How to measure AI traffic to a hotel website
Most visits that start inside an AI assistant arrive at your website with nothing attached to them, so your analytics files them as direct and the channel disappears. Here is the method that gives every assistant its own column, and what the numbers look like once it is running.
AI sessions measured across the sites we track, last 30 days.
Quick answer
To measure AI traffic to a hotel website you need a first party pixel that classifies every session by referrer and campaign source against a list of AI assistants, because most of these visits arrive without a usable referrer and end up filed as direct traffic. The panel at the top of this page carries the live 30 day split across the websites we measure, which are our own travel and hospitality sites and not hotel clients: ChatGPT at around 90 percent of the channel, then Copilot, Perplexity, Gemini and Claude. Every one of them is a session our own pixel recorded rather than a figure modelled from an external panel.
Why AI visits end up in the direct bucket
An assistant answers in its own interface and the traveller clicks out at the very end, often from a link the assistant rewrote. What arrives at your server is a visit with no referrer, or with one that no default analytics configuration knows how to name. The visit is real, the channel is invisible, and the hotel concludes that nobody comes from AI.
The fix is not clever attribution, it is a list. Every incoming session is checked against 27 known assistants by referrer and by campaign source, the two signals are deduplicated so a single visit cannot be counted twice, and the result is a column that exists rather than a bucket that hides.
The same principle applies to the consent layer. Our pixel runs in anonymous mode without cookies, so it is not gated behind a banner. That is a measurement decision before it is a privacy one: a banner that blocks the script removes most of the data before anyone gets to look at it.
What a real 30 day window looks like
Concentration is the headline. ChatGPT accounts for around 90 percent of the AI sessions in that window and the other four assistants share what is left. Copilot comes second, then Perplexity, Gemini and Claude, and the distance between first and second is more than twenty to one. The counts behind those ratios are the ones in the panel at the top of this page, read live from the same database.
That shape has a practical consequence. A channel where one source carries nine visits out of ten is a channel where a single change on that source moves your whole number, so it has to be watched weekly rather than reviewed quarterly. It is also small enough per property that an estimate would be worse than useless, which is why the count has to be a real session and not a model.
Volume alone does not tell you what to do, though. The same visits need their landing page attached, because that tells you which of your pages the assistants are actually reading.
Traffic and mentions are two separate scoreboards
The measurement that surprised us most is that the assistant that sends the most visitors is the one that names properties least. In our pass of 3 and 4 September 2026, ChatGPT produced a mention on only 10 of the 57 questions we measured, the lowest of the five surfaces, and yet it sent around nine of every ten AI sessions of the 30 day window. That is one pass, and a single reading proves nothing on its own: the honest unit is a weekly series.
Perplexity is the mirror image: mentions on 22 of the 57 questions, the highest of the five, and a sliver of the traffic. Reading either number alone leads you somewhere wrong. Traffic without mentions tells you that you are being found but not why, and mentions without traffic tell you that you are being recommended into a conversation that never becomes a click.
Both live in the same panel for that reason. Brand monitoring covers the mention side, AI traffic tracking covers the visit side, and the pillar guide covers what to do once both are on screen.
FAQ
By the session referrer and the campaign source, checked against a list of 27 known AI assistants and deduplicated so one visit never counts twice. Both signals are recorded by our own pixel on your pages, which is why the AI sessions in the panel at the top of this page are sessions and not an estimate.
Still a minority of visits and heavily concentrated. The panel at the top of this page carries the live 30 day split. Read it for the shape and not as a forecast for your property: behind it are our own travel and hospitality websites, not hotel clients. The shape is that ChatGPT is around 90 percent of the channel and the other four assistants share the rest.
Measured. Every AI session in the panel is a session our own pixel recorded on your website, with its landing page and its source. Nothing is modelled from an external panel, which matters because the whole channel is still small enough that a modelling error would be larger than the channel.
No. The pixel runs in anonymous mode without cookies, so it loads on every visit and does not sit behind a consent wall. A banner that blocks the pixel costs most of the measurement before you ever see a number.
We measure the split, not the cause. What the 30 day window shows is ChatGPT at around 90 percent of the channel, with Copilot, Perplexity, Gemini and Claude sharing the rest. Any explanation beyond that would be a guess, and a guess written into a report gets read later as a measurement.
Yes. The pixel is one script that runs alongside anything else without conflict. It adds the AI channel your current setup files as direct traffic, it does not replace your reporting.
Treat it as a measurement suspicion before a business conclusion. The check is end to end: load a page in a real browser, confirm the session appears in the database, then delete the test row. Our own worst measurement bug returned a clean zero across an entire surface and looked completely plausible.
No, and the September pass shows the gap. ChatGPT produced a mention on only 10 of the 57 measured questions and still sent around nine of every ten AI sessions of the last 30 days, while Perplexity produced mentions on 22 of 57 and sent a sliver. Mentions and visits are two scoreboards and both need reading.