Expedia processed $119.6 billion in gross bookings in 2025.

Around 415 million room nights. Roughly 720 million site visits.

How much of that infrastructure is now being consumed by AI agents that never book?

A property detail page on Expedia runs to several hundred kilobytes of rendered HTML. For a human, that resolves into a price, a star rating, and a booking button. For an agent, it costs four to twelve thousand input tokens to parse, with three to seven properties consumed per comparative query. The session terminates inside the agent. The booking, if it happens at all, arrives with no recoverable attribution.

This is the agentic shadow load. For Expedia specifically, the cost shows up in three layers.

Layer one. The parse itself.

Direct serving cost is small per page and modest in aggregate. Assume two to six percent of Expedia's annual property page views are now agent-initiated. At roughly half a cent per dynamic page assembled (bandwidth plus backend compute on pages whose pricing and availability cannot be edge-cached), the direct cost runs between one and ten million dollars per year. Real, but not the headline.

Layer two. The broken attribution loop.

When an agent passes a deep link through to a booking page, UTM tags, affiliate cookies, and referrer headers do not survive the handoff from headless agent session to clean browser tab. Expedia records the booking as direct. Their paid marketing channels and meta-search partners go on getting paid for clicks that are no longer doing the work. Marketing budget is misallocated against an attribution model that has quietly stopped reflecting where demand originates. For a company spending several billion annually on direct sales and marketing, even a one to two percent attribution distortion is a nine-figure misallocation surface.

Layer three. The recommendation gap.

This is the layer that matters. Across our audits, the average brand elimination rate before final AI recommendation runs at 87 percent across multi-turn reasoning sequences. The brand was cited. The brand was not chosen. The model lost confidence in the evidence under conversational pressure and substituted elsewhere.

If three to five percent of Expedia's would-be 2026 gross bookings are decisioned inside an agent, and Expedia's legibility under that probing is materially weaker than its inventory and rate competitiveness predict, the revenue exposure runs into the hundreds of millions of dollars. Five percent of $119.6 billion is $5.98 billion in gross bookings exposed. At the disclosed 12.3 percent take rate, that translates to approximately $735 million in revenue exposure. Recover even a third of that gap through legibility remediation, and the math closes quickly.

The position paper out today names this Agentic Brand Control. The discipline of governing how a brand is read, cited, and recommended by AI systems. The instrument is brand.context. The diagnostic is AIVO Meridian.

Expedia is not the only OTA exposed. It is the cleanest example because the scale is legible.

The shadow load is the cost of being read badly. At this scale, that cost is not theoretical.

Empirical cohort paper, auditing five booking surfaces under PSOS resilience testing, follows.

The Shadow Load: Who Pays for Agentic Travel Search?

Agentic search consumes online travel agency infrastructure at scale without the booking signal that historically justified the serving cost. A single property detail page costs an agent four to twelve thousand input tokens to parse, with three to seven properties typically consumed per comparative query. The cost falls on the model provider in inference, on the OTA in serving, and on the user in latency. No party recovers value through the conventional booking funnel. This paper defines the agentic shadow load, examines the inversion of finder and found in agent-mediated search, and proposes a paired diagnostic and remediation architecture. The AIVO Meridian PIM diagnostic, scored across RCS, PSOS, and CSR, identifies where product evidence is illegible to agents under conversational pressure. brand.context, a signed JSONLD evidence layer, restores legibility and reduces token cost by an order of magnitude. Together they constitute the operational practice of Agentic Brand Control.

ZenodoAIVO Journal and de Rosen, Timothy