75% of What AI Mode Cites in Travel Is Guide Content.

Across 100 unbranded travel searches, Google AI Mode built its answers by citing 2,558 separate sources drawn from 1,229 different domains. That averages about 26 citations per answer, and one answer pulled in 53. Three quarters of everything it cited, 75.3% to be exact, was a guide, a tips piece or an itinerary, the exact content format that most teams are quietly cutting from the plan. The clicks to that content have fallen sharply as AI answers absorb informational queries. The content itself is doing more work than ever. Those two facts sit next to each other in the data, and most travel teams are only looking at the first one.

We ran this study in-house to settle an argument we keep having with clients: if the traffic is gone, is the content still worth making. The short version is yes, and the numbers below show why. The full method, the caveats and how to get the raw data are all here.

How we ran the study

On July 9 2026 we sent 100 unbranded travel prompts to Google AI Mode through the SerpAPI Google AI Mode endpoint, which returns the answer text and the full ordered citation list for each query. The set leaned toward destination guides, 52 of them, with 24 itineraries and 24 travel-hack and tips queries, phrased the way people actually search rather than the way marketers write, so "best things to do in Tokyo" and "how to find cheap flights" instead of anything keyword-shaped. We set no location, so the answers reflect Google's default, US-leaning inference. For every prompt we saved the full answer text and logged each cited source in order, then classified the domains into source types using a documented rule set that a human on our team reviewed.

99 of the 100 prompts returned an AI Mode answer. The one that did not, "things to do in Paris for first time visitors," came back empty on two separate attempts, which is worth remembering the next time someone tells you AI answers are inevitable for every query. The 99 that did answer carried those 2,558 citations, and the answers themselves ran long, 1,854 words on average. AI Mode is not writing a snippet. It is assembling an article, and it needs a lot of raw material to do it. For scale, Google said at I/O 2026 that AI Mode passed one billion users, so the way it sources answers is not a fringe question.

AI Mode harvests, it does not rank

The old mental model is ten blue links and a fight for the top three. Throw it out for this surface. With 26 sources feeding the average answer and 1,229 domains earning at least one citation across a single snapshot, AI Mode behaves less like a ranking engine and more like a harvester. It takes a fact here, a recommendation there, a price from a third page, and stitches them into one response.

No single site owns it. YouTube was cited most often, appearing in 76 of the 100 answers, and Reddit followed at 70. Behind those two sits a long, fragmented tail where the biggest travel publisher in the set, Condé Nast Traveler, showed up in 15 answers and Lonely Planet in 11. The concentration everyone fears at the top of classic search is not there. A single well-matched page from a site nobody has heard of can land in the same answer as YouTube, which cuts both ways for every brand in the category.

The content everyone is cutting is the content it cites

Here is the finding that should change a content budget. Independent travel blogs and content sites were the single largest source type in the entire dataset, 56.8% of all citations. Not OTAs, not the big publishers, not tourism boards. Ordinary travel content, the guide-and-itinerary kind, more than half of everything Google's AI quoted.

Sit with that last row for a second. The official tourism boards and government sites, the ones with the real authority and the domain strength, earned 1.2% of citations. The big brand-name publishers earned 4.2%. Independent travel content earned more than ten times as much as either. In this dataset Google's AI rewarded whichever page answered the question best, and domain size did not predict who got cited. That should worry the brands who assumed authority was a moat, and it should encourage everyone smaller than them. We are describing a pattern in one snapshot, not a ranking law, so read it as a strong signal rather than a guarantee.

Clicks left, citations stayed, and those are not the same loss

The reason this matters to anyone watching a Search Console chart is the gap between two things that look identical on a traffic report. A guide that pulled 5,000 visits a year ago may pull a fraction of that today, and the obvious conclusion is that the content stopped working. Often the opposite is true. That guide is now one of the two dozen or so sources AI Mode reads to answer the query, and the searcher gets satisfied on the results page instead of clicking through. The demand did not disappear. The visit did. This is the zero-click shift we covered in detail when 68% of US Google searches ended without a click, and a traffic report cannot tell you which clicks were lost demand and which were simply absorbed into an answer.

We watch this exact divergence in client dashboards, impressions climbing while clicks flatten, and it looks like decline until you separate the two. The value moved from the visit to the mention. The goal moves with it, from ranking a link to being one of the sources the model quotes.

Guides, itineraries and tips all still get pulled

None of the three formats is being left out, which is the reassuring part. Itineraries pulled the most sources per answer at 29.4 and produced the longest responses, guides sat in the middle, and tips content was slightly leaner at 21.8 citations but still deep. Guide-shaped content made up roughly three quarters of the citations in every category.

The one real difference is where each format sources from. Destination guides and itineraries leaned hardest on independent travel blogs and tour platforms like GetYourGuide, Klook and kimkim. Tips and hacks queries leaned more on first-hand experience, so Reddit, Quora, NerdWallet and The Points Guy surfaced far more often on "how to find cheap flights" than on "best things to do in Tokyo." When the question rewards lived experience, the AI goes looking for people who have lived it, which is a useful thing to know before you brief another generic listicle.

What the cited pages have in common

The pages that get lifted share a shape, and it is not the shape of a beautifully written 3,000-word essay. AI Mode is harvesting parts, so it favors content built from parts: clear headings, specific named recommendations, real prices and times, day-by-day or step-by-step structure, and claims that stand alone one sentence at a time. Long unstructured prose can be completely correct and still lose the slot to a thinner page that is shaped like an answer, because the model is not reading for prose quality, it is reading for liftable facts.

That is the practical brief. Name the actual restaurant, quote the actual entrance fee, give the actual walking time between two stops, and let each fact live in its own line rather than buried three clauses deep. The content that reads as a little too structured for a human is exactly the content the machine finds easiest to quote.

What this study can and cannot tell you

Two limits belong next to every number here. Our source classification is a heuristic based on domains, so a small share of sites will sit in an adjacent bucket, a tour operator counted as a booking platform for instance, and we reviewed the edge cases by hand rather than trusting the rules blindly. And this is one snapshot from one day with no location set, so the answers skew to Google's default US inference. AI Mode is non-deterministic and personalized, which means the exact domains in any single answer will shift by user and by week. The stable findings are the patterns, citation depth, the dominance of independent content and the format that gets lifted. The specific brand in slot four is the part that moves, so treat the category-level shares as the reliable output and re-run for your own market before acting on the domain names.

The full dataset behind this post, all 100 prompts with every answer and citation, is available on request. We would rather you interrogate the numbers than take them on trust.

Keep publishing guides, itineraries and tips. They are the most-cited content format in AI Mode travel answers in this study, and cutting them removes the surface area you have to be cited on. Then shift the work from ranking to being quotable, structure pages so a model can lift a clean answer, put real numbers and named specifics where generic copy used to sit, and stop measuring that content on sessions alone. The scorecard that fits this surface is citation share, how often your domain shows up as a source for the queries you care about, which is the same thing this study measures at the category level.

About this study: Search Agency is a search and generative-engine-optimization consultancy that measures and improves how brands appear in AI answers across Google AI Mode, AI Overviews and the major AI assistants. We ran and analyzed this study in-house using the SerpAPI Google AI Mode endpoint for data collection and our own reviewed classification for source types. Questions or a request for the raw data can go through our contact page.

(Raw data available here)