How to Prompt-Track Your Hotel’s AI Visibility

Random, or just structured?

For twenty years we measured search with rank trackers: one keyword, one position, checked daily. It worked because Google’s results were stable enough to pin down. Most hotels now try to measure AI the same way — one prompt, run once, scored as in or out — and conclude it is hopeless.

They are half right. A single AI answer really is close to a coin flip. But “variable” is not the same as “random.” When I measured hotel recommendations across dozens of city-and-tier combinations, the instability was not noise — it was a pattern. Constrained queries (“family hotels in Berlin”) were remarkably stable. Broad, competitive ones (“boutique hotels in Paris”) wobbled hard. Same engine, same week — the structure was in the question, not the dice.

Why one run lies (and why it is still measurable)

These numbers are from my own hotel rankings-consistency study — the same prompt, repeated, across many markets:

Read those together and the conclusion writes itself. One run tells you almost nothing. Monthly tracking watches a surface that has already turned over since you last looked. And any single “AI visibility score” that blends engines is an average of things that should never have been averaged.

But the spread between 17% and 96% is the hopeful part: the variance is governed by things you control or can hold fixed — the city, the guest persona, the phrasing, the engine, the country you ask from. Pin those down, repeat the question, and the noise collapses into a signal. The rest of this guide is how.

Where prompt tracking goes wrong

Most prompt-tracking setups make the same handful of mistakes. Each one is fixable, and the fixes are the six steps that follow.

Build a persona × location panel

A panel is a fixed set of prompts you run forever, so movement means something. I build hotel panels from the two dimensions a real guest varies: who they are and how wide they cast the net.

Personas. Phrase each prompt the way a specific guest asks, with their need as the reason your hotel comes up — never your amenity list as keywords. “I’m a keen cyclist visiting Paris, any hotels set up for that?” is a persona prompt. “Paris hotel with a smart-trainer studio” is keyword-stuffing that only flatters your own metric. A workable controlled set: couples, solo leisure, families, solo business, luxury — plus the niche you are genuinely the best answer for (a cyclist, a dog owner, a wellness traveller).

Locations. Guests zoom in and out. Track a nearzone (the walkable neighbourhood — “Le Marais,” “the Gothic Quarter”) and a wide zone (“central Paris”). Your hotel is a strong match for both, against very different competitive sets — and, as the consistency data shows, the narrower one is far easier to win and to measure.

Cross the two — a handful of prompts per persona × location bucket — and you land around 40–60 discovery prompts that read like real planning questions. Add a small branded tier (“Tell me about Hotel X”) purely as a sanity and data-quality check. Then freeze the set and version it: only add prompts in batches, never edit an old one, or you lose week-over-week comparability.

Repeat every prompt, and keep the engines apart

This is the single change that turns noise into signal. Run every prompt several times — five is a sensible floor — per engine, every week, and keep ChatGPT, Perplexity, Gemini and Google AI as separate lines. Repeats average out the run-to-run wobble; weekly matches how fast the engines churn their sources; separate lines stop one strong engine from hiding four weak ones.

Two settings move the answer as much as the wording does, so fix them and write them down:

• Country / IP. A French IP and a Japanese IP get different hotel answers. Decide whose view you are measuring — your home market plus your top one or two origin markets — and route from those countries deliberately, rather than letting a tool pick at random.
• Clean, logged-out sessions. Personalised history skews the result. Measure the cold-start answer a new guest sees, and treat it as a floor, not the whole truth.

By hand this does not scale past one hotel. At volume you trigger the prompts through a scraping layer that captures the real chatgpt.com answer surface from a chosen country — which is exactly what my own pipeline does behind the scenes.

Report rates, not booleans

Once each prompt has run several times, stop reporting yes/no. Report a rate: appeared in 3 of 5 runs is a 60% mention rate. With only five runs the margin is wide, so the rule is simple — do not trust a single week’s number, trust the trend across three or more weeks. One green week is weather; three climbing weeks is a result.

Record five things on every run, not just “were we there”:

• Mentioned — your name appears in the answer text.
• Cited — your own website is linked (and at what position).
• Sentiment — are you the recommendation or the warning?
• Attributes — the words attached to you: “quiet,” “great location,” “pricey,” “dog-friendly.”
• Co-cited domains — every other site the answer leaned on (more on that below).

When you score zero, prove the zero is real

A hotel that never appears looks invisible. But a zero can be an artefact — a weak prompt, run-to-run variance, or the wrong country — rather than genuine absence. Before you accept “the AI has never heard of us,” climb a short ladder. This is the part I find hoteliers skip, and it has saved me from the wrong verdict more than once.

You are truly invisible only if you stay at zero all the way up the ladder. Otherwise, record the level that surfaced you — “invisible to default prompts, appears only when named from a French IP” is a precise, fixable diagnosis, not a shrug.

Measure the booking journey, not just turn one

Guests do not ask one question and book. They have a conversation, and AI is built for exactly that. A flat prompt list only ever measures the opening line. Take your highest-intent prompts and run them as a short journey, as one conversation, scoring every turn:

The metric a one-shot tracker can never see is persistence: are you still in the answer at Selection if you were there at Exploration? Surfacing once near the top and vanishing by the booking turn is a very different outcome from carrying all the way through — and only the journey view tells them apart.

Track which sources each engine trusts

This is where measurement turns into a to-do list. Record the full set of domains every answer cites, and a pattern appears fast: each engine grounds hotel answers on different places. One leans on review and comparison sites, another on your Google Business Profile and Maps, another on its own retrieval of your site and blog.

That co-citation list is the most useful by-product of the whole exercise. It is a leaderboard of who gets cited instead of you — the OTAs, TripAdvisor, the listicles — and it tells you which lever to pull per engine: review velocity and comparison content to win one, schema and Google Business Profile to win another, your own structured pages to win a third. A flat “you rank 6th” never tells you that.

A version you can run this week

You do not need a pipeline to start. You need discipline and a spreadsheet.

1. 1
   Write 12 prompts

   Two branded, six persona (the guests you genuinely serve), two neighbourhood, two generic. Freeze them.

2. 2
   Pick your engines

   ChatGPT and one of Perplexity or Gemini to start. Separate columns, always.

3. 3
   Run each 5×, logged out

   Same day each week. Note the country if you switch it. Yes, it is tedious; that is the point.

4. 4
   Score five fields

   Mentioned, cited, position, sentiment, co-cited domains — per run.

5. 5
   Report rates, not booleans

   Mention rate per prompt per engine. Watch the trend over three weeks before you believe anything.

6. 6
   Change one thing, then watch

   Adjust one item on your site, log the date, and see if the line moves over the next runs. Never claim cause — show timing.

FAQ

Further reading

This guide is the measurement layer of a larger body of hotel AI-search work. The variance figures above come from my own rankings-consistency study; the source-grounding split runs through the whole research library.