AI referrals drive under 1% of Booking Holdings room nights as internal code output rises 30%

Referrals from large language models, paid and unpaid, represent under 1% of total room nights at Booking Holdings, according to the company's chief financial officer Ewout Steenbergen. Speaking at Fortune's AIQ Summit at the New York Stock Exchange on October 1, 2026, Steenbergen outlined how the travel giant is managing artificial intelligence costs while the direct consumer impact remains minor.

The travel conglomerate, which owns platforms including Booking.com, Agoda, Priceline, and OpenTable, is balancing internal productivity gains against massive external technology investments. While the companies building the foundational models spend heavily, Booking Holdings is focusing on using existing infrastructure to optimize its operations and marketing.

What the data shows

The financial figures surrounding artificial intelligence reveal a stark contrast between internal corporate efficiency and external consumer adoption. On the consumer side, the volume of bookings driven directly by artificial intelligence remains low. Steenbergen reported that referrals from large language models, including both paid and unpaid sources, accounted for under 1% of total room nights during the company's second-quarter earnings call in August.

In contrast, internal applications of the technology show clear operational gains. Booking Holdings employs roughly 9,000 engineers. These engineers are currently getting about 30% more code into production. This measurement counts only merge requests that successfully pass the company's testing and quality control protocols.

Operational costs are also shifting within customer service. Booking Holdings reported that its total bookings are up at a high-single-digit rate, while customer service costs are slightly down. This combination has caused the cost per booking to fall by a lot, even as customer satisfaction metrics rose over the same period.

The company's marketing distribution remains heavily reliant on traditional digital channels. About one-third of Booking Holdings customers arrive through paid channels, which include search engines, social media platforms, and metasearch engines. Booking Holdings spends $8 billion to $9 billion a year on these paid acquisition channels. The remaining two-thirds of its customers access the platform directly. To retain these direct users, the company rolled out artificial intelligence features in 2025 aimed at travel discovery and customer support.

The performance of Booking's proprietary consumer-facing tools shows modest positive movements. Travelers who use these tools take a little less time to book their trips, convert at slightly higher rates, and cancel their reservations slightly less often. However, Steenbergen described these trends as early-stage developments based on a small amount of data.

Broader corporate metrics show steady performance across other initiatives. Connected trip transactions, where travelers book more than one travel vertical for the same trip, grew at a low-double-digit rate at Booking.com. Merchant bookings reached about 73% of gross bookings. Additionally, Level 2 and Level 3 Genius members made up more than 30% of active customers.

Meanwhile, the broader technology sector is spending heavily on infrastructure. According to projections by JPMorgan, global artificial intelligence capital expenditure is expected to reach $800 billion in 2026 and rise to $1.1 trillion in 2027. JPMorgan also estimates that artificial intelligence capital expenditure will consume about 93% of hyperscalers' operating cash flow in 2026, a significant increase from the 33% recorded in 2023. Financial analysis firm Morningstar has cited a potential payback period of only three to five years for this infrastructure, meaning these massive investments must show returns before 2030.

How it was measured

The operational and financial metrics detailed by Booking Holdings were disclosed by Ewout Steenbergen, the company's executive vice president and chief financial officer, during the Fortune AIQ Summit on October 1, 2026. The figure regarding large language model referrals representing under 1% of total room nights was first reported during the company's second-quarter earnings call in August.

The engineering productivity metric, which shows an approximate 30% increase in code production, is measured internally by tracking merge requests that pass automated testing and quality control. Booking Holdings did not publish the specific technical parameters of these tests or the baseline period used to calculate the increase.

The external capital expenditure figures and cash flow percentages were calculated and published by JPMorgan. The estimated payback period of three to five years for infrastructure was calculated and published by Morningstar. Booking Holdings did not publish its internal methodology for tracking customer satisfaction or the precise formulas used to determine that the cost per booking has fallen by a lot.

Managing the token bill

To prevent technology expenses from eroding its profit margins, Booking Holdings has implemented a strategy it calls effective model cost routing. Rather than running every query through the most advanced and expensive artificial intelligence models, the company routes simpler tasks to basic or open-source models. More expensive, proprietary models are reserved strictly for complex processes.

The company manages its engineering teams by measuring total IT cost per merge request. This metric combines the cost of human labor with the cost of artificial intelligence tokens. Under this system, token spending is allowed to rise, provided that the overall cost of each merge request reaching production declines.

“The returns will be there for those processes that are being redesigned end to end, but you also have to make sure that those costs on the other hand are not going out of control.”

This disciplined approach to cost routing is designed to offset the high fees associated with advanced large language models. By establishing a clear metric that balances human developer hours against token consumption, the company seeks to maintain a predictable cost structure as it rebuilds its processes from a white sheet.

The shift to high-frequency travel

Booking Holdings is attempting to transition its business model from a transactional platform to a high-frequency service. Currently, travelers visit about five platforms on average before making a booking. The company aims to use artificial intelligence to reduce this research phase and manage travel disruptions actively, a concept it refers to as the connected trip.

If a traveler experiences a flight delay that impacts a restaurant reservation, the company plans to use automated tools to reschedule the booking proactively. For example, if a traveler has a trip to Paris and the weather forecast for Wednesday is poor, the system could proactively offer to swap outdoor walking tours for indoor museum visits, such as a tour of the Louvre.

This level of proactive engagement is intended to build loyalty and brand value, converting occasional bookers into frequent users. Early signs of this transition are visible in the low-double-digit growth rate of connected trip transactions on Booking.com, where customers book multiple travel verticals for a single trip.

The CFO's AI coach

To better understand the technology, Steenbergen has adopted artificial intelligence tools into his own daily workflow. He revealed that he has an AI coach and intentionally speaks about this arrangement inside the company to encourage continuous learning among staff. He also employs two automated agents to assist with his corporate responsibilities.

The first agent, which he calls a strategic thought partner, helps Steenbergen prepare board presentations and develop long-term strategic plans. The second agent, a critical equity research agent, helps him prepare for the company's quarterly earnings calls. Steenbergen highlighted that adopting these tools is part of a broader corporate push to normalize artificial intelligence education.

“No one knows this. If I have to learn, and I have a coach, it's very normal. Everyone has to learn.”

Who the spending affects

The dynamics of artificial intelligence spending affect public market investors, short-term rental platforms, and property managers in different ways. For investors holding shares in major technology hyperscalers, the projected capital expenditure eating up about 93% of operating cash flow represents a significant risk. This high level of spending limits the capital available for direct shareholder returns and leaves these companies vulnerable if market demand fails to meet expectations.

For travel platforms and short-term rental operators, the situation is different. Companies like Booking Holdings can benefit from the advanced models funded by hyperscalers without absorbing the immense capital risk of building them. However, this reliance introduces a dependency. If hyperscalers face pressure to monetize their models quickly due to the three to five years payback window cited by Morningstar, they may alter their licensing fees or access terms.

For property managers and hosts listing inventory on Booking.com, Agoda, or Priceline, the shift toward merchant bookings and connected trips alters how bookings are secured. With merchant bookings reaching about 73% of gross bookings, the platform is taking greater control over the payment process. If artificial intelligence tools begin actively rescheduling trips due to weather or flight disruptions, hosts may see automated adjustments to check-in dates or activity bookings, requiring closer integration with the platform's automated systems.

What the data does not tell you

While the reported figures highlight operational improvements, they do not reveal the exact monetary cost of Booking Holdings' artificial intelligence initiatives. The company has not disclosed its total spending on model licensing, token consumption, or external cloud infrastructure. Without these specific figures, investors cannot independently verify the net financial impact of the 30% increase in code production.

The referral data is also limited. While referrals from large language models are under 1% of total room nights, this figure does not capture how many travelers use external artificial intelligence search engines for early-stage inspiration before navigating directly to Booking's platforms. It remains unclear whether these models are acting as a primary discovery tool that eventually redirects to traditional search channels.

Finally, the modest gains reported in booking speed, conversion rates, and cancellation reductions are based on early-stage data. The company has not published the absolute numbers behind these metrics, making it difficult to assess whether these improvements are statistically significant or sustainable over a longer holiday booking cycle.

What to watch

As artificial intelligence tools are integrated deeper into short-term rental distribution channels, hosts and managers should monitor several key operational areas:

• The share of bookings originating from direct channels versus the $8 billion to $9 billion paid marketing channels.
• Changes to reservation terms and cancellation policies driven by automated connected trip rescheduling features.
• The ratio of merchant bookings on major platforms, which currently stands at about 73% of gross bookings at Booking Holdings.
• The pricing and licensing terms of third-party artificial intelligence models as hyperscalers face pressure to show returns within the three to five years payback window.