Airlines are queuing up to get into ChatGPT. That is the easy part. The hard part is that most airline product data was never written for a machine to understand, and nobody wants to talk about it.
This week gave us both halves of the story. easyJet and Southwest launched ChatGPT plugins. And airline data company ATPCO quietly showed how far the industry still has to go before an AI can answer a simple question properly.
The plugin rush
On 5 October, easyJet launched a ChatGPT plugin for flights, package holidays and live flight status. Travellers search in a chat, see live prices and availability, then finish the booking on easyJet.com. easyJet says organic traffic from ChatGPT is up 400% year on year for flights and 125% for holidays. As Airways points out, it gave no absolute visitor numbers or conversion figures, so treat those percentages with care.
A day later, Southwest followed with its own plugin, which we touched on earlier this week. According to PhocusWire, they join Virgin Atlantic and Turkish Airlines, while Skyscanner, eDreams ODIGEO and Ixigo launched ChatGPT apps earlier in 2026.
So the shop window is filling up fast. My view: within a year, an airline plugin in ChatGPT will be about as special as an airline app. Everyone will have one. The edge will come from what sits underneath it.
“A comfortable seat” is a hard question
The most important travel AI story this week is not a launch. As PhocusWire reported on 9 October, ATPCO tested how to match what a traveller asks for in everyday language with the structured data airlines actually hold. Anand Mishra, ATPCO’s VP of technology, put it simply: AI assistants are already good at understanding what a traveller wants. The problem is linking that intent to airline data.
His example is a traveller asking for a comfortable seat. A human gets it straight away. But the data has fares, cabins, brands, seat maps and aircraft types, stored in separate systems. ATPCO used Gemini to map both the request and its own product descriptions into the same “space”, then pulled back the closest matches. The hardest part, Mishra said, was stitching separate systems into one record per product.
“The human intent doesn’t know that this is how we’re structuring the data,” he said. “They just have a semantic goal.”
The bit that should worry you
ATPCO tested 12 scenarios, including some where no good match existed. In some of those, the system still returned a record anyway. In Mishra’s words: “Even if that other point is really, really far away, it’s going to pick it.”
Ask for something an airline does not sell, and the system can still hand back something. Confidently.
ATPCO is open about this. It is looking at showing confidence levels, and Mishra said he expects the issue can be fixed. Fair enough, it was a proof of concept. But it shows where the real risk sits: not in AI misunderstanding the traveller, but in the data letting it down.
Southwest seems to know this too. Its website says ChatGPT responses “may contain errors or omissions” and that Southwest “is not responsible for the accuracy or completeness of information provided by ChatGPT”, as TravelPulse reported. Sensible legal cover. Also an admission that, once your fares are in the chat, you do not control the sales pitch.
Your product descriptions are now data
Here is the part most travel marketers will miss. In ATPCO’s test, the text that got matched to the traveller’s request was the description of what cabins and seat sizes mean, plus images and video. In other words, the stuff marketing teams usually own.
For years we wrote cabin and room descriptions like “relax in style” for people skimming a web page. Now those words are being turned into numbers and compared with what a traveller asked for. Vague copy does not just read badly. It matches badly.
Mishra also said ATPCO’s own data rules are currently written for human experts, and need to become machine-readable. If the industry’s data standards body has that job ahead of it, most airline, hotel and tour operator websites certainly do. I made a similar point in Hilton’s AI agent problem is a gift to the OTAs: if the machine cannot read your product clearly, it will sell someone else’s.
What I would do now
- Audit your product language. Take your cabin, seat, room or tour descriptions. Would a machine know what “premium”, “extra space” or “family-friendly” actually means? Add the facts: seat pitch, bed size, what is included, what is not.
- Test the awkward questions. Ask ChatGPT and Gemini for things you do not offer. If they still recommend you, find out which page or data source is feeding them.
- Do not judge success on traffic alone. easyJet’s 400% is a growth rate, not a sales figure. Track what AI visitors book, as I set out in yesterday’s piece on AI referral traffic.
- Give marketing and data teams one brief. The people who write product copy and the people who manage product data rarely sit together. They need to.
The debate
The industry is celebrating each new ChatGPT plugin as if it were the finish line. I think it is the starting gun. A plugin gets you into the conversation. Clean, specific, machine-readable product data is what gets you recommended for the right reasons.
So which matters more for your brand right now: being in ChatGPT, or being described correctly once you are there?
Follow the newsletter: Dan Taylor on LinkedIn