Ask an AI assistant to help book a flight and it will give you an answer. The question for airlines is where that answer comes from.
ATPCO commissioned research with Bain & Company and Profound to examine how AI assistants represent airline products when travelers ask for help finding flights. The findings, published in the report How AI understands airline products, raise questions about how airlines are represented in AI-mediated shopping and how much control they retain over that process.
ATPCO CEO Brett Burgess, who took the role in February 2026, addressed the research findings, the implications for airline distribution, and the company’s plans for making airline retailing work in an AI-mediated environment.
“The results weren’t comfortable”
“We recently engaged Bain to test how a dozen major carriers show up when a traveler asks an AI assistant to help book a flight,” Burgess says. “We ran five routes per airline, in local languages, across six AI platforms in eight countries. That produced more than 12,000 answers and roughly 64,000 cited sources, and the results weren’t comfortable.”
He points to three findings.
The first is omission. “Airline-owned sites accounted for only about 6 percent of citations, even on each airline’s own five busiest routes, while OTAs were cited roughly three times as often. An airline’s flights could appear in an answer without its website being used as a source at all.”
The second is errors. “When assistants described airline products, they sometimes got them wrong, presenting a premium economy cabin as business class or denying that a carrier offered a lie-flat seat when it did.”
The third is volatility. “Home-market visibility ranged from above 80 percent on one assistant to 67 percent or below on another, with large swings for individual carriers on the same prompt.”
The cause, in his view, is straightforward. “An assistant that can’t find structured, authoritative content about a fare fills the gap from aggregator pages, marketing copy, and inference.”
His conclusion is about data, not about the assistants. “The systems don’t have access to the structured, trustworthy data they need to get it right. That’s the kind of gap ATPCO exists to close, and we’re working on how to do that at scale for the industry.”
A choice the industry has to make
Burgess says momentum is building behind general-purpose agentic commerce standards, among them the Agentic AI Foundation under the Linux Foundation, Anthropic’s MCP and Google’s Universal Commerce Protocol and Agent Payments Protocol. None is written specifically for airlines.
“The industry has a choice to make,” he says. “It can let those standards define how agentic shopping works for airlines by default, or contribute airline-specific requirements while they’re still being written. The default isn’t neutral for airlines.”
He cites ATPCO’s research on where assistants send travelers next. “When AI assistants sent travelers onward to shop or book, roughly 75 percent of those referrals went to OTAs and about 25 percent to airlines.” ATPCO’s report describes its referral estimates as indicative. Burgess adds that results shifted with context. “A carrier that was visible in its home market became much less visible when the same route was searched from a transit country or in the opposite direction.”
His argument follows from that. “There’s no single AI channel an airline can optimize for, so the requirements need to sit in the shared standards every platform builds on.”
What ATPCO is building
Burgess says ATPCO is building toward three things.
“First, we’re compiling agent-ready data, which means expressing what airlines already file in a form a machine can read. Second, we’re testing a shared translation layer that maps the way travelers ask for a trip to the way the industry files it. Third, we’re evaluating AI-channel benchmarking, so airlines can see what’s represented in AI channels compared with what they filed.”
He describes a two-week proof of concept to test whether an agent could take a request like “find me a flight with a really comfortable seat” and match it to a real, priced, bookable product using trusted airline data. “It worked, but the more useful finding was what it took to get there: reconciling four ATPCO data sources that each identified the same product differently into one record an AI system could reason over. That’s the work that has to happen before agentic shopping is reliable at scale, and it’s where ATPCO adds the most value.”
He does not present this as a departure. “None of this is new for us. It’s the same trusted-data, standards-setting role we’ve always played, applied to a new kind of consumer.”
Agents are a channel, not a replacement
On whether agentic shopping is a new problem or the same one as NDC distribution, Burgess is precise.
“We see agentic AI as the next channel, not a replacement for the plumbing underneath it. Whether a shopper is typing into a GDS, calling an NDC API, or asking an agent to book a trip, the same questions have to be answered correctly every time: what’s the offer, what’s included, what’s the price for this specific traveler. ATPCO’s role is to stay the neutral, trusted source for that answer regardless of which channel is asking.”
NDC and agent-readability are related, he says, but not identical. “NDC solved how offers and orders move between airlines and sellers in a standard format. Agent-readability is more about how that same data gets exposed and described so a non-human consumer can reliably interpret and act on it, where structure, context, and trust signals matter more than they did for a human agent reading a GDS screen. We’re treating it as a layer on top of NDC and our existing API/subscription infrastructure, not a parallel distribution system built from scratch.”
He also raises load. ATPCO’s report puts look-to-book at roughly 1,000 to 1 a decade ago and about 10,000 to 1 today. “If AI use in trip planning keeps climbing, we estimate the overall ratio could reach 60,000 to 70,000 to 1. Agents don’t stop searching the way people do, so standards need to separate exploratory requests from high-intent shopping before that load reaches live pricing and availability systems.” In the report, that figure is a scenario built on stated assumptions, not a forecast.
Why ATPCO open-sourced NDC Exchange
Asked about the commercial logic of releasing the NDC Exchange code, Burgess says the decision predates him. “We open-sourced the NDC Exchange code some years ago, before my time at ATPCO, but the logic behind it makes a lot of sense to me: ATPCO’s business runs on network effects.”
He explains the reasoning plainly. “Keeping that source code behind a paywall doesn’t help us if it slows down NDC adoption industry-wide, and a smaller, slower NDC ecosystem is worse for the industry (and therefore ATPCO) than a faster, bigger one. Our core business is being the trusted, neutral data source every one of those connections ultimately needs, whichever technology stack carries it.”
So the move was a deliberate bet. “Removing a cost and complexity barrier grows overall NDC adoption faster than keeping that code proprietary would have grown our own position.” As for who can use it: “The code is there for anyone who wants or needs it.”
Africa, the Middle East and Latin America
TDN asked whether ATPCO is building anything to help carriers in these regions leapfrog rather than follow the multi-year path Europe and North America took.
Burgess is direct about what exists and what does not. “We don’t yet have a program explicitly branded or scoped as a leapfrog initiative for emerging markets, and I don’t want to overstate that we do.”
What he points to is the Product Catalog solution ATPCO has started building, “designed to plug into an airline’s existing systems rather than require a ground-up rebuild.” His case is that the groundwork now exists. “The multi-year path European and North American carriers took was multi-year largely because they were building the standards and the infrastructure at the same time as adopting them. That groundwork now exists, which is exactly what should let carriers in Africa, the Middle East, and Latin America move faster.”
ATPCO works with airlines, systems and channels across 148 countries, he says, and its tools “are available to any airline in our network, not gated by region or size.”
TDN’s take
Two things stand out.
First, ATPCO is publishing findings that raise uncomfortable questions about airline visibility in AI-mediated shopping, while arguing that the industry needs to shape the standards governing this emerging channel. That is a more useful contribution than another broad claim that AI will transform retailing.
Second, for carriers in emerging markets, the implication is practical. If assistants fill gaps from aggregator pages and inference whenever structured airline content is missing, airlines with the thinnest machine-readable data are the most exposed. That is TDN’s reading, not a finding of ATPCO’s research, which covered 12 anonymized airlines and did not break results out by region.
The standards debate is already underway. For airlines that want their requirements reflected in the next generation of AI-mediated retailing, the question is whether they will help shape that process or adapt to frameworks developed without them.
