One AI answer tells you very little

AI recommendations aren't fixed search results. Ask the same assistant the same question again and the answer may change. Reword the question and different places may show up. Ask a different model entirely and the ranking can look completely different.

So we don't try to measure AI visibility from a handful of prompts. We measure patterns across repeated observations.

How we observe AI

We ask travel-related questions across ChatGPT, Claude, Gemini, and Perplexity — covering restaurants, hotels, attractions, itineraries, family travel, romantic trips, budget travel, seasonal travel, hidden gems, and more. Rather than relying on one exact prompt, we use several phrasings of similar questions, and we ask them repeatedly over time.

The questions go to each assistant's API model (currently GPT-4o, Claude Sonnet 5, Gemini 3.6 Flash, and Perplexity Sonar), which can answer differently from the consumer apps. Our pages name the assistant; AI Oddities name the exact model, since each one quotes a single answer.

Each response gets analyzed to identify the destinations, businesses, attractions, experiences, and concepts the AI recommends. Repetition is what lets us tell a place that came up once apart from something AI consistently associates with a destination.

What that lets us measure

Recommendation frequency — how often a place appears across relevant observations. A restaurant recommended in 15 of 20 responses is a different signal than one that appeared once.

Ranking — where a recommended place tends to land in an answer, not just whether it's mentioned at all.

Consistency across phrasing — AI answers can be sensitive to wording. Asking the same thing several different ways lets us tell recommendations that persist apart from ones that only show up under one specific prompt.

Model agreement — ChatGPT, Claude, Gemini, and Perplexity don't always see a destination the same way. We track where they converge and where they don't.

Change over time — we keep collecting rather than treating any single response as permanent, so we can see rankings, recommendations, and themes shift as the dataset grows.

Broader destination presence — we also ask questions that aren't about one specific destination, like the best weekend trips in a state or the best food destinations in a country. That tells us whether AI thinks of a destination even when it wasn't asked about directly.

The TakeScout score

The score is a simple summary of three signals we track:

Model agreement — the share of a destination's questions where every assistant that answered named the same place first. A question counts once at least three assistants answered it with a clear first pick, and a destination needs at least five such questions. Higher agreement means a more consistent view across models.

Confirmed-open rate — among the places AI recommends most strongly that our follow-up checks have looked at, the share confirmed still open. A seasonal place checked out of season counts as open. This is a rough proxy for how current the recommendations actually are.

Presence rate — how often a destination shows up in broader regional or national questions that weren't specifically about it. A recommendation counts fully; a plain mention counts as half. It is measured for cities and resort towns only: those questions ask for places to go, so a state, province or country can't be the answer, and presence is left out of its score.

If we don't have enough observations yet to calculate a metric reliably, we leave it out of that destination's score rather than counting it as zero. The overall score is an unweighted average of whichever metrics have enough data, and is shown only when at least two do. Only cities and resort towns are ranked against each other.

What the data does — and doesn't — mean

We don't claim to know what any individual traveler will see when they ask an AI assistant a question. AI output is probabilistic — results shift with wording, model, timing, context, personalization, and changes the AI providers make on their end.

That's exactly why we don't try to draw conclusions from one answer. We're looking for patterns that hold up across many, and the more we observe, the more those patterns are worth trusting.

How we measure themes

Theme pages answer a different question from intent pages: not what AI recommends, but what it associates with a destination — and on what evidence.

A comparable grid. Every destination is asked the same in-destination questions by each assistant. For theme numbers we take a snapshot: for each destination, question and assistant, the latest answer. Counting all answers ever collected would favor destinations that happened to be asked more often (our August 2026 rollout reached destinations alphabetically). A destination is included once at least 80% of its snapshot answers are in; an answer older than 60 days, or asked under an outdated version of the destination's name, leaves its slot empty rather than being replaced by an older one.

The tagger. Each snapshot answer is read by an AI model against a closed list of 12 themes, each with a written definition of what counts and what doesn't — the same definition shown on each theme page. For every theme it finds, the model must copy the sentence that supports it. We then check that the sentence appears word for word in the answer; if it doesn't, the link is discarded. The model also records whether the theme describes the destination itself, a named place, or neither (a passing mention), and whether the sentence names something specific.

Validation. The tagger's links have not been checked against a hand-labeled sample. Every link on a theme page comes with the sentence it rests on, so each one can be checked against the answer it was read from. Answers are re-tagged every quarter; each theme page shows the dates of the answers it uses.

Echo questions. Some questions ask about a theme directly — romantic hotels, for Romantic Experiences. Their answers are excluded from a theme's headline rate and shown separately.

The destination itself, or its places. Destinations are ranked, compared and highlighted only on answers that tie a theme to the destination itself. An answer recommending a charming town near a city says something about that town, not about the city, so answers that tie a theme only to a named place are counted separately and listed with the places they name, leaving out the destination's own neighbourhoods.

Lift compares a destination with others of the same type only — cities with cities, regions with regions — because a country and a city answer the same question at very different scales.

Small samples. We don't show a rate backed by fewer than 3 linked answers; pages say “not enough observations yet” instead. Lift needs at least 5 linked answers and 5 published destinations of the same type.

Snapshot

For every destination, the latest answer from each assistant to each of the same in-destination questions. Every destination is asked the same questions, so rates compare like with like. A destination is included only once most of its answers are in; answers older than the snapshot's age limit, or asked under an outdated destination name, leave their slot empty.

Linked to

An answer links a theme to a destination when it ties the theme to the destination itself or to a named place in or near it, backed by a sentence quoted from the answer. Passing mentions ("if you're on a budget…") don't count.

Association rate

Answers that tie the theme to the destination itself, out of all answers about the destination in the snapshot, shown as "k of N answers". Destinations are ranked and compared on this rate. Answers that link the theme only to a named place in or near the destination are counted separately, as are answers to questions that ask about the theme directly.

Only through places

Answers that link the theme to a named place in or near the destination but not to the destination itself, out of all its answers. Shown beside the association rate, never ranked. The places listed are the ones those answers name most often, leaving out the destination's own neighbourhoods.

Describes the destination

Among linked answers, the share where the theme describes the destination itself ("one of the most romantic cities") rather than only a specific place in it.

Concrete

Among linked answers, the share whose quoted evidence names a specific place, business, community, nation, event, dish or practice.

Questions that ask about the theme

Some tracked questions ask about a theme directly — romantic hotels for Romantic Experiences, for example. Answers to them would inflate the rate, so they are excluded from the headline and shown as their own rate.

Lift

A destination's association rate (answers describing the destination itself) divided by the average rate of published destinations of the same type (cities with cities, regions with regions). Shown only when at least 5 destinations of that type are published and the destination has at least 5 linked answers.

Intent rate

For one traveler question type, the share of all its answers (across published destinations) that link the theme. Intent lift divides it by the theme's overall rate.

Limits. Answers to destination-specific questions show association, not that AI recommends one destination over another. How often AI brings up a destination unprompted for a theme — the cross-destination recommendation tier — is not measured on theme pages yet.

Applying these measures to your own destination: How to measure AI visibility for tourism destinations.

Questions about how a score is calculated, or something that looks off? Get in touch.