How the Agent Web Gets Built
Why incumbents will toll themselves out of relevance
AI agents are becoming the new discovery layer, sitting above Booking.com, LinkedIn, and the platforms that currently own consumer demand. The structural move that hotels spent fifteen years trying to make is now available to any startup with the right infrastructure. The categories where it works, and where it collapses on contact, are written in the loudest defensive moves incumbents are making right now.
At the happy hour we (NextView) hosted in San Francisco last week, I was catching up with one of the Gondola cofounders. Gondola is one of our portfolio companies, a consumer travel startup. Their pitch is simple: search for a hotel, compare cash and points across every major chain, book in one click, get the loyalty credit as if you’d booked direct with Marriott or Hilton.
I use AI agents heavily in my own work and home life, including a personal trip-planning project where I tried to get an agent to compare rewards calendars and monitor pricing. I quickly gave up the attempt because travel websites are some of the most locked-down on the consumer internet: bot detection, fingerprint checks, session tokens, captcha walls.
So I was very curious how Gondola does it.
He described what’s underneath the consumer UI as “a whole architecture in the background.” The outcome is publicly visible on Gondola’s MCP: every booking goes direct to the hotel (Marriott, Hilton, Hyatt, IHG, Accor, Wyndham), preserving member rates, loyalty point accrual, status nights, and credit card rewards. The product is the one-click consumer UX; everything underneath stays invisible.
The question that conversation raised wasn’t how they did it. It was: why did a consumer travel startup have to build infrastructure that big just to ship a booking flow?
The framework underneath
A decade ago, Ben Thompson named the structural rule of the internet age: aggregators own demand. Once digital distribution became free, the contest shifted to who controlled the consumer interface. The winners (Google for content, Booking.com for hotel rooms, OpenTable for restaurant tables, LinkedIn for professional identity) all followed the same template. They captured demand; suppliers came onto the platform on the aggregator’s terms.
Thompson updated the theory in 2019 to acknowledge a blind spot: where supply is concentrated, aggregators can’t commoditize it. Where supply is fragmented (hotels, restaurants, individual professionals), they can.
The framework governs content, networks, and transactional marketplaces alike. What’s new in 2026 is that AI agents are becoming the next demand-side layer. When your agent searches for a hotel, the agent IS the new Booking.com. The question is which seats at this new game are still open.
What hotels couldn’t do alone
Take lodging. Hotels have spent fifteen years trying to escape Booking.com and Expedia. In 2016, Hilton ran “Stop Clicking Around,” its biggest marketing campaign in 97 years. Marriott launched “It Pays to Book Direct” the prior September. Every major chain offered exclusive lower rates for loyalty members who bypassed OTAs.
The campaigns failed structurally. From May 2016 to 2017, Marriott’s US online-bookings share fell 37%, Hilton’s fell 6%, and Booking.com kept growing 22% year-over-year. Expedia spent $4.3 billion on marketing in 2016; Priceline Group spent $3.5 billion on performance advertising alone. Hotels spent a fraction of that. They couldn’t win the discovery layer because they couldn’t outspend the discovery layer.
The OTA’s moat was never price or convenience; it was demand discovery. As long as consumers searched through Booking.com, the OTA tax held even when hotels offered the same rate directly.
By 2026, two structural shifts had landed. A September 2024 European Court of Justice ruling1 weakened the contracts Booking.com used to keep hotel pricing uniform across channels, finally letting hotels undercut OTAs on their own websites. And AI agents had started becoming the way consumers search.
This is what the Gondola team built around. Agents are the new discovery layer, but they don’t reach consumers as raw developer tools; they reach consumers wrapped in products like Gondola. Hotels don’t need to outspend Booking.com on marketing if a consumer’s agent layer routes the booking directly to the hotel. The structural move that hotels couldn’t make alone in 2016 is now possible because the discovery interface itself moved.
When this works, when it doesn’t
Three questions determine whether the Gondola pattern holds in a given category.
One: who controls the supply? Hotels existed long before Booking.com. They have brand.com sites, direct phone lines, and loyalty programs of their own. Booking.com was a tax layer on top of pre-existing supply, never the supply itself.
Two: is the irreplaceable supply re-aggregatable through a different layer? Marriott isn’t replaceable; any credible hotel platform must carry Marriott. But Marriott IS reachable directly. The same supply that Booking.com aggregated can be aggregated again, by a new layer, through the chains’ own direct channels.
Three: can the agent tech stack actually do the work? The effort Gondola had to put in was non-trivial. The real challenge was building the infrastructure layer that lets agents reliably and repeatedly integrate with hotels on the web. That's where trust is built with the end customer, and it also answers whether the technical move is doable for a startup, not just possible in theory. When all three hold, agents do what hotels alone couldn’t: bypass the discovery layer without having to outspend it. The move collapses if any one fails.
Airbnb and Uber are harder cases because the aggregator created the supply category alongside its marketplace. There’s no parallel pool of Airbnb-style hosts organized around a trust infrastructure that Airbnb doesn’t control, and Uber drivers don’t exist as a service category outside the platform. But the opportunity isn’t categorically gone. A meaningful chunk of Airbnb’s supply runs through professional property managers who are individually reachable, and Airbnb’s own consumer search experience is dated enough that an agent layer wrapping a better UX could compete from above. The work is bigger than the lodging case: condition two (supply re-aggregability) is partial rather than clean, and condition three (technical lift) compounds with a demand-side UX rebuild. The real driver of difficulty isn’t whether the aggregator created the supply; it’s how identifiable and grouped that supply is. Truly atomized single-unit individuals stay hard.
Amazon’s stronghold isn’t supply lock-in; it’s demand-side experience. Most of Amazon’s catalog is third-party sellers with their own direct channels, so supply is technically reachable. What holds Amazon together is the demand-side UX: superior logistics compounded by Prime membership benefits that most consumers would not want to live without. When Perplexity’s Comet agent began making purchases on Amazon’s marketplace last fall, Amazon sued, and a federal court issued a preliminary injunction blocking Comet in March (currently under appeal). When the aggregator’s demand-side experience is genuinely superior, consumers won’t leave even if supply is reachable, and condition two effectively fails.
Independent retailers don’t fit because the technical work isn’t doable yet. OpenAI shipped Instant Checkout in fall 2025 with a small set of retail partners and pulled it back in March 2026, pivoting to merchant-native apps instead. Catalog accuracy, multi-item carts, sales tax, and checkout reliability all broke against the long tail of independent retailers. Condition three fails for the merchants that don’t have real-time programmatic infrastructure ready.
Four patterns from different answers
Different answers to those three questions produce four visible positions. Two are startup moves; two are incumbent moves, with most categories producing a fight between them.
From below: disintermediate the aggregator with willing supply. Gondola is doing this in lodging: hotels are pre-existing supply with direct channels and (now) genuine incentive to escape OTA economics. The agent does what the supplier alone couldn’t: route around the discovery layer without having to outspend it. The moat is the invisible work: the year of integration logic Gondola spent on the architecture that makes every booking land as a direct booking, with the right loyalty credit, across every major chain.
From beside: rebuild parallel where the supply IS the network. Boardy (disclosure: a NextView portfolio company) is attempting this for professional networking. LinkedIn’s value isn’t the platform; it’s the graph: your professional network, your reputation, your second-degree connections. Because LinkedIn’s graph isn’t extractable in practice anymore2, Boardy is building parallel: voice-mediated agent introductions among opt-in participants, a connection layer that doesn’t extract LinkedIn’s graph but builds an alternative one through new participation. (Boardy recently celebrated its 100,000th connection, gathering richer signal through voice, chat, text, and email than any static profile can.)
From within: go headless on your own terms. Shopify is the cleanest case. Their customers were always merchants, not consumers; Shopify has been infrastructure, not a consumer brand, from day one. Extending that infrastructure to be agent-readable (MCP integration, the Universal Commerce Protocol co-developed with Google, agentic storefronts available to non-Shopify merchants too) is the natural agent-era progression of the merchant-infrastructure strategy. The bet: when buyers come through agents, the rails merchants run on become the strategic surface.
From above: tollgate the agent traffic. SAP issued a policy prohibiting customers from using OpenClaw and other AI agents that “plan, select, or execute sequences of API calls” without official sanction. ServiceNow launched Action Fabric, explicit metered access for agents accessing app data. JPMorgan’s Mark Murphy described ServiceNow’s pricing as “effectively a tax on customers using outside AI agents to interact with data they already store in ServiceNow’s apps.” The bet is that switching costs keep customers paying the toll rather than leaving. For decades, that bet has worked: replacing an ERP is painful enough that customers absorb pricing changes rather than migrate. The agent era changes the math. When the agent experience inside SAP or ServiceNow is degraded by policy, customers don’t just pay; they route AI-native workflows through tools that work better at the application layer above. Over time, those tools accumulate the data the system of record used to own. The toll gets extracted in the short term, but the underlying system gets disintermediated from above in the long term.
Most incumbents will pick #4 because it feels like preserving optionality: you don’t shut the door, you charge for it. But #4 is the worst long-term position: it announces “we are the obstacle,” which invites a Gondola-style response from below or a Boardy-style rebuild from beside. The winners actively pick #3 (build the rails on your own terms) or get rebuilt around without a fight.
The harder question
This post answers which categories admit the disintermediation move. Who earns the right to build the user-facing agent surface is the harder question, for a separate piece. The demand-side build isn’t trivial: the new agent has to deliver a step-functionally better value prop than the incumbent, not just a better UX. The incumbent’s marketing budget, funded by the very take rate the disintermediator is trying to break, will outspend anything merely good. The bar is qualitative superiority on the fundamental value, not the surface, on both consumer and B2B agent interfaces.
Gondola’s hidden architecture isn’t infrastructure. It’s a head start.
Previously in Ground Truth: The Build-vs-Buy Reset (workflow innovation from the top of the application layer) · The Context Gate (agent access constraints, the prior framing of the access-route question).
The ECJ held that rate-parity clauses fall within the scope of antitrust review under Article 101(1) TFEU, removing a procedural shield Booking.com had used to block national courts from examining them. The ruling didn’t outlaw the clauses outright; it sent them to national courts for case-by-case examination.
The Ninth Circuit affirmed in 2022 that scraping public LinkedIn data does not violate the CFAA, but hiQ v. LinkedIn ended in a 2022 settlement with a permanent injunction against hiQ. LinkedIn’s contractual enforcement (Terms of Service, breach claims) and technical enforcement (rate limits, CAPTCHA, account flagging) make systematic graph extraction operationally untenable regardless.