If you cannot measure your AI visibility, you cannot improve it — and right now most independent hotels are flying completely blind. You can pull a Google Search Console report in ten seconds, but ask “how often does ChatGPT recommend my hotel when someone plans a trip to my city?” and the honest answer is usually a shrug.
This is a guide to fixing that shrug. Not with a magic dashboard, and not with a vendor pitch. With a repeatable monthly process you can run yourself in under an hour: a prompt panel that measures your share-of-mention across ChatGPT, Gemini, and Perplexity, tracked over time so you can actually see whether your work is moving the number.
I will show you exactly how to build one, what to record, how to avoid the traps that make the data lie to you, and how to read the results without fooling yourself.
Why a panel, and not a single check
Here is the thing everyone gets wrong first: they open ChatGPT, type “best boutique hotel in [their city],” see their name (or not), and draw a conclusion. That is not measurement. That is a coin flip you watched once.
Large language models sample their outputs. Ask the identical question twice and you will often get two different lists. Ask it on a Tuesday versus a Saturday, logged in versus logged out, with search browsing on versus off — different answers again. A single query tells you almost nothing about your real standing.
A panel solves this by trading depth for breadth. Instead of asking one question and over-reading the answer, you ask 20 to 40 realistic questions and record how often you show up across all of them. The individual answers stay noisy. The aggregate — your share-of-mention — gets stable and honest. That stability is what lets you compare July to August and trust the difference.
If the idea that AI assistants are now a real discovery channel still feels abstract, start with Is your hotel invisible to ChatGPT? — this guide assumes you already believe the channel matters and want to put a number on it.
The goal is not to “rank #1 in ChatGPT.” There is no such thing, and anyone selling you that is selling AI-slop. The goal is a trustworthy trend line: are you mentioned more this quarter than last, for the prompts that actually send you guests?
Download this study as a one-page PDF
Step 1: Write prompts a real traveler would type
Your panel is only as good as its prompts. The mistake is writing prompts you wish people asked (“best luxury independent hotel with rooftop bar and sustainability certification in [your city]”). Real travelers are lazier, vaguer, and more intent-driven than that.
Build your prompt list around the jobs a guest is trying to do. A good panel spreads across five buckets:
- Discovery / non-branded. “Where should I stay in [your city] for a first visit?” / “Good boutique hotels in [your neighborhood]?” This is the big one — these are the prompts where you are competing to be discovered at all.
- Attribute-led. “Family-friendly hotels near [local landmark] with a pool.” / “Quiet hotels in [your city] good for remote work.” Map these to your genuine strengths. If you run an aparthotel, lean into the queries covered in aparthotel and extended-stay marketing.
- Comparison. “Is [Competitor A] or [Competitor B] better for a couples weekend?” Include prompts that name your rivals but not you — those reveal where you are being left out of the conversation entirely.
- Trip-planning / itinerary. “Plan a 3-day trip to [your city], including where to stay.” These long, multi-part prompts are increasingly how people actually use Gemini and ChatGPT, and being the hotel that gets slotted into the itinerary is gold.
- Branded. “Tell me about [Your Hotel Name].” / “Is [Your Hotel Name] worth it?” These test whether the models know you accurately — wrong pet policy, wrong location, hallucinated amenities. Cheap to check, embarrassing to ignore.
Aim for roughly 25 to 35 prompts total, weighted toward discovery and attribute buckets because that is where new demand lives. Write them once, in a spreadsheet, and then freeze them. Changing prompts month to month destroys your ability to compare. The panel only works if the questions stay constant.
Step 2: Decide what “a mention” means before you start
You need scoring rules written down before you run anything, or you will unconsciously grade generously on the months you want to look good. Define these fields for every prompt-and-platform combination:
| Field | What you record | Why it matters |
|---|---|---|
| Mentioned? | Yes / No | Your core metric. Were you named at all? |
| Position | 1st, top-3, later, buried | Being first named is worth more than being tenth. |
| Sentiment | Positive / neutral / negative | ”Overpriced but central” is a mention you may not want. |
| Accuracy | Correct / minor error / wrong | Wrong location or amenities are a fixable content problem. |
| Cited source | URL the model linked (Perplexity especially) | Tells you which pages are feeding the answer. |
| Competitors named | List them | Your share-of-mention is relative to who else shows up. |
The citation column is the one people skip and the one that pays off most. Perplexity and ChatGPT’s search mode often show their sources. When you see the same OTA destination page or the same review aggregator cited again and again, you are looking at the exact surfaces you need to influence. That is not a coincidence — it is the mechanism behind how OTAs quietly steal your search traffic, now playing out inside AI answers.
Step 3: Run the panel the same way every time
Consistency of method matters as much as consistency of prompts. Small setup differences change answers, so lock down a protocol and follow it every month:
- Use fresh, logged-out or temporary sessions. Your own chat history personalizes results and flatters you. Use a signed-out window, a temporary chat, or a clean profile so you see something closer to what a stranger sees.
- Test across all three platforms. ChatGPT, Gemini, and Perplexity behave differently — Perplexity leans hard on live citations, Gemini pulls Google’s index and Maps, ChatGPT varies by whether browsing fires. Record each separately. Do not average them into one blurry number.
- Run each prompt two or three times and record the majority behavior. Because of sampling, one run can mislead. If you show up in two of three runs, that is a “yes” with a note. This is your defense against reading too much into a single lucky or unlucky roll.
- Set location deliberately. If you serve a local “hotels near me” intent, the searcher’s location changes everything. Note whether you tested from your city or neutrally.
- Timestamp everything. Same date each month (say, the 1st), same rough time of day. Write the date in the sheet.
Budget 30 to 45 minutes. Yes, it is manual. That is a feature the first few months — doing it by hand is how you learn what the models actually say about your market, which no dashboard will teach you.
Step 4: Turn the runs into a share-of-mention number
Now compute the metric that goes on your trend line. The simplest honest version:
Share-of-mention = (prompts where you were mentioned) ÷ (total prompts) × 100, calculated per platform.
So if you appear in 9 of 30 discovery-and-attribute prompts on Perplexity, that is 30 percent. Track that number monthly. A few refinements once the basic version is running:
- Weight by position. Count a first-named mention as 1.0, a top-three mention as 0.6, a buried mention as 0.3. This “weighted share-of-mention” reflects that being listed tenth barely gets clicked.
- Segment by bucket. Your branded share should be near 100 percent — if the models cannot describe your own hotel correctly, that is a content and structured-data emergency. Your discovery share will be much lower; that is the number to grow.
- Track competitor share alongside yours. If a rival appears in 24 of 30 prompts and you appear in 9, you now know the gap in concrete terms, and you can watch it close.
One number to watch above all: discovery share-of-mention on non-branded prompts. Branded queries just confirm the models know you exist. Non-branded discovery is where genuinely new guests find you instead of your competitor. That is the metric that maps to revenue.
Step 5: Read the results without lying to yourself
A trend line invites over-interpretation. Guardrails:
One month is a data point, not a trend. Expect noise of several points month to month from model updates alone. Do not celebrate or panic over a single move. Look for direction across three or more months.
Correlate with what you changed. The panel is worthless if you cannot connect it to actions. Keep a simple change log next to it: “May — published neighborhood guide,” “June — 40 new Google reviews,” “July — fixed schema on room pages.” When share-of-mention moves, you want a candidate cause. This is measurement in service of experimentation, which is the whole point of this cluster.
Watch citations as a leading indicator. Often your cited sources shift before your share-of-mention does. When the models start pulling your own pages instead of an OTA’s, movement in the mention rate usually follows. That is your earliest sign the content work is landing.
Separate accuracy problems from visibility problems. If you are mentioned but described wrong, that is a fixable data issue — often solved by tightening your Google Business Profile and your on-page structured content so machines read the right facts. If you are not mentioned at all, that is a harder visibility problem that needs an AEO and GEO strategy, not a quick edit.
What to do when the number is bad
It probably will be, at first. That is fine — you now have a baseline instead of a shrug. The levers that tend to move AI share-of-mention for independent hotels:
- Publish the content the models want to cite. Neighborhood guides, honest comparison pages, and specific answers to real traveler questions give the models something to pull from. Our mega-guide to OTA destination landers that rank covers the page types that double as AI training fodder.
- Fix the fact layer. Accurate, structured, consistent information across your site, your GBP, and major directories is what lets a model describe you correctly and confidently. Inconsistent facts get you dropped or mangled.
- Feed the reputation signal. Volume and recency of reviews influence both what gets said and how positively. This is ongoing content and reputation work, not a one-time push.
- Understand the data supply chain. A surprising amount of what AI assistants “know” about hotels traces back to a handful of aggregated data sources — the dynamic we unpacked in the FlipTo and Spacetime sleeping-giant piece. Knowing where the models get their facts tells you where to intervene.
And keep the business context straight while you do it. Every guest an AI assistant sends to your own site instead of an OTA is a guest you did not pay commission on — the book-direct math is exactly why this measurement work is worth an hour a month.
The 45-minute monthly ritual, start to finish
To make this concrete, here is the whole loop:
- Open your frozen prompt sheet (25 to 35 prompts, five buckets).
- In clean, logged-out sessions, run each prompt 2 to 3 times across ChatGPT, Gemini, and Perplexity.
- Record mention, position, sentiment, accuracy, cited sources, and competitors named.
- Compute weighted share-of-mention per platform and per bucket.
- Log what you changed since last month next to the numbers.
- Compare to prior months; act on the discovery number, not the branded one.
Do that on the first of every month and in a single quarter you will know more about your AI visibility than 95 percent of hotels in your market — and you will have the trend line to prove whether your marketing is actually working.
If you would rather have this run for you — a scaled prompt panel across dozens of queries, tracked monthly, tied to a content and reputation plan that moves the number — that is exactly the kind of measurement-plus-action loop we build. Book a call and we will show you your current share-of-mention and where the biggest gaps are.