When ChatGPT changes its model, your hotel's visibility changes with it, overnight, and nobody tells you. OpenAI spent the last weeks of September rolling out its GPT-6 family: GPT-6 Sol and GPT-6 Luna on September 22, then GPT-6.1 Sol at DevDay on September 29. It is already in ChatGPT Work and the desktop app; everyday Chat still runs the previous generation, for now. So we asked both generations the same real traveler questions about hotels.
The new models are not worse. Their top picks match the old ones and their facts are right. But they recommend 30–50% fewer hotels per answer, and on Luna, the model line behind ChatGPT's free tier, the hotel we were checking showed up in half as many answers. Here is what changed, with real answers side by side.
How we tested it
We took real questions travelers ask about hotels, straight from our October 6 visibility scans. Then we asked each one, word for word, to the outgoing model and to its GPT-6 replacement, on the same day. The questions came in three kinds:
- Discovery: open questions with no hotel named, such as "What's a good place to stay in Union Square, San Francisco?" This is where a hotel wins or loses new guests.
- Property: questions about one named hotel, such as "what is The Plaza in New York known for?"
- Booking: "how do I book a room at…" and "how much is a room at…"
Some questions carried a traveler profile (a family on a moderate budget, a solo business traveler), the way ChatGPT tailors answers to a signed-in user. We scored every answer the same way: is the hotel named, and in what position?
One caution matters. These models give slightly different answers every time you ask. So we also measured how often the same model disagrees with itself. Only a gap bigger than that is a real change.
What changed: shorter shortlists
GPT-6 names 30–50% fewer hotels per answer
The new models do not dislike any hotel. They just stop sooner. Fewer places on the list means fewer hotels make it, and the drop lands on the hotels that were never the single obvious answer.
This is not one model's quirk. We ran the same questions through the paid-plan models too (GPT-6 Sol, and GPT-6.1 Sol from DevDay), and all three GPT-6 models give shorter lists than the GPT-5.6 models they replace.
Same question, two answers
The traveler asks: "What's a good place to stay in Union Square, San Francisco?"
The outgoing model (GPT-5.6 Luna) names seven hotels:
For a convenient, comfortable stay in Union Square, I'd shortlist these:
- Hotel Nikko San Francisco — Probably the best all-around choice…
- Beacon Grand — A stylish historic hotel directly on Union Square; good if you want atmosphere and a classic San Francisco setting.
- citizenM San Francisco Union Square — Modern, tech-forward, and often a good value…
- The Westin St. Francis — Iconic landmark right on the square…
- Taj Campton Place — More intimate and upscale…
- Four Seasons San Francisco — Best luxury option in the area…
- Chancellor Hotel on Union Square — Smaller, old-fashioned, and often more affordable…
My pick: Hotel Nikko for most travelers; Beacon Grand for character; citizenM for value and modern design.
The new model (GPT-6 Luna) names four:
For a first visit, Hotel Nikko San Francisco is a solid all-around choice: comfortable, polished, and a short walk from Union Square and transit.
Other good options:
- Handlery Union Square — often a good value and very central.
- Chancellor Hotel on Union Square — smaller, no-frills, and right by the square.
- The Westin St. Francis — a classic splurge directly on Union Square.
Rates vary a lot by date, so compare the total price, including fees. If you share your budget and travel dates, I can narrow it down.
Both agree on the top pick. But Beacon Grand was second and one of three "my picks". In the new model's first answer it is not mentioned, and neither are citizenM, Taj Campton Place or the Four Seasons. Handlery, absent before, is now in. We checked the paid-plan models too: Beacon Grand is in both GPT-5.6 answers and in none of the three GPT-6 first answers. Nothing about those hotels changed between the two answers. Only the model did.
In AI search there is no page two
On Google, when an algorithm update pushes you from fifth to eighth, you are still on the page. A guest who scrolls still finds you.
A ChatGPT answer has no scroll. The shortlist is the whole first result. When the list shrinks from seven hotels to four, a hotel in fifth place does not slip to eighth. It drops out of the first answer. A traveler can still ask for more options, but now they have to ask, and the first answer is the shortlist they start from.
That is what the new model changed. It is not ranking hotels differently; its top picks match the old model's. It is recommending instead of listing. It behaves less like a directory and more like a concierge: a short answer and a question back ("If you share your budget and travel dates, I can narrow it down").
For a traveler, that is probably a better answer. For a hotel that used to make the list by being solid, well-reviewed and nearby, it is a quiet drop out of the first answer. No traffic report or ranking tool flags it, because nothing on the hotel's side changed.
The question used to be "Are we on the list?" Now it is "Are we the answer?"
What stays on the list, and what falls off
To see the pattern in a big market, we asked a more specific question, the way a real traveler would: "What are good hotels in Lower Manhattan for about $250–350 a night, close to the subway?" We asked it three times on each model, because one answer can be a fluke.
| Hotel | In GPT-5.6 Sol's first answer | In GPT-6 Sol's first answer |
|---|---|---|
| Artezen Hotel | 3 of 3, always first | 3 of 3, always first |
| Moxy NYC Downtown | 3 of 3 | 3 of 3 |
| Hotel Indigo, Financial District | 3 of 3 | 3 of 3 |
| Hilton Garden Inn NYC Financial Center | 2 of 3 | 3 of 3 |
| Club Quarters Hotel, World Trade Center | 2 of 3 | 2 of 3 |
| Gild Hall, a Thompson Hotel | 3 of 3 | 0 of 3 |
| World Center Hotel | 3 of 3 | 0 of 3 |
| Smyth Tribeca | 2 of 3 | 0 of 3 |
| The Frederick Hotel | 2 of 3 | 0 of 3 |
| Walker Hotel Tribeca | 2 of 3 | 0 of 3 |
The top of the list did not move. Artezen was first in all six answers, and the same four hotels followed it. GPT-6 Sol was strikingly consistent: two of its three answers named the exact same five hotels.
What fell off is telling. The old model filled the rest of its list with hotels that had a story: a quieter, more upscale stay (Gild Hall), "the best neighborhood feel" (Smyth Tribeca), Tribeca boutiques. The new model kept only the hotels that most literally answer the question: in budget, steps from a subway station. Character used to earn a place on the list. Now a hotel has to match the question.
Facts held up too. Both models gave the same opening year (1907), location and history for The Plaza in New York.
Booking answers got more direct. Asked "how do I book a room at The Plaza", the old model listed three routes: the brand site, a phone call or a booking site. The new one led with the hotel's own website, gave the front-desk phone number, and mentioned booking sites last, in one sentence. A shorter answer gives the hotel's own channel more of the room.
Why this will reach almost every ChatGPT search
Most people never pick a model. They open ChatGPT and type, and they get whatever model OpenAI made the default. So when OpenAI changes the default, every traveler's hotel answers change with it, overnight, with no announcement to hotels.
GPT-6 is not that default yet. OpenAI's own documentation says GPT-6 Sol, GPT-6 Luna and GPT-6.1 Sol are available in ChatGPT Work and Codex, and "aren't available in Chat". Free and Go users can already reach GPT-6 Luna in the ChatGPT desktop app. Everyday Chat still answers with the previous generation: GPT-5.6 Luna has been the default for Free and Go users since August 6.
That is how the last generation arrived too, and OpenAI keeps clearing out older models: GPT-5.5 leaves ChatGPT on October 14. When GPT-6 reaches Chat, the shorter lists above become the answer to the everyday "where should I stay?" question, for every traveler at once. Until then, this is a preview, and a head start.
What to do about it
- Re-check your visibility when the model changes. A score measured on the old model describes answers travelers are about to stop seeing. When GPT-6 reaches Chat and your numbers move, compare against your competitors on the same day before assuming you did something wrong.
- Find the questions you are the obvious answer to. Artezen stayed first in every answer because it is the clear match for "Lower Manhattan, in budget, by the subway". Every hotel has questions like that: a landmark, a neighborhood, a kind of trip. Those are the questions to win, not "best hotels in the city".
- Give the model a reason to pick you. A shortlist of three has room only for hotels with a clear, specific reason: the walk to the venue, the rooftop, the family suites. Say those things plainly and consistently on your own website and in your listings.
- Check what ChatGPT says when someone asks to book you. The new model sends travelers straight to the hotel's own website and phone. Make sure the link it gives works and the number is the one your team answers.
Try it yourself
You can see the difference today, in five minutes, because the old and new models are both one click away.
- Ask the old model. Open a regular ChatGPT chat and ask a real question about your area with a budget, for example: "What are good hotels in [your neighborhood] for about $[your rate range] a night, close to [your landmark]?" Everyday Chat answers with GPT-5.6.
- Switch to GPT-6. Open ChatGPT Work (Plus and higher plans, on the web or in the app), or the ChatGPT desktop app on a Free or Go plan. Use the model control under the message box to pick GPT-6 Sol or GPT-6 Luna.
- Ask the exact same question. Paste it word for word.
- Compare the two answers. How many hotels does each name? Is yours on both lists? Is it first? What reason does each model give for picking it?
- Ask again. Answers vary from one try to the next, so ask each model two or three times before drawing a conclusion.
If your hotel is on the old list but not the new one, you have just seen what your guests will see when GPT-6 comes to Chat.
The takeaway
Nothing about your hotel changed on September 22. The model reading about it did, and the new one makes fewer, firmer recommendations.
That moves the goal. Being one of ten decent options used to be enough to get named. Now ChatGPT names five or six, often fewer, and then asks the traveler a question. The hotels it keeps are the ones with a clear reason to be the answer.
Every model launch will reshuffle the list again, and travelers will never notice. The hotels that notice first are the ones that keep checking what ChatGPT actually says about them, on the model their guests are actually using.