One of my sites gets 43% of its traffic from AI. Another gets 2%. The difference is not technical.
On a network of 60 travel sites, one gets 43% of its traffic from AI and another 2%. The difference is whether your content answers a fact or supports a decision. The AI keeps the fact and sends nobody. The decision gets referred to you, because the user needs to check and buy.
- On my network of more than 60 travel sites, the best case gets 43.4% of its sessions from AI assistants.
- The worst gets 2.0%. Same engine, same stack, same person behind it.
- The winners answer decision questions: which one do I pick, is it worth it, how do I get there, what does it really cost.
- The losers explain what a place is: the AI answers that on its own and keeps the visit.
- It is not a matter of schema or robots. It is what you write about.
The numbers
I run a network of travel sites with the same generator, the same hosting and the same technical criteria. When I measured where the traffic came from, the gap between sites was so wide that at first I thought it was a measurement error. It was not.
| Site | What it answers | Traffic from AI |
|---|---|---|
| Site A | A day trip to an island | 43.4% |
| Site B | Excursions to a waterfall | 41.7% |
| Site C | One specific cave | 2.0% |
How it is measured: first-party analytics on each site, classifying by referrer the sessions that arrive from AI assistant domains. These are not impressions or mentions: they are people who landed on the site coming from a generated answer. AI bots are measured separately, at the edge, because they do not run JavaScript and counting them here would inflate the figure.
Why it happens
The three sites are equally well built. What changes is the question each one answers.
Site C is a cave. Whoever asks about it wants to know what it is, where it is and how much it costs to get in. Those are closed facts. The AI gives them in the answer and the user's journey ends there: there is no reason to click. My page is perfect and irrelevant.
Site A is a decision. There are twenty operators, prices that double from one to the next, departures from three different cities and a real difference between going by catamaran or by speedboat. There the AI cannot close the answer: it has to lean on something, cite where the prices come from and send the user to check. That click is mine.
Put another way: the AI competes with you on facts and collaborates with you on decisions. If your content is an encyclopedia, you have built free training material. If it is a criterion, you have built a source it needs to cite.
What the rest of the industry says
My network is a small sample, so I checked it against the large published studies. Three things worth knowing before deciding where to invest:
ChatGPT cites far more than the rest. Semrush's visibility index (January to April 2026, 126 million prompts) measures an average of 15 sources per answer in ChatGPT versus 3 in Gemini. If your bet is getting cited, ChatGPT is where there is the most room at the table.
Each engine leans on different surfaces. According to Profound, over more than 680 million citations, the most cited domain in ChatGPT is Wikipedia (7.8%), while in Perplexity the king is Reddit. Tinuiti puts numbers on it: Reddit is more than 5% of ChatGPT citations, close to 24% in Perplexity and 0.1% in Gemini. There is no such thing as optimising "for AI" in the abstract: each engine rewards different things.
And it is volatile. That same Semrush study recorded Reddit falling from 60% to 10% of ChatGPT answers in six weeks. Anyone selling you a strategy based on a single surface is selling you a still from a video.
The upside is that the traffic that does arrive pays off. Seer Interactive measured conversion rates of 15.9% from ChatGPT versus 1.76% from Google organic: around nine times better. Fewer people arrive, but they arrive with their mind made up, which is exactly what I see on my own sites.
Be careful when comparing other people's figures: "Reddit is 5% of ChatGPT citations" and "Reddit appears in 60% of ChatGPT answers" look contradictory and are not, because they measure different things (share of total citations versus frequency of appearance). A good part of the noise written about GEO comes from mixing the two.
What to do with this
Before writing the next page, ask yourself whether an assistant can answer it entirely without sending anyone anywhere. If the answer is yes, that page will not bring you traffic no matter how much schema you add.
The angles that do work are the ones that force a choice: comparisons between real options, origin (how do I get there from my city, the angle everyone forgets), real price with the catch included, logistics and the classic is it worth it or not answered for real, with a criterion, not with an "it depends".
And something I learned by tripping over it: a metric at zero is almost never a business conclusion, it is a reason to suspect the measurement. Before deciding that AI sends you no traffic, check that you are measuring it. One of my AI traffic lines read zero for weeks because of a column that did not exist in the database.
Questions
What share of a travel site's traffic can come from AI assistants?
On an own network of more than 60 travel sites, the best case gets 43.4% of its sessions from AI assistants and the worst 2.0%. Most sites sit much closer to 2% than to 43%.
Why do some travel sites get traffic from ChatGPT and others do not?
Because of the kind of content. Pages that answer a closed fact (what a place is, where it is, how much it costs) are answered by the AI itself, which keeps the visit. Pages that support a decision (which option to choose, how to get there from your city, the real price with the catch included) get cited and clicked.
How is traffic coming from AI assistants measured?
With first-party analytics on each site, classifying by referrer the sessions that arrive from AI assistant domains. AI bots are measured separately at the edge, because they do not run JavaScript and counting them here would inflate the figure.
Sources
Semrush AI Visibility Index 2026 · Profound, citation patterns · Tinuiti via CMSWire · Semrush, most cited domains. The data from my network comes from first-party analytics, measured in September 2026.
Joan Sanz. Fifteen years in travel and hotel tech, running a network of more than 60 sites where I test these things before recommending them to anyone. More on my track record at /about.
Does your content answer facts or decisions?
If you want to look at it with your numbers on the table, write to me and we will go through it.
Message me on LinkedIn