-length sensitivity. AI summaries appeared for 53% of searches containing ten words or more but only 8% of searches containing one or two words.

That difference is directly relevant to conversational travel planning. “Four-bedroom beach house in Kiawah Island June 12” is structurally more exposed to synthesized answers than “Marriott Marquis,” although Pew did not isolate accommodation queries and therefore cannot prove that exact travel effect.

Pew's broader March 2025 telemetry also found that 13% of the 900 participants visited an AI chatbot during the month, rising to 20% among adults aged 18 to 29. This documents meaningful usage but not substitution of AI for travel search.

Commercial datasets reinforce the click-reduction concern but disagree on magnitude and evolution. Ahrefs selected 150,000 keywords that produced an AI Overview and 150,000 informational keywords that did not, then used aggregated Google Search Console data to compare click-through rates before and after rollout. Its April 2025 estimate was a 34.5% reduction in position-one CTR associated with an AI Overview. Repeating its method using December 2025 data produced an estimated 58% reduction.

That is not a randomized experiment. Ahrefs constructed a counterfactual using matched categories and historical CTR changes. The initial sample was also overwhelmingly informational because nearly all AIO-triggering terms in its sample were informational at that time.

Semrush's measurements complicate the story. Its 2025 analysis reported that the proportion of tracked keywords showing AI Overviews rose sharply during the year before falling from its summer peak. More significantly for lodging, the composition broadened: informational queries represented 91.3% of AIO-triggering queries in January 2025 but 57.1% by October, as commercial, transactional and navigational exposure expanded. Its same-query comparisons did not find the straightforward increase in zero-click behavior implied by some other studies.

By 2026, Semrush found AI Overviews increasingly coexisting with paid ads and analyzed more than 600,000 U.S. desktop keywords to document their expansion into commercially valuable searches.

The apparent disagreement is partly methodological. Pew measures actual browsing sessions. Ahrefs focuses heavily on CTR to the first conventional organic result. Semrush classifies keyword and traffic behavior across its proprietary databases. A result can reduce clicks to position one while still generating a click elsewhere. None of those measures is the same as “AI causes fewer hotel bookings.”

Direct chatbot referrals remain far smaller than search traffic. Similarweb estimated that AI platforms generated about 1.13 billion referral visits to the top 1,000 websites in June 2025, up sharply year over year, while Google generated approximately 191 billion. Semrush's April 2026 analysis of more than 50,000 websites across 17 industries found AI referral traffic growing 66% during 2025 but accounting for less than 0.15% of total visits; organic search still generated more than one trillion visits in its measured dataset.

The studies therefore support two statements that can coexist:

1. AI-generated search answers can materially reduce traditional outbound clicks.
2. Standalone AI systems have not yet replaced conventional search as a source of website referrals at comparable scale.

Neither statement proves which booking channel captures the resulting transaction.

What the assistants actually do

The implementation evidence is unusually revealing because it exposes where the hard infrastructure resides.

Booking.com's AI Trip Planner, initially launched in 2023 and developed using OpenAI models, maps natural-language requests onto Booking.com's proprietary structured data and retrieves real-time accommodation availability and prices from Booking's database.

That architecture is almost the reverse of the leveling thesis. The intelligence layer is general-purpose; the commercially useful real-time inventory layer remains aggregated.

OpenAI subsequently made Booking.com and Expedia launch partners in the app ecosystem introduced inside ChatGPT in 2025. Apps can respond within conversations and expose interactive interfaces rather than requiring a user to begin on the partner's own domain.

Expedia has pursued the same idea from the supply side. In 2025 it announced integrations with OpenAI Operator and Microsoft Copilot Actions. In 2026 Expedia B2B previewed an AI toolkit and “Intelligent Experience Platform” intended to let outside partners incorporate Expedia supply and capabilities into their own AI workflows. On July 31, 2026, Expedia acquired Layla, an AI-native conversational trip planner, explicitly describing the combination of Layla's interface with Expedia's supply, first-party data, marketplace and booking technology.

The direction of travel is therefore not only “OTAs add chatbots.” It is also “OTAs make themselves available as infrastructure to other chatbots.”

General-purpose browser agents provide a different route. OpenAI introduced Operator in January 2025 as a browser-using agent, then folded its capabilities into ChatGPT agent that July and announced the standalone Operator product would be sunset. This reversal is instructive. Product form remains fluid even when the underlying agent capability persists.

Google is pushing more aggressively into web action. Gemini Spark's July 2026 Chrome integration can use logged-in accounts, research flight options and begin the booking process. Google nonetheless says sensitive actions such as payments are returned to the user. In August 2026, Google announced Gemini connected apps for services including GetYourGuide, Localiza, OpenTable and Ticketmaster, covering experiences, rental cars, dining and events. A generic accommodation transaction partner was not among the named integrations in that announcement. That is a different surface from Google Vacation Rentals, which already routes structured lodging inventory through approved connectivity partners rather than letting individual operators list themselves.

Perplexity demonstrates that an AI company can move into checkout in retail. Its Instant Buy system, launched with PayPal in November 2025, permits U.S. consumers to purchase eligible merchandise through Perplexity, while the merchant fulfills the order and remains merchant of record. In the official Perplexity materials reviewed for this report, I did not find an equivalent disclosed accommodation transaction partnership.

The commercial terms between the AI companies and travel partners are generally not disclosed in these announcements. Consequently, there is presently no public evidentiary basis for claims about whether ChatGPT, Google or another assistant receives hotel commissions at a particular rate, whether inventory partners pay for ranking, or whether eventual economics will resemble affiliate fees, advertising, API fees or merchant commissions.

Layer two: mechanism. These product choices suggest a division of labor.

Large language models are becoming good at converting an imprecise human request into structured intent: dates, party size, geography, amenity preferences, price sensitivity and trip purpose. An OTA has spent decades solving the inverse problem, converting millions of fragmented supplier records into normalized inventory that can answer those structured requests.

Connecting the two is cheaper than recreating the second system from scratch.

AI browsing agents weaken this argument somewhat because they can interact with supplier websites without a formal API. But browser automation does not remove inventory staleness, rate-plan inconsistencies, cancellation semantics, fraud, payment disputes or the need to recover from a failed booking. The fact that Google keeps the user involved at sensitive payment stages is an example of the difference between technical navigation and assumption of transactional responsibility.

Layer three: what follows. I have high confidence that AI will reduce the strategic value of ranking first on an ordinary blue-link results page. I have only medium confidence that it will materially reduce OTA transaction share because the two effects are not equivalent.

The twelve-month falsifier is straightforward. By August 2027, independent accommodation analytics should show that AI-originated visitors to lodging suppliers convert to direct transactions at enough scale to offset lost conventional organic traffic. If AI referrals remain tiny while outbound search clicks continue falling, then better “AI visibility” will not have become a direct-distribution engine.

Natural experiments and prior disintermediation cycles

The strongest case against forecasting from intuition is that travel distribution has already experienced several moments in which the dominant discovery interface changed.

Layer one: measured.

The Digital Markets Act

The European Union's DMA requires designated gatekeepers not to prefer their own services unfairly. The Commission opened proceedings against Alphabet in March 2024 over concerns that Google Search favored vertical services including Google Hotels.

Google made substantial European search changes during early 2024. Mirai's initial study compared traffic for 3,450 hotels in DMA and non-DMA markets from January through April and reported that Google Hotel Ads traffic in affected EU markets fell approximately 30%, with bookings attributed to the direct channel through those ads falling 36%.

Taken alone, that sounds like a large blow to direct distribution.

Mirai then broadened the measurement. Its October 2024 follow-up used Google Analytics 4 data, more than 3,000 properties and eight months before and after implementation. Google's hotel-metasearch contribution to direct bookings fell from 13.4% to 8.9%, but hotels recovered most of the loss through paid Google traffic, organic search and other metasearch. Mirai calculated the residual effect on total direct reservations at roughly negative 0.8%.

That follow-up is more consequential than the initial headline. A structural interface change produced severe disruption within one direct-acquisition channel but almost no aggregate change in direct bookings because demand moved elsewhere.

Mirai inferred that some unrecovered traffic probably went to OTAs. Its data did not directly observe those OTA transactions, so that part should remain an inference rather than a measured fact.

The experiment also never reached a stable policy state. Google subsequently proposed additional European result changes, and the Commission ultimately found in July 2026 that Google continued to give preferential treatment to its own services, including hotel results. It imposed €460 million of the €890 million announced DMA penalties for the Search self-preferencing violation.

This makes the DMA evidence less like a controlled laboratory intervention and more like a moving natural experiment. It remains valuable because the immediate hotel response falsified the simplest theory: weaken Google's existing travel interface and direct suppliers automatically gain.

They did not.

Rate parity

Booking.com was designated a DMA gatekeeper on May 13, 2024. From November 14, 2024 it was required to comply with obligations including a prohibition on parity clauses. The Commission specifically stated that hotels and other travel providers using Booking.com could offer better prices or conditions on their own websites or other channels without being punished by increased commissions or delisting.

Separately, on September 19, 2024, the Court of Justice of the European Union ruled in Case C-264/23 that Booking.com's price-parity clauses could not be treated as restraints objectively necessary to the operation of the platform.

The empirical literature is stronger than the legal commentary.

Ennis, Ivaldi and Lagos used hotel-chain transaction data and European policy changes as natural experiments. They found that after broad parity obligations were restricted, direct hotel sales became relatively cheaper than OTA sales for mid-level and luxury hotels. More extensive elimination of parity in France and Germany produced substantially less uniform additional effects.

A 2026 paper in the International Journal of Industrial Organization used a synthetic-control strategy and transacted price and occupancy data. It estimated that restricting broad parity while leaving narrower parity reduced hotel prices by about 1.5% and increased occupancy by roughly one percentage point. Relative to narrow parity, complete bans reduced prices by between zero and 4% but produced no detectable occupancy effect.

The policy did what direct-channel advocates said should be valuable: it increased suppliers' ability to differentiate price.

What has not been established is a large resulting migration of transactions to supplier websites.

That gap matters because it separates supplier freedom from consumer channel choice. A hotel may gain the right to undercut Booking.com and use it. The traveler can still choose Booking.com because loyalty, convenience, stored payment credentials, comparison, support or perceived recourse are worth the price difference.

Previous “disintermediation” moments

OTA economics have long contained an apparent contradiction. Hotels dislike intermediary commissions, yet intermediaries can create incremental demand even when the final booking happens direct.

Cornell's early work on the “billboard effect” found evidence that hotel visibility on an OTA could increase bookings on the hotel's own site. One expanded study examined 1,720 InterContinental Hotels Group transactions and found substantial overlap between OTA exposure, search behavior and eventual brand-site purchase.

A later Cornell analysis concluded in 2017 that the billboard effect had persisted despite hotels' increasingly aggressive direct-booking campaigns. Its observed hotel shopper visited many travel-related pages before purchase, reinforcing the view that discovery and transaction channels cannot be cleanly assigned to one another.

Metasearch was supposed to strengthen direct sales by presenting a hotel's own rate alongside OTA offers. Research has nevertheless shown that positioning within metasearch continues to affect clicks, meaning an apparently neutral comparison layer creates another scarce ranking surface rather than eliminating distribution competition.

The mobile transition likewise did not make aggregators obsolete. Instead, OTAs built high-quality mobile apps that combined information, communication, payment and servicing in one interface. Contemporary academic work continues to treat mobile OTA reservation as a major consumer channel rather than a historical residue.

Layer two: mechanism. These precedents share a pattern. New interfaces reduce one kind of friction but often increase the value of another intermediary function.

Metasearch lowered price-comparison cost, but it created competition for metasearch position.

Mobile lowered the friction of reaching supplier websites, but it increased the value of saved credentials, app notifications and compact multi-property comparison.

Rate-parity reform increased hotels' pricing freedom, but it did not eliminate the consumer's reasons for preferring an intermediary.

The DMA weakened particular Google travel placements, but users largely rerouted through other search and comparison mechanisms rather than moving wholesale to hotel-owned channels.

AI is more capable than any of those technologies, so the analogy should not be treated as destiny. What history does remove is the burden-free assumption that lower information costs necessarily transfer transaction power to suppliers.

Layer three: what follows. The historical base rate raises the threshold for believing in a major direct shift. A forecast of substantial disintermediation should identify a function that AI eliminates which previous technologies did not merely relocate.

Discovery alone does not meet that threshold.

A real break with history would be visible within twelve months if assistants begin to normalize direct supplier availability, cancellations and checkout across thousands of independent properties without requiring an OTA or comparable wholesaler. That would be a qualitatively different event from generating better recommendations.

Confidence: medium-high.

Structural economics: what discovery does not solve

The decisive question is not “Can an AI find a property?” It is “What organization must exist between the moment the AI recommends the property and the moment the traveler safely completes and, if necessary, unwinds the transaction?”

Layer one: measured. Current intermediaries have built substantial machinery around this problem.

Airbnb describes a trust system incorporating reviews, account protection, risk scoring, secure payment, fraud and scam prevention, booking restrictions and safety support. AirCover includes host damage protection up to $3 million for eligible property damage, host liability insurance up to $1 million per occurrence and guest assistance when serious reservation problems arise.

Booking Holdings' filings likewise discuss protections for accommodation partners and liability insurance of up to approximately $1 million per occurrence for eligible partners in relevant jurisdictions. The company also notes that many alternative accommodations are individual or small operators and that it cannot systematically verify every property's safety, quality or legal compliance.

Expedia reports more than 3.6 million lodging properties on its platforms as of 2025, including roughly 2.4 million online-bookable alternative accommodation listings associated with Vrbo. Its risk discussion specifically identifies fraud situations in which travelers are induced to pay outside Expedia's systems.

These are not search-engine features.

They are transaction infrastructure.

Intermediary function	What AI can plausibly do	What still requires infrastructure	Likely near-term effect
Demand discovery	Interpret open-ended intent; search and synthesize many sources.	Ranking policy, source freshness and commercialization still matter.	Strong potential to weaken traditional search advantage.
Inventory aggregation	Query APIs or browse multiple websites.	Normalization, duplicate resolution, availability freshness and rate-plan semantics.	Likely dependence on aggregators for broad real-time supply.
Trust and identity	Summarize reviews and external reputation.	Verified identity, anti-fraud systems and contractual accountability.	AI improves information but does not itself create recourse.
Payment	Fill forms or invoke payment systems.	Merchant relationships, authentication, compliance, settlement and chargebacks.	Technically automatable; economically non-trivial.
Refunds and disputes	Automate intake and policy interpretation.	Authority to return money and absorb unrecoverable losses.	Requires a financially accountable counterparty.
Damage / insurance	Help assess evidence or administer claims.	Insurance underwriting and balance-sheet risk.	AI is a tool for the function, not a substitute for the risk bearer.
Cancellation / rebooking	Search for alternatives rapidly.	Inventory access, compensation rules, customer service and funding.	Favors broad inventory owners when disruption occurs.

Layer two: mechanism.

Follow the loss

Consider the hardest case, because it reveals the channel's actual structure.

A traveler tells an assistant to book an independently operated villa. The villa's website shows availability, the assistant completes the form, and the traveler pays $4,000. On arrival the villa is occupied, fraudulent, unsafe or materially different from the listing.

Who returns the $4,000?

If the assistant merely operated the supplier's website as the traveler's tool, contractual liability may remain principally with the supplier and whatever payment protections apply. The assistant has displaced search and form filling, but not intermediation in the economically important sense.

If an OTA supplied the listing and booking rails, the OTA can administer the refund or re-accommodation according to its contractual model, then seek recovery from the supplier. Its ability to do so rests partly on scale, supplier contracts, payment control and access to replacement inventory.

If the AI company promises the traveler that it will make the situation right, it has crossed a line. It now needs a claims policy, fraud controls, supplier agreements, customer support, payment authority and capital or insurance sufficient to absorb cases in which the supplier cannot reimburse it.

That is possible. It is simply a different business from answering travel questions.

This is the central reason that “AI can browse any website” is an incomplete distribution thesis. A browser agent can make the booking. The difficult economic question begins when the booking fails.

Why hotels and STRs diverge

A Hilton or Marriott can itself occupy the trusted-counterparty position. Its brand gives the assistant a direct endpoint with known fulfillment and support. The hotel chain can also maintain centralized availability, loyalty data, payment systems and customer-service operations.

An individual short-term rental cannot replicate those functions at the same unit cost. AI can verify some external signals and potentially reduce fraud through better cross-checking, but it cannot guarantee that a dispersed owner has enough money to refund a guest or that replacement inventory exists on a sold-out weekend.

This predicts a bifurcated market. AI can be pro-direct without being pro-small-supplier.

Commercial incentives

The final missing piece is monetization.

Booking spent $8.2 billion on marketing in 2025. Expedia spent a similarly large amount under its much broader selling-and-marketing definition. Those budgets reflect the economic value of controlling high-intent demand.

There is no reason to assume that conversational interfaces will remain economically neutral simply because their output looks different from a search-results page. AI companies face compute costs and business-model incentives. They can potentially monetize travel through subscription, advertising, referral payments, transaction fees, partner economics or combinations of these. The commercial terms of the lodging integrations reviewed here are not public, so choosing among those models today would be speculation.

The procurement incentive is clearer. An assistant that wants to serve millions of lodging queries can integrate once with an entity offering millions of properties, normalized availability, support and payment capability, or negotiate and maintain interfaces with a fragmented supplier base. Expedia's B2B strategy explicitly seeks to make the first option easier.

That does not guarantee that Expedia or Booking wins. Open standards, channel managers and property-management systems could aggregate supplier-direct feeds without becoming consumer-facing OTAs. This is one of the most important open questions because such infrastructure could preserve direct supplier economics while solving AI's fragmentation problem.

Layer three: what follows. The decisive leading indicator is therefore not the number of hotels cited in AI answers. It is the contractual architecture beneath AI checkout.

Watch who supplies the inventory, who charges the card, whose terms the traveler accepts, who handles cancellation, and who writes the check when fulfillment fails.

Confidence: high that these variables will predict bargaining power better than citation share or AI referral traffic.

The forecast would be wrong if assistants achieve meaningful lodging volume while consistently routing payment directly to fragmented suppliers and still provide credible guarantees through an independent insurance or payment mechanism that leaves the OTA outside the transaction.

Scenarios, indicators, limitations, and source notes

The scenarios below are deliberately different in mechanism rather than being optimistic, base and pessimistic versions of the same trajectory. Probabilities are subjective assessments based on the measured evidence above. They sum to 100% because they describe the dominant distribution regime at the end of the 24-to-36-month horizon, not every feature that may coexist.

Scenario	Assessed likelihood	Core mechanism	What must become true	12-month falsifier
Interface changes, channel shares mostly hold	45%	AI becomes important for discovery, but suppliers, OTAs and search platforms all adapt; transactions continue through familiar channels.	Consumers value AI planning without demanding that the assistant become merchant or guarantor. Direct and OTA integrations both proliferate.	Credible channel data shows a sustained shift of more than roughly five percentage points toward direct or intermediary share attributable to AI.
OTAs become the infrastructure behind AI	35%	Assistants outsource inventory, availability, checkout and servicing to scaled travel platforms; OTA brands become less visible while their transaction role grows.	Booking, Expedia and peers expose attractive agent-ready APIs; assistants favor reliable supply and fulfillment over web-wide direct crawling.	Major assistants instead establish direct supplier-feed standards plus independent payments and guarantees, with little OTA participation.
Selective supplier-direct resurgence	15%	AI materially lowers discovery cost, channel managers expose structured direct inventory, and supplier websites offer differentiated pricing or benefits.	Assistants reliably compare direct rates and cancellation terms; payment/trust services become available independently of OTAs.	AI-originated supplier traffic remains negligible, direct share is flat, or assistants increasingly source hotel results through OTA inventory.
A new AI-native intermediary captures the transaction	5%	An AI company stops acting principally as an interface and becomes the booking counterparty or comprehensive agent, contracting inventory and owning service recovery.	It develops seller-of-travel compliance where required, payment rails, supplier contracting, refund authority, support and a guarantee product.	By August 2027 the major AI assistants still redirect, browser-automate, or depend on established travel partners rather than assuming transaction liability.

Continuity is the highest-probability scenario

Layer three: forecast, confidence medium. AI can alter the top of the funnel much faster than it changes mature transaction relationships. In this scenario, conversational discovery becomes commonplace and ordinary search loses traffic, but hotels continue to cultivate direct customers, OTAs continue to aggregate comparison shoppers, and STR travelers continue to buy substantial trust and recourse from marketplaces.

The evidence supporting continuity is unusually concrete. Large OTA economics remain stable. Booking has successfully increased the share of consumers who arrive directly at Booking itself. Rate-parity reform changed relative pricing without demonstrating a wholesale channel migration. Google's DMA changes caused large movement inside the direct-acquisition mix but only a small measured net effect on total direct bookings in Mirai's eventual analysis.

The evidence against continuity is that AI interfaces are spreading faster than earlier distribution technologies and increasingly possess agentic capability. Google AI Mode, Gemini, ChatGPT agent and in-chat apps can compress search, comparison and action into a single session in ways metasearch never did.

The falsifiable claim is that, by 2028-2029, aggregate lodging-channel shares remain within a few percentage points of their pre-agentic-AI trajectory even though the source of discovery traffic changes substantially.

OTA infrastructure gains are the strongest directional alternative

Layer three: forecast, confidence medium. In this scenario, the consumer increasingly thinks they are “booking with an assistant,” while the economic plumbing belongs to Booking, Expedia or another travel intermediary.

The causal chain is:

AI captures conversational intent → the assistant needs trustworthy live lodging supply → a handful of travel platforms can provide millions of normalized inventory records through one integration → those platforms also supply payment and post-booking servicing → the assistant externalizes expensive travel-specific operations → OTA bargaining power over suppliers persists or increases even if OTA consumer brands receive fewer initial visits.

Booking.com's integration of language models with its proprietary availability database and Expedia's explicit effort to expose its supply through agent-oriented infrastructure support this mechanism. Expedia's acquisition of Layla further demonstrates that the travel platform itself views conversational discovery and existing supply infrastructure as complementary assets.

What cuts against it is the growing competence of browser automation. If an agent can query direct sites reliably, the supplier does not need to participate through one of the large OTAs simply to be technically reachable.

The key unknown is whether browser reachability is commercially sufficient. My forecast assumes it often will not be because availability freshness, support and failure recovery matter disproportionately once the system moves from recommendation to confirmed travel.

A decisive 12-month indicator would be the announcement of large-scale hotel inventory agreements between major AI assistants and Booking, Expedia, Airbnb, wholesalers or channel-management networks, particularly when the partner also handles modifications and servicing.

Direct gains are possible, but likely concentrated

Layer three: forecast, confidence medium-low. Direct distribution has a credible path when three conditions coincide: the supplier has a trusted identity, it exposes authoritative inventory in a form assistants can reliably consume, and it can make a direct offer economically better than the intermediated one.

European parity reform improves the third condition. AI improves the first-stage discovery problem. Property-management systems, channel managers and structured web standards could improve the second.

The likely winners would therefore be branded hotel groups, recognized independent resorts, professionally managed vacation-rental portfolios and operators with substantial repeat business. Their problem is principally discoverability and customer acquisition, not the absence of an accountable transaction counterparty.

The one-property vacation-rental host faces a different hurdle. A direct website that an AI can read does not itself create consumer confidence equivalent to Airbnb's marketplace protections.

The direct scenario becomes materially more likely if a neutral infrastructure layer emerges that lets independent suppliers publish live availability and direct rates while a third party provides identity, payment protection and traveler guarantees at a cost materially below OTA commissions. That company might be a payment processor, insurer, PMS network or new travel protocol rather than an AI-model developer.

Within twelve months the evidence should be visible in conversion data, not citation anecdotes. Hotels and professional STR managers should be able to show that AI-originated demand produces material incremental direct bookings rather than tiny referral counts. Reputable channel panels should show direct share rising beyond ordinary year-to-year noise.

A new AI-native intermediary remains possible

Layer three: forecast, confidence low. The fourth outcome is frequently omitted because discussion is framed as “direct versus OTA.” An assistant might instead become the next OTA.

Technically, the building blocks are becoming visible. OpenAI's agent can operate websites, ChatGPT supports embedded partner apps, Google's agents can act through Chrome, and Perplexity already supports integrated checkout in retail.

Economically, however, becoming an accommodation intermediary means volunteering for precisely the obligations general-purpose AI companies can currently avoid by routing transactions elsewhere.

The scenario becomes credible when one of them does four things publicly: signs lodging suppliers or wholesalers at meaningful scale, owns or controls checkout, promises defined post-booking remedies, and builds a support and risk operation capable of funding those remedies.

Until then, “agentic booking” should not be confused with “new travel intermediary.”

Indicators worth watching

The most useful 2026-2027 indicators are not AI model benchmark scores.

First, watch transaction architecture. Does a traveler presented with a lodging option in ChatGPT, Gemini or another assistant ultimately accept the supplier's terms, Booking's terms, Expedia's terms or the AI company's own terms?

Second, watch payment ownership and servicing. The entity authorized to issue refunds, re-accommodate guests and absorb unrecoverable losses is likely to possess more bargaining power than the entity that generated the prose recommendation.

Third, watch B2B intermediary growth. Expedia's B2B gross bookings already reached $35.7 billion in 2025, versus $83.9 billion in B2C gross bookings, making invisible distribution a substantial business before AI agents become a mature acquisition channel. If AI partnerships accelerate the B2B side disproportionately, that would support the infrastructure scenario.

Fourth, watch structured supplier-direct connectivity. A widely adopted protocol through which hotels and STR managers can expose authoritative price, inventory, cancellation and identity data directly to agents would materially strengthen the direct thesis. Mere schema markup for descriptive webpages would not be enough.

Fifth, watch independent channel-share data. The research problem will remain underdetermined until STR and hotel analytics providers publish comparable transaction-level data separating OTA, supplier-direct and AI-assisted origin, with consistent definitions and panel methodology.

Finally, watch regulation. The European Commission's July 2026 finding that Google continued to self-preference hotel and other vertical results shows that the search distribution rules are still moving rather than settled. Booking's DMA obligations likewise make Europe a continuing test of whether contractual freedom translates into supplier channel gains.

Limitations and open questions

The most important limitation is attribution. A traveler may discover a property through an AI assistant, inspect it on Booking.com, visit the hotel's direct website, ask another AI question and then book using a loyalty app. Assigning that transaction to a single acquisition channel can misrepresent the causal path. The older billboard-effect literature already demonstrated this problem before generative AI arrived.

A second limitation is that most zero-click research measures informational web publishing rather than travel commerce. Pew's study is excellent behavioral evidence but does not isolate hotel queries. Ahrefs' original sample was heavily informational. Semrush shows commercial-query exposure increasing, but commercial exposure is not equivalent to travel-booking conversion.

Third, AI referral measurement is proprietary. Similarweb, Ahrefs and Semrush make their money partly from digital-traffic intelligence and search optimization. Their datasets are valuable, but their global traffic estimates cannot be independently reproduced from public information.

Fourth, the STR direct-booking market is particularly opaque. Hostfully, Lodgify, Hospitable, Hostaway and Key Data observe different slices of operators. Their products can themselves select for more professional or more direct-oriented hosts. Figures from these datasets should therefore not be averaged into a synthetic “true” direct share.

Fifth, public companies do not expose enough granular economics. Booking and Expedia do not report a standalone vacation-rental P&L comparable with Airbnb's company-level disclosures. Nor do they provide a standardized measure of paid versus unpaid customer acquisition. Booking's room-night direct percentage is useful, but it measures direct acquisition to Booking's platform rather than the accommodation's direct channel.

Sixth, the financial data cannot identify an AI effect causally. The period from 2021 to 2025 includes pandemic recovery, inflation, international travel normalization, product changes and shifting advertising conditions. The series is a baseline, not an experiment.

Seventh, the commercial agreements that may decide the issue are mostly private. Public announcements establish that OpenAI, Expedia, Booking and others integrate. They rarely disclose who pays whom, commission percentages, ranking commitments, refund obligations or ownership of the resulting customer.

Finally, technology forecasting in travel has a poor record when it mistakes a change in interface for a change in industry structure. Search engines, OTAs, metasearch, mobile apps and supplier-direct campaigns repeatedly altered where consumers looked without cleanly eliminating the functions that existing intermediaries performed. The current forecast should therefore be revised aggressively when transaction evidence arrives rather than defended because the technology narrative remains attractive.

Key source notes

Source	Date / period	What it contributes	Commercial or institutional position
Booking Holdings SEC filings	FY2021-FY2025	Gross bookings, revenue, marketing, direct mix, AI/search risk language, accommodation-marketplace operations.	Primary issuer disclosure; management has an interest in the company but filings carry securities-law obligations.
Airbnb SEC filings	FY2021-FY2025	GBV, revenue, sales and marketing, AI/search risk, trust and AirCover structure.	Primary issuer disclosure.
Expedia Group SEC filings	FY2021-FY2025	Gross bookings, B2C/B2B split, selling and marketing, lodging inventory, AI-search risk.	Primary issuer disclosure.
Pew Research Center	March 2025 telemetry; published May and July 2025	Observed Google search click behavior with and without AI summaries; chatbot-site visit rates in the same 900-adult panel.	Nonprofit research organization; no travel-distribution commercial position.
Ahrefs	2025 and 2026 studies	Keyword-level estimates of AI Overview impact on organic CTR.	Commercial SEO software vendor; lower search clicks are relevant to its customer base.
Semrush	2025-2026	AI Overview prevalence, intent mix and AI referral-channel estimates.	Commercial digital-marketing and SEO intelligence vendor.
Similarweb	June 2025	Estimated AI referral traffic compared with Google referral traffic.	Commercial web-analytics company; methodology relies on proprietary estimation.
European Commission	2024-2026	DMA designation, Booking parity obligations and enforcement against Google's search self-preferencing.	Regulator and enforcement authority.
Ennis, Ivaldi & Lagos, Journal of Law and Economics	2023	Natural experiments on broad and narrow hotel price-parity restrictions.	Peer-reviewed academic research.
International Journal of Industrial Organization study	2026	Synthetic-control estimates of parity reform effects on transacted prices and occupancy.	Peer-reviewed academic research; authors note institutional views are their own.
Mirai	2024	Hotel-level measurement of Google DMA search changes and direct-booking attribution.	Hotel technology company whose products are designed to increase direct sales; useful proprietary data but commercially interested.
OpenAI, Google, Booking.com and Expedia product announcements	2023-2026	Evidence of products that actually shipped, assistant capabilities, partnerships and agent architecture.	Primary product sources describing their own systems; useful for existence and architecture, not neutral evidence of economic impact.

Bottom line. The strongest evidence available today does not show AI transferring lodging-distribution power to small suppliers. Search visibility is becoming less dependent on the old ranked-link model, but website traffic generated by AI remains small, AI answers often suppress outbound clicks, and the accommodation implementations that have reached the market frequently rely on existing aggregators' structured supply and transaction systems.

The strongest evidence also does not yet justify declaring that OTAs will capture a dramatically larger share. Regulatory experiments show substantial adaptation and surprisingly limited aggregate channel movement, while direct pricing freedom is real and AI browser agents can technically reach supplier sites.

What the evidence does justify is a hierarchy of claims. Discovery is becoming easier to reintermediate. Transactions are harder. Liability is harder still. Until AI systems or an adjacent infrastructure layer take responsibility for the latter two, the economic center of gravity in short-term rentals is likely to remain with parties that aggregate not merely information, but inventory, money, trust and recourse. For branded hotels and professional operators, AI creates a more credible direct opportunity because those suppliers already possess much of that infrastructure themselves.

The next decisive dataset will not measure which domain an AI cites. It will measure who takes the booking, who keeps the customer, and who pays when the stay goes wrong.

How we know

The baseline is fiscal-year 2021–2025 gross bookings or GBV, revenue, and marketing or sales-and-marketing expense from Booking Holdings, Airbnb, and Expedia Group Form 10-K filings, read as of August 24, 2026. Implied revenue ratios are accounting quotients, not contractual commissions. Risk-factor language was compared across successive annual filings to identify when AI moved from a technology risk into a distribution risk.

Search behavior is taken first from Pew's March 2025 opt-in browsing telemetry (900 U.S. adults; 68,879 Google searches), published in May and July 2025. Ahrefs, Semrush, and Similarweb keyword and referral estimates are included because they cover larger samples; they are proprietary vendor measurements, not audited internet-wide counts.

Natural experiments were preferred to commentary: European DMA changes to Google travel search (Mirai hotel-level attribution, 2024) and European hotel price-parity restrictions (Ennis, Ivaldi and Lagos, 2023; Klopack and Pierri, 2026). Product architecture is taken from first-party announcements by OpenAI, Google, Booking.com, Expedia Group, and Perplexity for systems that actually shipped.

Excluded because they could not be measured reliably: a representative time series of vacation-rental direct-booking share; an audited split of AI-assisted lodging transactions by final channel; paid-versus-unpaid traffic for the large platforms; and the commercial terms of AI–OTA distribution contracts. Scenario probabilities are judgments, not measurements.

Sources

• Booking Holdings Inc. Form 10-K for the year ended December 31, 2025
• Airbnb, Inc. Form 10-K for the year ended December 31, 2025
• Expedia Group, Inc. Form 10-K for the year ended December 31, 2025
• Pew Research Center — Google users are less likely to click on links when an AI summary appears (July 22, 2025)
• Pew Research Center — What web browsing data tells us about how AI appears online (May 23, 2025)
• Ahrefs — AI Overviews reduce clicks by 34.5% (April 17, 2025)
• Ahrefs — Update: AI Overviews reduce clicks by 58%
• Semrush — AI Overviews’ impact on search in 2025
• Semrush — AI Overviews are expanding across commercial intent search
• Semrush — How AI is reshaping traffic channels
• Similarweb — AI referral traffic winners by industry (June 2025 estimates)
• Booking.com — AI Trip Planner launch (June 27, 2023)
• OpenAI — Introducing apps in ChatGPT (October 6, 2025)
• Expedia Group — Expedia in ChatGPT
• Expedia Group — Acquisition of Layla (July 31, 2026)
• OpenAI — Introducing ChatGPT agent (July 17, 2025)
• Google — Gemini Spark Chrome browsing integration (July 30, 2026)
• Google — New connected apps coming to Gemini (August 12, 2026)
• Perplexity — Shopping that puts you first
• PayPal — Instant Buy with Perplexity (November 25, 2025)
• Airbnb Help Center — Getting protected through AirCover for Hosts
• European Commission — Booking must now comply with the Digital Markets Act (November 14, 2024)
• European Commission — Google fined €890 million for DMA breaches (July 23, 2026)
• Court of Justice of the European Union — Press release, Case C-264/23 (September 19, 2024)
• Ennis, Ivaldi, and Lagos — Price-Parity Clauses for Hotel Room Booking (Journal of Law and Economics, 2023)
• Klopack and Pierri — Broad and narrow price parity agreements (International Journal of Industrial Organization, 2026)
• Mirai — DMA implementation sinks 30% of clicks and bookings on Google Hotel Ads
• Mirai — DMA impact on hotels: 0.8% loss of direct reservations
• Skift — Hostfully 2022 survey on direct bookings (December 11, 2022)
• PhocusWire — Hospitable and Hostaway 2026 host surveys
• Skift — Airbnb, Booking, and Expedia/Vrbo 71% global STR share in 2024
• Anderson — Search, OTAs, and Online Booking: An Expanded Analysis of the Billboard Effect (Cornell, 2011)
• Anderson and Han — The Billboard Effect: Still Alive and Well (Cornell, 2017)