Seba BlancoI help independent hotels sell more effectively and operate smarter by combining hotel technology with sales and marketing.
Table of Contents
Do not index
Do not index
A couple in Buenos Aires is planning a January road trip south. One evening, one of them opens ChatGPT and types: somewhere quiet by a lake, with cabins, good for kids, close to the road we are driving. ChatGPT takes a few seconds. It opens a handful of websites, reads them, and writes back three suggestions with a line about each.
Maybe my lodge is one of the three. Maybe ChatGPT opened my homepage to write that line. The couple reads the answer, picks a place and goes to bed.
My monitor tells me, with reasonable confidence, what ChatGPT would answer if a traveller asked that question. What it can't tell me is whether any traveller actually asked. This post is about how to find out.
What the monitor already tells me
The lodge I run in Patagonia, with rooms, cabins and a campsite, gets a weekly check on how AI assistants describe it. A scheduled routine puts the same set of traveller questions to ChatGPT, Claude, Gemini and Perplexity. I did not invent those questions. They come from the searches that already bring people to the site, taken from Search Console and reworded the way a guest would ask an assistant. Each reply is then compared with a fact sheet in Notion that holds the hotel's correct details.
Three readings come out of that.
Discovery: does the assistant name the hotel at all?
Accuracy: is what it says true?
Sources: which pages was the answer built from?
If you want the full build, I walked through it in the first post on AI visibility for hotels.
The two questions it cannot answer
The monitor only sees the answers to questions I ask. That leaves two blind spots.
First, when a real traveller asks an AI about places like mine, does the AI open my website to answer? Second, how many of those travellers come to the site afterwards, and how many book?
So I added two layers. The simplest way to picture the system is three windows onto the same thing. The monitor asks the assistants what they say. A counter on the website sees whether the assistants come to read. GA4 sees whether people arrive.
Three kinds of AI robots
Not every AI visit means the same thing. The companies behind the assistants send different robots (crawlers, in the jargon) for different jobs, and they document them. OpenAI, Anthropic and Perplexity each publish their list.
Type | What it does | Names to look for |
Reads because a person asked | Opens your page at that moment to answer someone's question | ChatGPT-User, Claude-User, Perplexity-User |
Indexes for the AI's search | Builds the index the assistant searches later | OAI-SearchBot, Claude-SearchBot, PerplexityBot |
Collects data for training | Gathers pages that may be used to train future models | GPTBot, ClaudeBot |
Only the first row is a direct sign of a real question. When ChatGPT-User opens your page, someone just asked ChatGPT something and the answer needed your site. The other two rows still matter, because they decide whether you can be found at all. But they run on the AI company's schedule, not on a traveller's.
Where to see it
If your website runs through Cloudflare, look for AI Crawl Control in the dashboard. It used to be called AI Audit, and Cloudflare lists it as available on all plans, the free one included. The Crawlers tab shows each AI robot and how many requests it made. The Metrics tab shows which pages they asked for most. The free plan has two limits, according to Cloudflare's own setup guide: the analytics only go back 24 hours, and robots are identified by the name they announce, without a deeper check.
If your site does not run through Cloudflare, you probably cannot see this yourself, but your web or hosting provider can. Send them three questions:
- How many visits did ChatGPT-User, Claude-User and Perplexity-User make last month?
- Which pages did they read?
- Did any of them get errors or blocks?
The third question matters more than it looks. A firewall rule or a security setting can turn these robots away without anyone at the hotel ever deciding it.
For people, use GA4. Open the traffic acquisition report, look at session source, and filter for chatgpt.com, perplexity, gemini, claude and copilot. That is your count of humans who arrived from an assistant.
What I saw in one week
The lodge's site already ran through Cloudflare. I set up a counter that saves the AI robot visits every day, because the free plan keeps only a short history, and a weekly report that compares each week with the one before and with GA4.
Here is the first week, 25 September to 1 October 2026, counting verified visits only.
Type | Robot | Visits |
Reads because a person asked | ChatGPT-User | 179 |
Claude-User | 62 | |
Perplexity-User | 0 | |
Indexes for the AI's search | OAI-SearchBot (ChatGPT Search) | 350 |
PerplexityBot | 76 | |
Collects data for training | ClaudeBot | 489 |
GPTBot | 167 |
The reads triggered by questions went mostly to the homepage and to blog articles.
On the human side, GA4 counted 64 sessions from AI assistants between 1 July and 30 September, out of 2,759 in total. That is 2.3%. ChatGPT sent 58, Gemini 4 and Perplexity 2. They landed on the homepage (28), the campsite page (11) and the cabins page (5). In one recent week of September there were 18, of which 14 came from ChatGPT. Bookings attributed to those sessions over the quarter: zero.
Reads are not visits, and visits are not bookings
Put those numbers side by side and one thing stands out. In a single week, ChatGPT opened the site 179 times to answer someone. In the whole quarter, 58 people came from ChatGPT.
The assistant reads on the traveller's behalf and answers inside the chat. Very often the traveller never clicks through. The couple from Buenos Aires got three suggestions and a line about each. They had no reason to open three websites.
That has a practical consequence. For that traveller, what the AI says about your hotel is your website. It is the only version they see. This is why the accuracy part of the monitor matters so much. If the AI reads your page and repeats a wrong fact, the traveller never reaches the page that could have corrected it.
It is tempting to divide one number by the other and call it a conversion rate. Don't. The time windows are different. One question can trigger several reads. And part of the traffic that comes from assistants arrives without a referrer, so GA4 files it as direct. The ratio would look precise and mean nothing.
The zero bookings deserve the same care. They do not prove the channel is worthless. They prove you cannot measure it by looking only at the last session before a booking. Someone can read about the lodge in ChatGPT on a Tuesday, search its name on Google on Thursday and book by WhatsApp a week later. None of that is credited to ChatGPT.
Where the monitor and the counter meet
This is the part I find most useful. Each layer is limited on its own. Together they cover each other's blind spots.
Cited sources versus pages actually read. The monitor records which URLs the assistants cite. The counter records which pages they really open. The most-read pages were the homepage and the blog posts, so blog posts are sources too. An old fact in an old post travels. My rule now: when the monitor flags a wrong fact, I check the pages the AI reads most before anything else.
The monitor leaves footprints. It runs Monday to Thursday, one model a day, always at the same hour. Some of the reads at that hour may be the monitor itself. The counter marks that hour and I discount it.
Noise, with the same caution in both layers. The monitor only treats a change as real when a percentage moves by 15 points or more. For the counter, my rule is simpler: under 10 visits, I read nothing into a rise or a fall. Perplexity-User at zero is just zero, for now.
How to read it without fooling yourself
It is a floor, not a total. Many answers come straight from the assistant's index without reading the page again. The number of questions that touched your hotel is higher than the reads you see.
Count only verified robots. Some programs borrow these names, so if your tool cannot verify them, treat the count as approximate.
Discount your own tests. Every time you ask ChatGPT about your hotel to see what it says, you generate reads. So does the monitor.
Write the number down every week. The free history is short, and Cloudflare's free panel shows a single day. Note it on the same day and at the same hour each week. A one-day sample taken consistently shows a trend that a single look never will.
What it is not
The counter started on 1 October 2026, with seven days of history behind it. The first full weekly report comes out on 6 October, and the first monthly report from the monitor is also due this month. So far I can show that the assistants read the site. I cannot show that this turns into bookings, and I will not pretend otherwise.
Costs are modest. The monitor comes to about nine US dollars a month in API calls. To start with the counter, Cloudflare's free panel is enough. Mine runs on the Cloudflare setup the hotel was already using.
The point of all this
Travellers are handing more of their research to assistants, and the assistants do the reading for them. Most of those conversations will never show up in your analytics as a visit. You cannot watch them happen. You can see the trace they leave: the robot that came to read your page because someone asked.
Put the three layers together and you get a reasonably honest picture. What the assistants say about you, whether they come to read, and whether anyone follows them back. None of it proves bookings yet. All of it tells you where to look when something is wrong.
If you want the monitor and the counter running for your hotel, built on your own Search Console data, your fact sheet, your website and your GA4, Hotel Marketing Point sets them up for independent hotels.
Written by
Seba Blanco
I help independent hotels sell more effectively and operate smarter by combining hotel technology with sales and marketing.
