The question travel keeps asking has the threat wrong. Your moat will very probably survive disintermediation – which is the wrong reassurance. What it must survive is the new layer settling above you, and the customer who never comes back.
The question that ran through Phocuswright Europe this year, in almost every room I sat in, was blunt and personal: what is my moat, and will it protect me from disintermediation? It is the right instinct pointed at the wrong threat. Because the honest answer is that most moats worth the name probably will protect you from disintermediation, and that is not the reassurance the room imagines it to be.
I have made a version of this argument before. In Winning travel’s AI shift I wrote that the worry about being cut out was receding, that the big four would endure and the niche players would hold wherever they owned first-party data or an exclusivity advantage. What I would add this year is sharper. Asking whether your moat stops you being disintermediated is asking whether your defences hold against a siege that is not actually coming. The event to prepare for is not removal but reintermediation: a new layer settling on top of the existing stack and charging for access to the customer you thought was yours. A moat built to prevent removal does very little against a toll, and less still against the quieter loss beneath it, which is the customer who never comes back to you because the agent on their phone has become the thing they return to instead. Telling these apart is now the whole of the strategy.
The moat is being built against the wrong siege
Begin with why disintermediation is the wrong thing to fortify against, because the evidence here is unusually settled. A study that traced the web APIs of the online travel ecosystem across twenty-six years found the same move repeated in every cycle: an incumbent is challenged by a new entrant and then reconfigures rather than dies. The global distribution systems were partly displaced by the online agencies, then re-mediated themselves by opening their APIs to those same agencies. The metasearchers first bypassed the agencies by scraping them, then negotiated cooperative APIs, then replicated the agencies’ booking functions and began competing directly. The pattern is coopetition, not clean substitution, and nobody in it was ever simply erased.
The most instructive part is what the incumbents did next. Amadeus, Booking.com and Expedia began monetising their own APIs directly, turning the plumbing of intermediation into a revenue stream in its own right and, in doing so, lowering the barrier for new players to assemble marketplaces on top of them. Digital infrastructure does not remove the intermediary function; it diversifies who gets to perform it. The middle never disappears in travel. It multiplies, it relocates, and it finds new things to charge for.
The market reads the same way today. In July the investment bank BTIG concluded that the disintermediation risk to the online agencies had been exaggerated, and that pressure on margin was the likelier outcome, with AI acting as a customer-acquisition channel and the agencies positioned as partners of choice. That posture is expensive to hold. The combined sales and marketing spend of Airbnb, Booking, Expedia and Trip.com passed twenty billion dollars in 2025, up from under eighteen billion the year before. That is not the spending of businesses being cut out. It is the spending of businesses paying to stay in, which is a different thing entirely. So if your moat is designed to stop you being disintermediated, you are defending a gate no one is really storming. The attack is coming over another wall.
The real test of a moat is margin
Here is the wall that matters. Hotels already pay the online agencies transaction fees that the revenue-management literature puts at fifteen to thirty per cent, a range that alone should end any comfort drawn from mere survival. That same literature has been inventing metrics precisely because operators cannot see where their margin goes: NRevPAR, which nets out the acquisition cost of each channel, and RevPAC, which values a guest across every revenue stream rather than the room alone. The obstacle to adopting them is not arithmetic. It is the absence of industry benchmarks, weak sponsorship from the top, and a network-externality problem in which a new metric is only worth using once everyone else uses it too. The tools to see the leak exist; the discipline to look does not.
Now layer an agent’s cut on top of that existing fifteen to thirty per cent, and, absent any change, the margin compresses again. This is the exam most moats fail. A brand moat does not stop a take-rate levied above it. Breadth of undifferentiated inventory does not stop it either, since an aggregating layer commoditises that kind of supply first. Analysts modelling Booking Holdings arrive at the same place from the opposite direction, noting that operating margins can be squeezed even when gross booking value holds, and that Google’s AI Overview is already eroding the click-through that feeds paid listings. And the pressure is aimed. A study of Chinese hospitality firms found that AI adoption diversifies channel portfolios away from dominant partners, but that the diversifying force falls harder on the demand side, on the customer-facing channels where the agencies and the corporate intermediaries sit, than on the upstream suppliers. The squeeze has a target, and the target is the intermediary layer.
Which moats hold, and which are mirages
So the question the conference asked needs reframing. Not “will my moat stop me being removed?” but “does my moat give me leverage against the layer above me?” Those are different tests, and they sort the moats sharply.
The moats that hold are the ones the new layer cannot source, route around, or replicate for itself.
First-party data the layer cannot get elsewhere. A current, structured memory of the guest and of the product that an agent has to come to you for is leverage. Data an agent can reconstruct from public signals is not. This is the moat I argued for in Winning travel’s AI shift, and it is the same conclusion the early adopters of an AI spine keep reaching: that the thing which distinguishes them is not the size of their estate but what they actually control.
Unique or exclusive supply. If you are the only route to inventory a traveller wants, the agent has to deal you in on your terms, which is why the niche and the genuinely exclusive hold where the broad and the commodity do not. It is also the commercial logic behind selling access rather than assets, which I set out in To succeed in luxury travel, sell access not assets.
Legibility to the machine. This is the new moat, and it runs against instinct, because it is not about hiding from agents but about being readable by them. Booking-capable agents favour suppliers whose pricing and availability are machine-readable and verifiable, and the research is explicit that small and local providers risk exclusion from AI-mediated recommendation, and therefore from the transaction, unless their data is made accessible. If a machine cannot read you, you are not argued with. You are skipped, before a human ever sees your offer.
The direct relationship and the accountability that comes with it. The merchant-of-record standing that keeps you inside the contract, answerable to the guest and therefore paid, is a moat an intermediating layer cannot assume on your behalf without taking on the liability it is designed to avoid.
The moats that are mirages are the ones that only ever guarded against removal. Marketing scale is the clearest of them: the twenty billion dollars the big four spend is not a moat but rent, and rent rises. Sheer breadth of inventory without differentiation is another, since it is precisely what an aggregating layer absorbs and commoditises first. If your defence amounts to “we are too big to be cut out,” you have answered the disintermediation question and ignored the reintermediation one, which is the only question now being marked.
Being chosen once is not being remembered
There is a second question folded inside “what is my moat,” and it is the larger of the two. The first asks whether an agent picks you for a given trip, which is the transactional contest of being found, read and shortlisted. The second asks whether the customer comes back to you at all, or whether the thing they come back to is the agent itself. As general-purpose assistants take up residence on every phone and desktop, the default place a traveller now begins is not a brand’s app or website but whichever agent is already in their hand. That agent, used for everything and consulted by habit, accumulates the memory, the routine and the proactive nudges that were once the machinery of loyalty. Being chosen once is a transaction. Being remembered is a relationship, and the ever-present agent is quietly positioned to capture the relationship whether or not it ever processes a payment.
This is why retention, rather than acquisition, is where the agent layer does its deepest and least visible damage. A travel business can win the booking, deliver the trip well, and still lose the customer, not to a competitor but to the interface, because the next time that traveller wants to go somewhere they will ask the assistant on their device rather than return to the brand that served them last. The research shows the mechanism already forming. Studies of continuance intention find travellers building a habit of returning to the conversational AI itself, and work on proactive AI communication shows the agent making the first move, reaching out with a suggestion, which is precisely the gesture a brand relies on to earn a repeat booking. If the agent owns the remembered preference and the proactive contact, it owns the return.
The defence is not to fight the assistant for the device, which is a contest no travel brand can win, but to hold the things the agent cannot remember on your behalf. A memory of the guest that lives with you — rich, current and yours — is an asset the agent must consult rather than replace, which keeps you in the relationship rather than beneath it. Loyalty rebuilt as recognition, personalisation and recovery when something goes wrong, rather than as a points balance, gives the traveller a reason to return to the source instead of the interface, a shift I set out in Winning travel’s AI shift. And the human moment, the thing a guest actually remembers about how a place made them feel, is the one part of the relationship no assistant can hold, which is the argument of The AI Differentiator Dividend. The moat, then, is not only about being selected in the moment. It is about being the party the customer chooses to come back to, in a world where the easiest thing to come back to is the agent that is always already there.
The layer is forming above planning, not at booking
There is one more thing the evidence settles, and it explains where the toll will actually be levied. The transaction layer is not here yet. The two studies that examine agentic booking both design the purchase act out of scope: one benchmark states plainly that completing a purchase is not relevant to what it tested, while the other presents its autonomous, negotiate-and-book travel companion as a concept rather than a deployment, and places that capability only midway up its own maturity model, gated behind interoperable data standards, cybersecurity and unresolved rules on liability and auditability. That the leading work designs the purchase out is itself the finding. Autonomous booking is anticipated, not realised.
Set that beside the margin point and the picture resolves. The new layer is not forming at the moment of payment, where trust and liability still hold it back. It is forming above planning, at inspiration, discovery and the shortlist, where AI is already trusted and already used, for the reasons I set out in the companion to this piece on why AI is trusted to plan a trip but not to book one, and it can tax your access to the customer there without ever completing a booking itself. You can keep the transaction and still lose margin on the demand that feeds it. Which is exactly why the moat you build now has to defend the demand layer, not merely the checkout.
Auditing your moat against the right threat
The practical work, then, is to test your moat against reintermediation rather than removal, and three questions do most of it.
Does my moat give leverage against the layer, or only against removal? Be honest about which. Brand and scale mostly answer the disintermediation question. Data, exclusivity, legibility and the direct relationship answer the reintermediation one. You want the second kind, and you want to know which kind you actually hold.
Can a machine read me, and am I instrumented for it? The roughly one in ten travel companies that are genuinely agent-ready did not build cleverer AI; they changed what they measure, watching latency, room-mapping confidence and booking-success at speed, and they made their inventory machine-readable so the agent picks them. If your reporting stops at rank and conversion, an agent’s silent rejection never appears, because it looks like a competitor’s booking rather than an error.
Am I defending margin on purpose? The new take-rate is the strategic variable, not a cost to absorb quietly. Make channel profitability visible, which is what NRevPAR is for, model what an agent’s cut does to your unit economics before it is imposed rather than after, and decide deliberately whether you are the front door or the infrastructure, because you cannot fund both.
The bull case, in fairness
There is a genuinely optimistic reading, and it deserves stating. Travellers in the research value AI’s perceived commercial neutrality over the commission-driven motives of a human agent, so a layer trusted to be even-handed could convert better and match more honestly than the pay-for-placement funnels it replaces. A well-built layer could be cheaper than today’s stacked distribution chain rather than dearer, offering thinner margins, better matching and less drop-out, in which case it disintermediates the old toll rather than the supplier. Whether the new layer turns out to be a tax or an ally depends entirely on where you sit and how much of your value it can replicate, which is the very reason the moat question, properly framed, is the one that matters.
The honest answer
So, to the question the conference kept asking: yes, your moat will very probably protect you from disintermediation, and no, that is not the reassurance to want. Disintermediation was never the likely outcome. A new layer settling above you and billing you for the customer you assumed was yours is. The moat worth having is built for that: data the layer cannot source, supply it cannot route around, legibility so that it picks you, a direct relationship it cannot reach, and a reason to return that it cannot supply on your behalf. This is the same lesson the AI spine teaches from a different angle, that the firms which win are not the ones with the most AI or even the largest estate, but the ones clearest about what they control.
The transition is not hypothetical. IDC expects as much as thirty per cent of bookings to be executed by AI agents by 2030, and a majority of travel businesses are already experimenting with or scaling agentic AI. The layer is being built now, in public. When Phocuswright reconvenes in Fort Lauderdale in November, I suspect the sharper rooms will have stopped asking whether their moat keeps them from being cut out. The middleman is fine. A new one is arriving to sit above him, and the moat worth having is the one that decides how much rent you pay, and whether, once the trip is over, the traveller comes back to you or to the agent that is always in their pocket.
Sources
- The disintermediation-versus-margin reading of AI’s effect on the online agencies: BTIG, reported by PhocusWire, “AI risk to OTAs is margin squeeze, not disintermediation” https://www.phocuswire.com/news/online/ota-ai-margin-pressure-disintermediation
- The twenty-six-year account of reconfiguration and re-intermediation across the online travel ecosystem’s web APIs: R. Pujadas, E. Valderrama and W. Venters, “The value and structuring role of web APIs in digital innovation ecosystems: The case of the online travel ecosystem,” Research Policy 53(2), 2024, doi:10.1016/j.respol.2023.104931
- OTA transaction fees of fifteen to thirty per cent and the NRevPAR and RevPAC metrics: H. C. Boo, D. Remy and K. Lee, “Drivers, barriers, and challenges in NRevPAR and RevPAC adoption,” International Journal of Hospitality Management 127, 2025, doi:10.1016/j.ijhm.2025.104116
- AI adoption diversifying channels more sharply on the demand side: X. Liu, Y. Huang, R. Qi, L. Zhang and A. Alshalfan, “Centralization or diversification? Artificial intelligence application and supply chain configuration in hospitality enterprises,” International Journal of Hospitality Management 136, 2026, doi:10.1016/j.ijhm.2026.104643
- Agentic booking benchmarks that exclude the purchase act, and the machine-readability precondition for inclusion: S. Chemli, A. Vitale, Shekhar and M. Valeri, “When algorithms become travel planners: Benchmarking Agentic AI in Web 3.0 Tourism,” Journal of Engineering and Technology Management 80, 2026, doi:10.1016/j.jengtecman.2026.101959; and J. Yu, “Preparing for an agentic era of human-machine transportation systems,” Transport Policy 171, 2025, doi:10.1016/j.tranpol.2025.05.030
- The margin-compression reading of Booking Holdings and the AI Overview effect on paid listings: PitchGrade, “Booking Holdings: Online Travel AI and the Risk of Agent-Mediated Disintermediation” https://pitchgrade.com/research/booking-holdings-ai-margin-pressure
- The share of travel companies able to sell to an AI agent, and the IDC projection for agent-executed bookings by 2030: Gimmonix, “11% of Travel Companies Can Sell to an AI Agent” https://gimmonix.com/news/only-11-percent-agent-ready-2026
- Travellers forming a habit of returning to the conversational AI itself (continuance intention): H. C. Pham, C. D. Duong and G. K. H. Nguyen, “What drives tourists’ continuance intention to use ChatGPT for travel services?”, Journal of Retailing and Consumer Services 78, 2024, doi:10.1016/j.jretconser.2024.103758
- The agent making the first, proactive move in the relationship: M. Z. Zhou and T. J. Lee, “Helpful or hollow? Fit-to-feel alignment in proactive AI travel communication,” Journal of Retailing and Consumer Services 94, 2027, doi:10.1016/j.jretconser.2026.104999
- The contracting-layer framing and the optimistic, lower-cost reading of an agent layer: Travel Ecosystem, “The Agent That Books” https://www.travel-ecosystem.com/the-agent-that-books-who-earns-a-seat-at-the-hospitality-commerce-table-in-2026; and Nowah, “The Quiet Revolution: How AI Changes Travel Distribution” https://nowah.xyz/blog/quiet-revolution-ai-travel-distribution
Related on this blog: its companion piece, Why Is AI Trusted for Planning, but Not for Booking?, on why the barrier at the transaction is trust and infrastructure rather than capability (add the link once published); Winning travel’s AI shift on the first-party data moat and a switchable architecture; To succeed in luxury travel, sell access not assets on exclusivity as defence; The AI Differentiator Dividend on the human moment as the thing no assistant can hold; and my companion pieces on the AI spine, on why what you control matters more than what you own.
Analysis and interpretation are the author’s own. Figures are drawn from the cited peer-reviewed literature and public 2026 reporting and should be verified against the primary sources before republication.
