Why Can AI Disrupt Hotel Booking but Not Food Delivery?

Deep thoughts on AI and aspirations —— ByteDance Deep Thinking Circle

When Silicon Valley discusses how AI entry points impact platforms, the most widely circulated story carries a hint of irony, called the “DoorDash Problem”: in the future, you won’t open a food delivery app, but simply tell AI “order me a pizza,” and it handles everything, reducing the platform to a mere pipe. But when you run the numbers carefully, DoorDash itself is actually one of the safest platforms of all, while hotel booking platforms are the ones truly standing on the cliff’s edge. Both are platforms—why such different fates?

Platform Profits Are Essentially an Annuity

First, let’s look at how platforms make money.

Food delivery, ride-hailing, hotel booking, e-commerce—all these platforms burn money on the same thing in the first half: paying to acquire users. Once acquired, users develop the habit of opening the app directly, and every subsequent transaction costs nearly zero in customer acquisition. Upfront investment in exchange for long-term repeat purchases—that’s the platform’s annuity.

There’s another layer on top of the annuity: since users come on their own, advertising and recommendation spots on the page can be monetized. Amazon’s advertising brings in about $60 billion annually, contributing almost all of the e-commerce business’s profit, with over half of product display positions being paid promotions.

AI entry points threaten both layers. When users start asking AI instead of opening apps, the annuity breaks: platforms now have to pay traffic costs for every transaction all over again. What hurts more is the second layer—when transactions happen elsewhere, the advertising inventory disappears along with the transaction.

So far, no platform has actually died from this. What’s broken is the logic: all commercial designs that treat “users will come directly” as a permanent premise now find that premise shaken.

And this isn’t the first time this scene has played out. Search engines demonstrated the first half long ago: when Google did flight search, shopping, and local business search, it moved the value of “helping you find” directly into its own results page, slicing away the profits of platforms that lived on referral traffic. Those that survived were the ones that kept transactions and fulfillment in their own hands. AI entry points add something beyond search engines: they can not only find but also place orders for you. It’s completing the second half of a replacement that only went halfway last time.

The Dividing Line: Are You Selling “Finding” or “Backstopping”?

So why are hotel booking platforms first to break?

Break down a platform’s value into four pieces: helping you find (search, price comparison), helping you transact (ordering, payment), helping you backstop (refunds, guarantees, financial terms), and helping you fulfill (delivery, logistics, service delivery).

AI entry points have completely different capabilities for these four pieces. Finding is what they do best—understanding intent, multi-turn questioning, cross-source price comparison, all much stronger than a platform’s built-in search box. Transacting is being filled in; instant checkout capabilities are already rolling out. For backstopping and fulfillment, there’s no sign they want to do it. These two things mean building logistics networks, hiring drivers, tying up capital, handling disputes—all heavy-asset, low-margin hard work, the exact opposite of model companies’ light-asset, high-margin business model.

So to judge a platform’s safety, don’t look at industry buzz or tech narratives—just look at one thing: in its value proposition, how much is information and transactions, and how much is responsibility and fulfillment?

Two variables amplify or reduce risk: usage frequency and decision weight. Ride-hailing and food delivery are high-frequency, low-decision-weight activities—users open the app out of habit without thinking, leaving no room for AI to insert itself. Vacation trips or buying ski equipment are low-frequency, high-decision-weight purchases where users naturally want to research and compare—this is AI’s home turf. Even if you’re a loyal user of a certain platform, when booking a long vacation, you’ll likely ask AI first.

The Counterintuitive Ranking: The More Digitized, the First to Fall

Applying the above framework to reality yields a ranking that contradicts intuition.

The hotel industry was one of the earliest and most thoroughly internetized industries. Product standardization, data structuring, real-time inventory syncing—all done twenty years ago. Back then, this was a huge competitive advantage; now it’s precisely the weak point: the more standardized and readable the supply, the easier for AI to aggregate. When model companies want suppliers to open accounts and integrate new channels, hotels have ready-made enterprise customer teams to serve them, and both sides hit it off.

Compare that to food delivery and ride-hailing supply: millions of local restaurants, fragmented vehicle capacity, and every order needs someone managing delivery—on time, hot, and handling refunds when problems arise. Aggregating this kind of supply can’t be done with APIs alone. Vacation rentals fall between the two—more dangerous than hotels, safer than food delivery. Aggregating millions of rooms, each with its own layout and rules, took Airbnb many years and billions of dollars—photos, amenities, quality all had to be made readable one by one. This kind of hard work can’t be quickly copied by AI, but it also can’t defend against anything outside high-frequency scenarios: travel planning itself is low-frequency, high-consideration purchasing, and traffic on the entry point side will still leak out bit by bit.

In other words, industries that were most thoroughly digitized in the last era surrender the fewest defensive lines in this era. The standardization painstakingly built back then has now become an instruction manual readable by anyone.

There’s a second-order backlash hidden here, which I call the settlement of the advertising tax. Mature platforms almost all tax user experience: more and more sponsored positions stuffed into search results, rankings tilted toward higher bidders, commissions raised repeatedly. When there was no alternative, users endured it. Now AI entry points have appeared, and the first gift they hand users is a recommendation without ads. Every cent of tax platforms collected over the past twenty years has become a reason for users to leave. This is no longer a moral question—the competitive structure has changed: the portion of profit gained by degrading experience now has a clearly priced competitor for the first time.

Four Ways Forward for Platforms, and One Boundary

If you’re running a platform, or looking at platform investments, the response sequence is actually quite clear.

Do what AI won’t do. Deepen backstopping and fulfillment: insure transactions, build proprietary delivery, sign exclusive supply. These activities have low margins and slow returns, but precisely because of this, model companies won’t follow. This is the only place you can’t be moved from.

Keep search experience competitive. You don’t need to lead—keeping the gap within twenty or thirty percent is enough; make up the rest with delivery and after-sales. If you fall behind tenfold, users will leave completely.

Those with high market share should dare to drag their feet. Doing e-commerce entry points can’t bypass the top platforms, which hold inventory and fulfillment networks—they’re the needed party. At the negotiation table, slow is leverage; rushing to integrate means devaluing yourself. If you must cooperate, keep data and choice rights in your own hands.

The last one: while AI hasn’t acted yet, self-inflict the advertising tax wound first. Passively waiting for someone else to disrupt you versus proactively cutting monetization that degrades experience—the two outcomes are vastly different.

The boundary also needs to be stated clearly. This judgment has a premise: model companies stick to their light-asset DNA, only doing the understanding and decision layers, not touching fulfillment. If one day a model giant really invests heavily in building logistics networks, this safety ranking will need to be redrawn. But based on their current business models, the probability of that happening isn’t high.

Back to the opening question. Next time you look at a platform business, just ask one thing: is it selling “helping you find” or “helping you backstop”? Those selling finding—the lease is up. Those selling backstopping—can still collect rent for many more years.