I think a lot of the OTA vs Muse debate comes down to a difference in what can theoretically be built versus what is actually likely to be built. It is very easy to image a world where an AI Agent can scan every OTA and every single hotel's web listing then come back to you with the best options (especially price-wise). In this event $BKNG and $EXPE will loose some sales to direct traffic where pricing is more favorable. This is why Agentic AI is such a scary proposition: it makes the consumer a much more rutheless travel shopper while making them omniscient. However, it misses what is the path of least resistance to building an actually product today (and at least for the near future). It is extremely compute intensive to web scrab every hotels website. AI agents do not have a live inventory link, so you'd have to go and try to book each room to see if it is available... and this can of course change, so you need to keep updating it. Hotel pricing is also dynamic and can change very quickly. OTAs pricing especially changes fast and can often be discounted for logged in users. (Side note, but worth mentioning that the OTAs often block transact for rooms at discounts, so their rates can be lower than going direct. You usually have to be logged in to see the best discounted rates) All this means it is just a very inefficient way to build a product. Imagine you are looking for a book in a library. There is a librarian who knows where every book is and what every book is about. However, you refuse to talk to her. Instead, when you looking for books, you decide to pick up every book in the library and read it yourself. Only then do you decide what book you were looking for. It gets even crazier though. Because you have to do this every time you are looking for a book. Booking already built out the infrastructure so you can "know where every book is and what it is about" without having to do it yourself every time. So the AI agent / OTA debate isn't about what can theoretically happen. Obviously an AI agent in theory could find you any hotel and book it without OTAs. However, does that really make sense from a product standpoint to rebuild all of this infrastructure every time a user queries? Especially when we consider the compute-intensity of it and time it takes to do so. (Not to mention the fact that Booking does often have the same or lower prices. See my example the quoted tweet below) It's an interesting question because Muse already partners with Expedia. ChatGPT has Booking and Expedia partnerships. Aren't the AI companies saying they need the OTA piece to offer a better use experience today? That doesn't mean the OTA's sponsored ads business can't be under pressure, but there could be some tailwinds too: more unpaid traffic directed to them from AI and they can shift ad budget from Google Search to different AI channel. The diversity of AI channels is likely to result in lower ad costs than when they were beholden to just Google. So yes AI is scary and will change things in ways that are hard to foresee, but is this really an existential risk or just one of many risks businesses face?

Real example of Muse and the OTA risk. $BKNG $META I wanted to try what happened when I used Muse to book a stay in Florence, Italy. First it returned 5 options. The links to ALL FIVE were Booking .com listings. It did not prompt me to price check or go direct. I did prompt it to check if it was cheaper going direct. From the time I asked it to check to when it answered me 14 minutes had passed- sure I could do other things in the mean time, but if I was really in the mood to search travel, I would have realistically just gone to different sites or did the check myself. The conclusion was 2 out of the 5 listings were cheaper to go direct, with the other 3 cheaper on Booking .com. Although 1 of those was only cheaper because it was a non-refundable rate-- the refundable rate was the same as on Booking. (The one that was cheaper was because it had a special 15% discount if you stayed longer than 5 nights). So really more like 1 out of 5 was clearly cheaper going direct, 1 was a wash, and 3 were more expensive going direct. The interesting thing is people are acting like the abilit to price compare and shop rates is new with Muse. Yes, I get it makes it more automatic, but a lot of travelers check the direct rate before booking. If the price is similar enough, they may still prefer to book on a better known site like Booking because of the support they expect to receive if something is wrong. But also, who books hotels like this? No pictures, no maps, no reviews. I have to trust it figured out the best one for me. It makes sense as a place to start a search (better than Google perhaps), but for me at least this wouldn't satisfy me with enough info before booking. The long wait time inbetween prompts would have compelled me onto a more traditional OTA to speed things up. I get the interface and speed can/will improve in the future, but it still seems like the path of least resistance to just pull from Booking. It may keep price checking and some consumers will save money from that, but it isn't built to scrap the entire web everytime a user searches for a hotel. As Ben Thompson notes, in time they can build out their own web directory, but that still doesn't solve the live hotel inventory link. I get AI Agents are scary and they will change the world in all sorts of ways that are hard to know ahead of time. On the other hand, it is very hard for a single business to be great at everything.

Real example of Muse and the OTA risk. $BKNG $META I wanted to try what happened when I used Muse to book a stay in Florence, Italy. First it returned 5 options. The links to ALL FIVE were Booking .com listings. It did not prompt me to price check or go direct. I did prompt it to check if it was cheaper going direct. From the time I asked it to check to when it answered me 14 minutes had passed- sure I could do other things in the mean time, but if I was really in the mood to search travel, I would have realistically just gone to different sites or did the check myself. The conclusion was 2 out of the 5 listings were cheaper to go direct, with the other 3 cheaper on Booking .com. Although 1 of those was only cheaper because it was a non-refundable rate-- the refundable rate was the same as on Booking. (The one that was cheaper was because it had a special 15% discount if you stayed longer than 5 nights). So really more like 1 out of 5 was clearly cheaper going direct, 1 was a wash, and 3 were more expensive going direct. The interesting thing is people are acting like the abilit to price compare and shop rates is new with Muse. Yes, I get it makes it more automatic, but a lot of travelers check the direct rate before booking. If the price is similar enough, they may still prefer to book on a better known site like Booking because of the support they expect to receive if something is wrong. But also, who books hotels like this? No pictures, no maps, no reviews. I have to trust it figured out the best one for me. It makes sense as a place to start a search (better than Google perhaps), but for me at least this wouldn't satisfy me with enough info before booking. The long wait time inbetween prompts would have compelled me onto a more traditional OTA to speed things up. I get the interface and speed can/will improve in the future, but it still seems like the path of least resistance to just pull from Booking. It may keep price checking and some consumers will save money from that, but it isn't built to scrap the entire web everytime a user searches for a hotel. As Ben Thompson notes, in time they can build out their own web directory, but that still doesn't solve the live hotel inventory link. I get AI Agents are scary and they will change the world in all sorts of ways that are hard to know ahead of time. On the other hand, it is very hard for a single business to be great at everything.

With everyone talking about $META’s Muse, I asked Gemini how my thesis that $BKNG is not at direct risk from AI Agents has held up. TLDR is that: 1) AI Agents may use Booking but they aren’t likely to want to replace what Booking already does. 2) BKNG pays for a lot of sponsored ads already so that money could go to Meta instead of Google 3) They have a big direct booking business (their app) with a loyalty program and those users are harder to move purchase activity from (however, displacement here means higher ad costs, lower margins, but not an existential risk) ~*~*~ Summary of Drew Cohen’s Argument In his deep dive on Booking Holdings, Drew Cohen addresses the prevailing bear case: autonomous AI agents will scrape hotel websites, bypass Online Travel Agencies (OTAs), and render Booking.com obsolete. His rebuttal argues that Booking.com’s moat is operational, regulatory, and logistical—not just a search interface: The Integration Barrier: Scraping the open web for live travel inventory is brittle and compute-intensive. In Europe (Booking’s core profit center), supply is heavily fragmented across 100,000+ independent boutique hotels. These properties lack standardized APIs or the technical capability to host server-side agent protocols (like Anthropic’s MCP). The "Merchant of Record" (MoR) Problem: Direct booking requires taking on fraud prevention, chargeback risks, cancellation management, dynamic foreign exchange (FX), and compliance with complex regional regulations (such as the EU's DMA, DSA, and GDPR). Cohen argues frontier AI labs have no desire to become the Merchant of Record. Customer Service & Operational Liability: Travel is a high-stakes, non-deterministic purchase. If an AI hallucinates or an overbooking leaves a traveler stranded in a foreign city at 2 AM, tech labs will not run a 24/7 human customer service operation to rebook them. The Intermediary Becomes the Infrastructure: Because of the friction above, frontier AI companies will opt to plug directly into Booking.com’s aggregated APIs and live inventory feeds rather than trying to rebuild the entire plumbing of global travel. Does Meta’s Muse Invalidate the Thesis, or Does It Hold Up? Meta’s launch of Muse—which operates inside a cloud virtual machine (Muse Secure VM), drives a desktop browser, and checks out using Stripe Link—directly tests Cohen's thesis. The thesis largely holds up, but Muse breaks some of Cohen's technical assumptions while validating his broader economic conclusions. Where Muse Challenges Cohen’s Arguments 1. It Bypasses the Need for APIs or MCP Cohen argued that 100,000 independent hotels would never adopt modern server-side agent protocols. Muse proves they don't have to. Because Muse controls a virtual browser that reads the DOM, clicks buttons, and fills out forms just like a human traveler, it can theoretically execute a reservation on any hotel's standard website without that hotel building an integration. 2. It Evaporates the "Merchant of Record" Barrier Cohen asserted AI labs wouldn't facilitate direct bookings because they refuse the liability of being the Merchant of Record. Muse sidesteps this entirely by acting as a client-side proxy: it deploys single-use virtual cards generated via Stripe Link, autofilling the guest's payment directly into the merchant's existing checkout form. Meta never acts as the Merchant of Record; the hotel (or OTA) retains that liability. Why Cohen’s Core Thesis Still Largely Holds Up Despite overcoming those technical hurdles, Muse runs squarely into the structural realities Cohen highlighted: 1. Muse Still Navigates to Booking.com If a user instructs Muse to "find and book a 4-star boutique hotel in Florence under $250 with free cancellation and great breakfast," Muse does not attempt to scrape and parse 80 separate Italian boutique hotel websites. Cont…

I just released my first investor letter. You can read it here: drewcohencapital.com/investo…

Me getting roasted on my own podcast Nevertheless, new Dialogue out this week on $ADBE