KAYAK is one of the more interesting companies to test in AI travel because it does not start from an empty page. It already has the thing many AI travel startups struggle to build: live prices, large inventory, filters, maps, partner links and a familiar booking handoff.

That should give KAYAK an advantage. If AI travel is partly about turning messy human intent into bookable results, then metasearch should be a natural place for it.

Kayak offers customers two products: Kayak.ai and side panel chat experience called Ask AI. So we tested KAYAK’s AI experience across flight search, hotel search, pet constraints, flexible dates, destination discovery and airport strategy. The goal was not to see whether it could answer a simple prompt. The goal was to see whether it could understand the trip.

The short answer: KAYAK’s AI is useful, but still limited.

It is strongest when the user already knows what they are looking for and needs help searching faster. It is weaker when the user needs judgment: which trade-off is worth it, which hotel actually fits the trip, whether a destination matches the weather brief, or whether a fast flight is really the fastest total journey.

That is the difference between search and advice.

KAYAK’s strongest AI move is not the chatbot text. It is the way the AI sits next to live travel inventory.

When we clicked on an option in KAYAK.ai, it opened more details side by side with the planning/chat context. We could inspect the flight, carrier, timing and price, then move into the normal KAYAK results page for deeper comparison and booking.

That is the right pattern for travel AI. A travel assistant should not hide the evidence.

The clearest example came in the flight test. We asked for Dubai to Munich around July 23, one adult, direct preferred, but one stop acceptable if it saved more than 25%.

KAYAK first showed direct Emirates flights at around $622. But cheaper one-stop options were available. When challenged, the AI found options around $237–255 and correctly calculated that they saved roughly 52–62%.

That is the problem in one test. The data was there. The tool could do the comparison. It just did not do it on the first answer. This matters because the user’s real request was not “show me direct flights.” The request was: “Show me the best trade-off between direct and much cheaper one-stop options.”

That is exactly where AI should add value. Normal search can list results. AI should read the condition, compare the options and say: “Direct is easier, but the one-stop saving is large enough to consider.”

If the assistant only gets there after being corrected, the user still has to supervise the search.

The same pattern appeared in hotels. KAYAK’s AI could produce useful hotel shortlists. In Kuala Lumpur, it surfaced central options with pools and kept the budget under control. In some cases, KAYAK.ai gave a better shortlist than the basic Ask AI side panel: more central, more relevant, less random.

But hotel advice is not just a list. A good hotel recommendation needs to understand why the hotel fits the trip. Is it best for Chinatown? Quiet stay? Pool quality? Business travel? Late checkout? Room size? Walkability?

KAYAK’s AI can work with filters and visible signals. It can say a hotel has a pool, a good rating or a central location. But the advisory layer still feels thin. It sometimes says a hotel avoids a “party” feel without showing much evidence. It can rank hotels, but the explanation does not always feel strong enough to trust.

This is not surprising. Metasearch is built for speed, breadth and price competitiveness. It is not always built on the deepest hotel content or verified property-level policy data.

AI exposes that structural limit.

The pet test was one of the more useful signals. We added a 12 kg dog to the Kuala Lumpur hotel search. That is a hard constraint. If the hotel does not accept the dog, the rest of the recommendation is useless.

KAYAK’s AI handled this better than expected in the stronger KAYAK.ai flow. It rebuilt the search and showed pet-friendly options under budget. It also added a needed caveat: pet-friendly does not always mean a 12 kg dog is accepted. Weight limits, fees, room restrictions and breed policies may still need hotel verification.

That is closer to a real workflow. A traveler adds a constraint late. The original shortlist breaks. The system rebuilds the path. But it also shows the risk. “Pet-friendly” is a weak signal unless the policy is explicit. For the traveler, this is not a data-quality nuance. It is the difference between a workable booking and a problem at check-in.

KAYAK can surface likely options. It cannot always guarantee the operational detail.

That is where travel AI becomes sensitive. The more the interface sounds like an advisor, the more users may expect certainty.

The weakest results came when we moved away from inventory search and into destination choice.

We asked for four nights in late August from Kuala Lumpur. The destination should be cooler than Singapore, direct flight preferred, good food, walkable, not too expensive and simple for a Malaysian passport.

KAYAK suggested places such as Taipei, Incheon/Seoul, Osaka and Tokyo. In earlier testing, similar suggestions included Hanoi and Hong Kong.

That is not a strong answer if “cooler than Singapore” is one of the main constraints. Taipei, Tokyo and Osaka can be hot and humid in late August. Hanoi and Hong Kong are also not obvious cooler-weather escapes. Seoul is more defensible, especially in the evenings, but still not a clean cool-weather answer.

When challenged, KAYAK’s AI corrected itself and admitted that several of the suggested cities were hot. That is useful. But again, the first answer is the product experience that matters.

The likely issue is that the system optimized for what KAYAK knows well: direct flights, price, popularity and city appeal. It underweighted the real-world travel condition: weather comfort.

That is the boundary of metasearch AI. It can find trips. It may still struggle to understand whether the trip is actually a good idea.

Across the tests, one pattern kept repeating: KAYAK’s AI often became much better on the second answer.

It missed the Dubai–Munich one-stop trade-off, then corrected it. It gave weak cooler-weather recommendations, then improved after being challenged. It initially handled airport strategy too much like flight search, then gave a better door-to-door comparison after a follow-up.

That is both encouraging and limiting. It means the system has enough capability to reason better. It can compare, recover, explain and refine. But the burden is still on the user to challenge it.

That is not how a strong travel assistant should work. A strong assistant should catch the important condition before the user points it out.

The strongest use cases were the ones closest to KAYAK’s core business.

Flexible-date flight search worked well. When we asked for a Dubai to London long weekend in September, flexible by three days, with a cheap but sensible itinerary and no terrible overnight layovers, KAYAK’s AI could search live fares and bring back useful options.

Airport strategy also worked reasonably well after refinement. For a New York to Bologna or Florence trip, KAYAK could compare Bologna, Florence, Pisa, Venice and Milan. The first answer was too flight-led, but the follow-up got closer to the right travel logic: Bologna is best for Bologna, Florence is best for Florence, Pisa is a backup, and Milan or Venice only make sense if the flight advantage is large enough to offset the ground transfer.

These are the use cases where KAYAK’s AI feels genuinely useful. The user knows the broad trip. The system helps translate messy preferences into live results. That is a good product direction, but it is not yet an AI travel agent.

KAYAK’s AI shows where travel search is going. The search box is becoming conversational. Users will increasingly describe what they want instead of manually clicking filters. But the hard part is not the conversation, it is judgment.

Can the assistant understand that a 60% saving should be highlighted even if the user prefers direct flights? Can it tell when a “cooler” destination is not actually cooler? Can it know when a pet-friendly hotel still needs verification? Can it compare total travel time rather than just flight duration? Can it explain why one hotel fits the trip better than another? Those are the questions that separate AI search from AI travel advice.

KAYAK has a smart foundation because it does not hide the underlying results. The user can still inspect inventory, compare prices and click through to the classic KAYAK flow. That makes the product safer and more useful than a standalone chatbot.

But it also reveals the current limit. KAYAK’s AI can make metasearch easier to use. It can help users start faster, refine faster and inspect options in a cleaner way. However when the task moves from finding options to deciding what makes sense, the product still needs supervision.

The verdict after testing: KAYAK’s AI is a better search interface, not yet a strong travel advisor. It is probably for clients who are really excited about the conversational search, but it does not give user a major benefit compare to regular search with filters and sorts.

That may be enough for now, but if AI travel is moving toward agents that understand context, preferences and trade-offs, then metasearch has a harder question to answer. Can it move beyond what is it today given limited content access, traditional for meta searches?

Or, as we have seen before, when metas were trying to experiment and expand into other segments - they should keep doing what they do best: compare prices in a simple interface that any user can understand.