The OTA model is facing the most serious structural threat in its history — and the threat is not coming from a competitor, it is coming from the architectural layer beneath the entire industry. AI agents are collapsing the distinction between discovery, research, and booking into a single interface. When that transition completes, the 15–20% commission margin that OTAs extract by owning that interface is directly at risk. Sarah Kopit (Skift) drew the comparison explicitly: ChatGPT's current share of Expedia traffic looks exactly like mobile's share of Expedia traffic in 2009. Everyone in the room in 2009 knew what happened next.
The disintermediation threat, however, is not inevitable — and this brief argues both sides with evidence. OTAs that control payment infrastructure, possess rich behavioral data, and move aggressively into AI-native distribution channels are structurally positioned to entrench their position, not lose it. Booking Holdings' $130 billion in transaction volume growing at 63% year-over-year is not a company being disintermediated; it is a company becoming the payment and data rails of the industry. The risk is concentrated in OTAs that sit in the middle of the stack without controlling either end — those that own neither the consumer relationship deeply enough to survive AI shortlisting nor the payment and data infrastructure required to be indispensable to suppliers.
Three protocol wars are now underway simultaneously — Google's Universal Commerce Protocol, Stripe's Agentic Commerce Protocol, and Alibaba's Open Claw — and the winner of the protocol layer will reshape distribution more fundamentally than the OTA revolution of the 2000s. OTAs are not bystanders in this race. They are both the most threatened incumbents and the most capable participants. Their data assets, supplier relationships, and transaction volume give them more leverage than any hotel chain or airline to shape protocol outcomes. The question is whether leadership teams understand this as a distribution infrastructure problem rather than a marketing technology problem.
The 90-day horizon is not one for watching. Fliggy processed 12 million AI-driven orders in six days. LLM booking conversion rates at forward-leaning companies moved from 1% to 12% in twelve months. Google is testing in-search hotel bookings inside AI Mode — a feature that, if broadly rolled out, inserts Google as the transaction layer between consumers and every OTA currently dependent on Google for traffic. The window for OTAs to establish AI-native distribution patterns before LLMs calcify their defaults is open now. Andrew Boch's warning from the eTravel stage is the brief's organizing principle: "LLMs learn from observed behavior — if OTAs continue to be where bookings happen, models will recommend OTAs. This pattern is being set now."
1. The AI traffic ratio gap is the most important number in travel distribution.
Google scrapes a page and sends back two visitors (1:2 ratio). OpenAI sends back one visitor for every 250 pages consumed (1:250). Claude's ratio is 1:6,000. US travel site visits from AI sources jumped 3,500% year-over-year in July 2025. The volume is already here. The revenue conversion from that traffic is not. OTAs have a narrow window to become the default booking destination that LLMs route to — or to be routed around entirely.
2. Booking.com is not being disintermediated. It is becoming the infrastructure.
$130 billion in transaction volume, 63% YoY growth, 109 payment methods, 130+ provider relationships. Daniel Marovitz's framing at ITB was explicit: "Our job is to inject financial products into the experience of travelers." The agency model is structurally exposed to AI disintermediation. The merchant model — owning the transaction, owning the payment, owning the data — is not. Every OTA leadership team should be asking which model they are actually operating.
3. The protocol layer is being built without you if you're not at the table.
Google's UCP launched in the US in January 2026. Stripe's ACP was open-sourced after six months of co-development with OpenAI. Alibaba's Open Claw was framed as "Linux but with a brain." All three are structuring how AI agents discover, evaluate, and transact travel inventory. Amadeus has invested in the MCP layer. Skyscanner has opened its REST+MCP API to third-party AI agents. The companies not doing this equivalent work are building on sand.
4. Your GDS-routed data is being truncated to 20 characters.
GDS note fields cap at 20 characters. Any contextual data that AI conversations elicit — hypoallergenic soap preferences, interconnecting room requirements, pet policies, accessibility needs — is architecturally truncated before it reaches the supplier. AI personalization that routes through legacy GDS infrastructure is not personalization; it is the illusion of personalization on top of structurally impoverished data. This is not a configuration problem. It is an architectural one with a multi-year remediation timeline.
5. The loyalty model faces an existential test it was not designed to pass.
Accor's ALL program has 100 million members who spend 2x versus non-members, with 1 in 3 bookings coming from a loyalty member. If an AI agent selects a hotel based on its interpretation of a traveler's preferences, does the loyalty signal survive the agent layer? Dan Ogen's warning from the conference was precise: "AI will not give you 150 options. It will give you three or four." If your loyalty program does not transmit through an AI decision layer, it is not a moat — it is a sunk cost.
Strategic Threats
Agentic Disintermediation
The mechanism is straightforward: AI agents do what OTAs do, but without charging 15–20% commission. They aggregate inventory, compare options, surface relevant alternatives, and — increasingly — complete the transaction. Bernstein's 60%-probability scenario from ITB: AI platforms access OTA and supplier inventory directly, hotels using Model Context Protocol bypass OTAs entirely, feeding rates straight to LLMs. Stuart Barwood projects 25% of bookings will be fully agentic direct-to-supplier within 3–5 years.
The compounding threat is the shortlist problem. In keyword search, OTAs win by surfacing the most results. In AI-mediated discovery, the model surfaces three or four options. Getting onto that shortlist requires structured, accurate, machine-readable data and established behavioral patterns in the training corpus. OTAs that have historically competed on breadth are suddenly competing on a different dimension they were not built for.
The urgency is real. Three speakers independently warned at the AI track that the industry has "less than one year" to establish positioning before agentic defaults calcify. The training feedback loop is running now: LLMs learn from where bookings happen. If today's booking patterns route through OTAs, tomorrow's models recommend OTAs. If they don't, they won't.
Protocol Fragmentation Risk
OTAs face a classic early-mover dilemma in a protocol race where the eventual winner is unclear. Committing to Google's UCP locks in a Google-adjacent distribution architecture. Stripe's ACP is the OpenAI partner stack. Alibaba's Open Claw has unmatched transactional scale data but is built for a Chinese ecosystem. Early protocol choices will create integration debt and potential lock-in effects. Companies that wait for a winner risk being locked out of the foundational infrastructure layer while it is still being built. Companies that back the wrong protocol face costly re-architecture.
The deeper structural risk: if generic retail protocols (built for fixed SKU reference numbers) become the travel industry default before travel-specific servicing logic — cancellations, schedule changes, dynamic reference number generation — is baked in, OTAs will spend years adapting to an architecture that was not built for their operational reality. Amadeus is the only company at ITB that clearly understood and articulated this risk.
Social Commerce and Discovery Fragmentation
Travel discovery now happens across five distinct modalities simultaneously: Google keyword search, Google AI Mode, TikTok visual discovery, YouTube long-form inspiration, and direct AI agent planning. No single channel dominates, and the emerging #1 inspiration source for travel is TikTok — with 81% of users who discover travel content booking within one month. OTAs built their customer acquisition model around Google paid search and SEO. That model faces structural disruption from both ends simultaneously: AI is eroding Google's query-to-click economics while TikTok is pre-capturing intent before it reaches search at all.
HolidayPirates is the proof-of-concept: 35 million followers, 10 markets, 90% organic traffic, zero paid acquisition. A single reel generating €30,000 in commission. The social-native OTA model operates at the discovery layer that traditional OTAs do not control. The full funnel now runs: social discovery → Google evaluation → OTA or direct booking. OTAs are capturing the end of a funnel they don't own the top of.
Direct Channel Resurgence via MCP
Every hotel with a website can now, in theory, have a Model Context Protocol server live within three months, enabling direct AI-native booking without OTA intermediation. The MCP deployment window is open. Andrew Boch's directive at ITB was explicit: any hotel can do this now. The strategic implication is the same structural threat OTAs faced when hotels built direct booking websites in the early 2000s — except the timeline is compressed and the interface is AI-native rather than web-native. The OTAs that solved the first round by building better search and lower friction need to solve this round by becoming AI platform infrastructure rather than just booking interfaces.
Strategic Opportunities
Data Moats That Actually Compound
The most important statement from the AI track: "Even Gemini 2.5 Flash achieves only a 23% satisfaction rate on specific travel tasks because it lacks local data. Private data and local knowledge are the true moats." Generic frontier models cannot serve travel adequately without proprietary behavioral data. OTAs with years of search, comparison, and booking data are sitting on training assets that no hotel chain, airline, or new entrant can replicate.
The opportunity is to make that data work. Context engineering — structuring product data for LLM discoverability — is the successor to SEO and SEM. The OTA that best structures its inventory for AI reasoning, and that contributes the richest behavioral training signals to AI systems, compounds its position. Sabre holds 50 petabytes of structured travel data. The question for OTAs is not whether they have data — they do — but whether it is structured for the AI era or locked in legacy formats that LLMs cannot reason about.
Payment Infrastructure as Strategic Moat
Booking.com's merchant model transformation is the clearest strategic playbook at ITB. $130 billion in transaction volume is a payment infrastructure play disguised as an OTA. The Booking.com/Adyen partnership produced 70% reduction in go-to-market time, 60% total cost reduction, and 2% card acceptance rate improvement. For OTAs, the payment layer is both the least glamorous and most defensible part of the stack: it handles FX complexity, dynamic package settlement timing, B2B supplier payments, and increasingly BNPL and embedded finance products. Travel payments are structurally complex because, as Mary Cotugno observed, "banking rails were conceived for linear commerce and travel isn't linear." OTAs that own the payment rails own something that AI agents cannot bypass.
The opportunity for OTAs willing to invest: build the "Adyen for travel" that the hypothesis tests identified as a $50B+ opportunity. The $130B Booking Holdings number shows it is not a $50B opportunity — it is potentially a $500B opportunity if the merchant model fully matures across the industry.
AI-Native Experience Products
The tours and activities market is a $271 billion market growing to $340 billion by 2029, with 72% of experiences still offline. LLM-native discovery generates significant new demand in previously invisible long-tail product categories. Fliggy's demonstration showed that natural language queries surface products that keyword search cannot find — "ice hotel plus animal kingdom plus breakfast" as a coherent request. This is a structural opportunity for OTAs to expand their addressable market by making the long tail discoverable.
The MICE sector is the clearest early beachhead: Flemings Hotels went from 9% to 18% lead-to-sale conversion after AI automation, 70% of hotel RFP requests currently go unanswered, and day-one proposals convert at double the rate of day-two proposals. An OTA that captures the MICE automation opportunity — structured lead-to-proposal-to-booking — is building in a sector where AI ROI is already demonstrated and the competitive moat is high.
Emerging Market Demand Pipeline
India, China, and the USA are identified as the three demand superpowers for the next decade. The middle-class demand pipeline is predictable: 2–3 years to domestic travel, 2–3 more for regional, up to 9 years for long-haul. Emerging markets contributed 47% of 2019 tourism arrivals. Google's Outlook projects top-5 destinations' share dropping from 37% (2000s) to 16% by 2050. OTAs with the payment infrastructure, language support, and local inventory relationships to serve emerging market travelers will capture the majority of global travel volume growth. OTAs optimized exclusively for European and American consumers will not.
The payment dimension of this opportunity is specific: APAC, LATAM, and MENA markets have prohibitive card acceptance costs. Virtual cards, which have dominated B2B travel for 20+ years, cannot serve these markets. Real-time payment rail adoption is accelerating in precisely the markets where OTA penetration is lowest. The OTA that solves emerging market payments first builds a distribution position that is extremely difficult to replicate.
The Agentic Commerce Question
The Case That AI Agents Disintermediate OTAs
The structural argument is clean. OTAs exist because they solve a search and comparison problem that is expensive for consumers to solve themselves. AI agents solve the same problem — search inventory, compare options, surface the best alternatives, complete the booking — without requiring a 15–20% commission margin to do it. The value proposition of the traditional OTA is directly substitutable by a capable AI agent with inventory access.
The evidence is accumulating. Bernstein's 60%-probability scenario explicitly models hotels feeding rates to LLMs via MCP and bypassing OTAs entirely. Bhanu Chopra at ITB predicted AI agents will condense the distribution chain, reducing intermediary "hops" between consumer and supplier. Google is testing hotel bookings inside AI Mode — which, if rolled out, makes Google itself the transaction intermediary rather than an OTA. The 1:6,000 Claude traffic ratio means LLMs are consuming OTA content at massive scale while returning almost no referral traffic. The economics of the current arrangement are already breaking down.
The timeline is also compressing faster than most OTA leadership teams appear to believe. LLM-to-booking conversion moved from 1% to 12% in twelve months at forward-leaning companies. Dr. Alex Chen's warning — "the industry has less than one year to prepare" — is not from an academic; it is from the CTO of Fliggy, a platform that already processed 12 million AI-driven orders in six days during Chinese New Year.
The Case That AI Agents Entrench OTAs
OTAs are not passive recipients of distribution change — they are the most data-rich, capital-intensive, and technically capable participants in the travel ecosystem. The disintermediation thesis assumes that AI agents can access supplier inventory directly as easily as through OTAs. This ignores the structural reality: OTAs have spent decades building supplier integrations, content normalization, pricing APIs, payment infrastructure, and review aggregation that no AI agent can replicate without those same integrations. The path of least resistance for an AI agent seeking reliable, structured, normalized inventory is an OTA's API — not a thousand fragmented hotel MCP servers, many of which will have impoverished data architectures.
Booking.com's data is the most compelling counterpoint to the disintermediation narrative. 30%+ of Booking.com orders are already multi-category; over 10% of GMV is multi-category. The OTA that best enables multi-category AI-native trip assembly — flight + hotel + experience in a single conversational interface — retains the value proposition that individual suppliers cannot replicate. An AI agent building a trip from scratch via direct supplier APIs faces a coordination problem across hundreds of disconnected integrations; an OTA with a unified API and transaction model solves that problem with a single connection.
The training feedback loop is also an OTA advantage. LLMs learn from observed behavior. If bookings happen through OTAs today — and they do, accounting for approximately 25% of hotel bookings — the training data reinforces OTA recommendation. The OTA that acts now to become the default booking destination in AI interfaces compounds this advantage daily. The ones that wait, lose it.
The verdict: Both scenarios are live simultaneously. OTAs with strong payment infrastructure, rich behavioral data, multi-category inventory, and aggressive AI platform positioning will entrench. Pure-play OTAs that compete only on search aggregation and have not built merchant model depth will face structural margin compression. The outcome is determined by decisions being made in the next 90 days, not the next 3 years.
The Data Moat vs Open Protocol Tension
This is the defining strategic dilemma for OTAs in 2026, and it has no clean resolution.
The data moat argument: Generic frontier models achieve a 23% satisfaction rate on specific travel tasks (Fliggy/Dr. Alex Chen). The only way to improve that is with proprietary data — behavioral patterns, inventory data, pricing histories, user preferences — that OTAs possess and that no open standard can transfer. The OTA that trains AI systems on its own data produces better recommendations, which drives more bookings, which generates more training data. This is a genuine compounding flywheel. Surrendering that data to open protocols is surrendering the flywheel.
The open protocol argument: Piero Sierra (Skyscanner) made the case that "companies that build walled gardens will be bypassed; those that offer open APIs with ML-derived intelligence will be indispensable." If every major AI agent platform routes to the most open and reliable inventory sources first, OTAs that refuse to expose structured data via open protocols will be invisible to the agents booking on behalf of consumers. The Skyscanner thesis: become the data and intelligence backbone for third-party AI agents, not the competitor to them. Rome2Rio framed the same position: "We look at LLMs as partners, not threats."
The cross-track synthesis identified this as the industry's central collective action problem: everyone benefits from open data for market-wide AI quality, but individual firms benefit from hoarding it. There is no natural resolution. The practical answer emerging from ITB: invest in structured, high-quality data that can be surfaced via open protocols while retaining behavioral and preference data that trains proprietary AI systems. The distinction is between inventory data (open it, because AI agents need it and locked inventory becomes invisible inventory) and behavioral data (protect it, because it is the training asset that generic models cannot replicate).
For OTAs specifically, the strategic sequence is: (1) ensure inventory data is structured for LLM consumption via open protocols before the defaults calcify, (2) build proprietary AI systems trained on behavioral data as the differentiated layer on top, and (3) own the transaction via merchant model infrastructure that AI agents must route through regardless of which discovery interface they use. The OTA that owns the transaction owns the data generated by the transaction. Discoverability is becoming commoditized; the payment and post-booking layer is where durable advantage accumulates.
The uncomfortable implication: OTAs that have historically competed by restricting supplier access to their customer data — refusing to share booking source data with hotels, limiting API access — are training for the wrong sport. The AI era rewards data openness in the inventory layer and data richness in the behavioral layer. The firms that understand this distinction will shape the next decade of travel distribution.
The Uncomfortable Truths
1. Your best AI customers are your competitors' best customers.
The LLM traffic ratios (Claude 1:6,000, OpenAI 1:250) mean that the same AI systems OTAs are racing to integrate are consuming OTA content at industrial scale and returning almost no traffic. OTAs are, right now, inadvertently training the systems most likely to displace them. This is not a hypothetical future risk — it is the current operating reality.
2. The K-shaped economy is your concentration risk, not your growth story.
The top 10% of US earners account for nearly half of all consumer spending at a record level, and the gap is "even more pronounced in travel" (Kopit). Hotels commanding $1,000+/night have tripled since 2019. OTAs celebrating premium RevPAR growth are celebrating a deepening dependency on a narrow customer segment that will still travel in a recession, "just not with the same gusto." The breadth of the OTA customer base was a strategic asset; OTAs that have implicitly abandoned it to chase premium yield have narrowed their moat, not widened it.
3. Your loyalty program probably doesn't survive the agent layer.
If an AI agent selects on a traveler's behalf, it selects based on data it can reason about — price, reviews, structured attributes, past behavioral patterns. Loyalty points are not a structured attribute that survives AI reasoning unless they are explicitly engineered into the data feed. The 100 million member loyalty programs built over the past decade may be the industry's most expensive irrelevant asset if they are not redesigned for AI-mediated decision-making.
4. You are running a 20th-century distribution architecture in a 21st-century discovery environment.
GDS 20-character field limits. Fragmented identity data. Non-machine-readable hotel content. "130+ room type combinations" that need machine-readable standardization. These are not technical footnotes; they are the binding constraints that will determine whether OTA data feeds are useful to AI agents or are ignored by them. The infrastructure gap between the OTA stack and what AI-native distribution requires is larger and more expensive to close than most engineering roadmaps currently acknowledge.
5. "Start now imperfectly" is good advice that kills companies if executed without discipline.
Andrew Boch's directive — every hotel can have MCP live in three months — is correct and dangerous simultaneously. The MCP server that trains LLMs on impoverished data does not improve AI recommendations; it poisoned them. The OTA that moves fast on AI integration without first investing in data quality trains AI systems to make bad recommendations about its inventory. Quality of data going into AI systems is not a launch detail — it is the entire competitive moat. The companies that get this wrong in the next 12 months will spend 3 years correcting the models they trained.
6. The US inbound collapse is not an anomaly to wait out — it is a demand rebalancing to act on.
Nearly half of travelers surveyed in 5 countries are less likely to visit the US due to Trump policies. US is the only major tourism economy to contract in 2025. Full US recovery to 2019 levels is not projected until 2029. European and Asian OTAs have a multi-year window to capture displaced demand that is looking for destinations that actively want it. The destinations rolling out the red carpet are capturing travelers the US is repelling. OTAs that respond to this shift with product and marketing agility will take disproportionate market share.
90-Day Action Plan
Week 1–2: Intelligence and Triage
Map your AI traffic reality. Pull all AI-source referral traffic (ChatGPT, Gemini, Claude, Perplexity, AI Overviews) for the last 12 months. Measure conversion rates from each source. Benchmark against the known ratios (Google 1:2, OpenAI 1:250, Claude 1:6,000). You need to know where you currently sit in the AI distribution ecosystem before making any investment decisions.
Audit your data architecture for LLM readiness. Commission a technical assessment of how much of your inventory data is structured, machine-readable, and accurate at the attribute level required for AI reasoning. The question is not whether you have data — you have enormous amounts of data — the question is whether AI agents can reason about it. Specific items to assess: room type standardization, property attribute completeness, dynamic pricing API coverage, review data recency, and multi-category inventory linkage.
Identify your merchant model depth. Map every transaction touchpoint where you do and do not own the payment. The agency model transactions — where you are a referral layer and the supplier owns the transaction — are the ones most exposed to AI disintermediation. The merchant model transactions — where you hold the payment, bear the settlement risk, and own the transaction data — are the defensible ones.
Weeks 3–6: Protocol Positioning
Establish MCP/protocol presence. You cannot afford to wait for a protocol winner. Deploy a Model Context Protocol endpoint for your inventory that exposes your highest-confidence, best-structured data to AI agents. Do not deploy this for all inventory simultaneously; deploy it for the inventory where your data quality is highest. A small, high-quality AI-native inventory surface is better than a large impoverished one.
Engage directly with Google, Stripe, and Amadeus on protocol development. The protocol layer is being built now. Your leverage to influence the servicing logic — the cancellations, changes, and rebooking functionality that retail protocols currently lack — is highest while the protocols are still being designed. After the standards calcify, your negotiating position is as a compliance cost, not a shaping partner.
Define your AI agent governance framework. Dr. Patrick Andrae's question from the AI track — "How do you govern what AI agents can do? It's like HR policies for humans" — is not rhetorical. Before deploying AI agents at scale, establish: what decision categories agents can make autonomously, what requires human escalation, what the liability framework is for agent booking errors, and how consumer data is handled in agent interactions. Gartner projects 40% of agentic AI projects cancelled by late 2027 due to poor governance. Do not be in that 40%.
Weeks 7–10: Channel and Product
Build GEO capability as a parallel workstream to SEO. This is not a replacement for your existing search spend — 94% of frequent LLM users are also frequent Google users. This is an additive investment in a new discoverability surface. Prioritize: structured data markup across all inventory pages, third-party review presence on platforms that AI systems cite (not just Google), factual accuracy audits across all property descriptions, and schema standardization for event, activity, and experience categories.
Develop a social commerce strategy with transaction completion. TikTok's booking acceleration data (81% book within one month of discovery) is a channel signal OTAs cannot ignore. The strategic challenge is that social platforms are discovery engines, not transaction engines — the traveler who discovers a destination on TikTok still has to go somewhere to book. The OTA with embedded booking capability, deep-linked from social content, captures the transaction. The one waiting for the traveler to navigate back to Google does not.
Prioritize MICE and B2B automation as near-term AI beachhead. Flemings Hotels' 9% to 18% conversion improvement from MICE AI automation is the clearest documented ROI case for agentic AI in travel. For OTAs with B2B exposure, the MICE sector has the highest near-term AI return: manual RFP processes, measurable conversion outcomes, relationship-driven sales culture that AI can augment rather than replace. A 12-month focused investment in MICE agentic automation produces ROI that funds the broader AI buildout.
Weeks 11–13: Payments and Emerging Markets
Accelerate merchant model conversion. Identify the top 5 markets or inventory categories where you are currently operating as an agency model and quantify the data and transaction value being lost to suppliers. Build a merchant model business case for each, including the payment infrastructure investment required. The Booking.com/Adyen case study is your template: 70% reduction in go-to-market time, 60% total cost reduction, 2% card acceptance lift.
Initiate emerging market payment capability assessment. Map the gap between your current payment method coverage and the payment rails dominant in your three highest-growth emerging market opportunities. Identify which account-to-account and local real-time payment rails you need to support. The virtual card model that has dominated B2B travel for 20 years cannot serve APAC, LATAM, and MENA at competitive cost. The OTAs that solve this first in each market build a distribution position that compounds.
Key Quotes
> "The online travel agency model faces the most serious structural threat in its history due to generative and agentic AI." — Sarah Kopit, Skift Editor-in-Chief
> "Less than 1% is roughly what mobile traffic to travel sites looked like in 2009. We all know what happened next." — Sarah Kopit, Skift (on ChatGPT's current share of Expedia traffic)
> "If we seed that territory to the OTAs, we will have another 20 years of dominance by those players." — Sanjay Vakil, Direct Booker (warning that LLM booking patterns are being established now)
> "In an A-to-A world, your beautifully designed website or app won't matter as much. AI doesn't respond to visual hierarchy, immersive imagery, or clever copy." — Mitra Sorrells, Future Track
> "Our job is to inject financial products into the experience of travelers and into the experience of our supply side partners that helps them reduce risk, reduce anxiety, and increase flexibility." — Daniel Marovitz, Booking.com SVP FinTech
> "Private data and local knowledge are the true moats. Even Gemini 2.5 Flash achieves only a 23% satisfaction rate on specific travel tasks because it lacks local data." — Dr. Alex Chen, CTO Fliggy/Alibaba
> "AI will not give you 150 options. It will give you three or four." — Dan Ogen (on the shortlist compression problem)
> "Banking rails were conceived for linear commerce and travel isn't linear." — Mary Cotugno, Travelsoft Pay
> "Priority and commitment are not the same thing." — Mitra Sorrells (on the gap between declaring AI a top priority and allocating budget to it)
> "If we spend years as an industry talking about who's the merchant, we're going to miss the boat." — James Lemon, Stripe
Data Points That Matter
| # | Data Point | Source | Strategic Implication |
|---|---|---|---|
| 1 | LLM-to-booking conversion: 1% → 12% in 12 months | Uta von Dietze, Wyndham (eTravel Track) | The trajectory is steeper than most OTA roadmaps assume; the 18-month window is closing |
| 2 | Booking Holdings: $130B transaction volume, 63% YoY growth | Daniel Marovitz, Booking.com | Merchant model depth is the viable response to disintermediation; this is what it looks like at scale |
| 3 | AI traffic ratio — Claude 1:6,000 / OpenAI 1:250 / Google 1:2 | Sarah Kopit, Skift (Adobe data) | OTAs are contributing content to AI systems that generate almost no direct traffic return; the referral economics have broken |
| 4 | US travel site AI-source visits: +3,500% YoY in July 2025 | Adobe data cited by Kopit | Volume is already here; monetization of that AI traffic is the urgent problem |
| 5 | Fliggy: 12 million AI-driven orders in 6 days, Chinese New Year 2025 | Dr. Alex Chen, Fliggy (AI Track) | At-scale agentic commerce is not a future scenario; it is live in the world's largest travel market |
| 6 | Gemini 2.5 Flash: 23% satisfaction rate on specific travel tasks without proprietary data | Dr. Alex Chen, Fliggy | Generic frontier models cannot serve travel; proprietary behavioral data is the defensible moat |
| 7 | TikTok: 81% of users who discover travel content book within one month | Adam Zarin, TikTok (Marketing Track) | OTAs without social commerce capability are losing top-of-funnel to a channel they don't control |
| 8 | Top 10% of US earners: nearly half of all consumer spending (record); gap "even more pronounced in travel" | Sarah Kopit, Skift (citing Moody's) | K-shaped demand is a concentration risk; OTAs optimized for premium are narrowing their moat |
| 9 | OTAs: ~25% of hotel bookings in 2024, equal to direct channels | Sarah Kopit, Skift | Parity is the starting position; the question is which direction it moves in the AI transition |
| 10 | Tours & Activities: $271B → $340B by 2029; 72% still offline | AI Track synthesis | Largest AI-unlockable TAM expansion opportunity in travel; OTA that cracks T&A distribution wins the next decade's growth |
| 11 | Booking.com/Adyen: 70% go-to-market time reduction, 60% total cost reduction, 2% card acceptance lift | lastminute.com/Adyen case study (eTravel Track) | The ROI on payment infrastructure modernization is documented and large; this is not a speculative investment |
| 12 | 61% of travel companies testing or scaling agentic AI; only 2% at widespread organizational use | Linda Fox session (AI Track) | The gap between pilot and production is enormous; the companies that close it first will shape defaults |
*Generated from ITB Berlin 2026 research corpus: 135 session transcripts, 17 track research memos, cross-track synthesis, and structured hypothesis testing across the full conference program.*