Why One AI Visibility Score Can Mislead Luxury Hotels

Written by Shamus Smith, Founder, LuxDirect

The problem with one number

Ask six different AI assistants to recommend a luxury hotel in Mayfair and you will not get six versions of the same answer. You will get answers built by genuinely different systems. Some assistants search or ground their answers in live web results before they reply. Others rely more heavily on model memory unless the product, setting or query triggers live retrieval. A few sit between those poles, blending remembered patterns with fresh sources depending on the question.

This matters because it means visibility is not one property of a hotel. It is six different properties, one per platform, each driven by a different mechanism. A hotel that a live-web reply surfaces easily can be absent from a memory-led one. Both facts are true at the same time. Averaging them into a single score does not summarise the picture. It erases it.

How platforms differ, and three tiers of visibility

There is a useful way to think about the platforms: not as a fixed split between those that search and those that do not, but by how often, how deeply, and from which sources they retrieve fresh evidence when they answer. Some lean heavily on the live web. Some lean more on what they remember. Most move along that spectrum depending on the question.

This hierarchy should be read as a diagnostic model, not a permanent label for each company. The same platform can behave differently depending on the query, region, product mode, account setting and whether live search or grounding is enabled. What follows is a way to reason about where your effort pays off, not a fixed ranking of who wins.

At one end sit replies that lean on the live web. If a hotel ranks well and is cited across the open web, these surface it readily, because they are drawing on that web as they answer. In the middle sit platforms that search their own narrower slice of the web, or that reach for live retrieval only when the query invites it. At the harder end sit the assistants, modes or query types that weight model memory most heavily and retrieve fresh evidence least consistently. These are the places where being already known matters most, and where a property that is not already famous has the steepest climb.

For a luxury boutique hotel, that hierarchy is not academic. It maps directly onto where your effort pays off and where it barely moves the needle.

Why this hits independent hotels hardest

The mechanism that separates these platforms is also the mechanism that disadvantages independents specifically. Replies that lean on model memory tend to over-recommend the names that were already everywhere when the model was trained. A globally famous chain or a household-name grande dame is more likely to be deeply represented in that memory. An independent boutique, by definition, usually is not.

So the memory-leaning replies carry a structural bias toward incumbency that no amount of quality can shortcut in the short term. The live-web replies are the opposite. They will surface a lesser-known property the moment the live web gives them a clear, consistent, well-cited reason to. That is the difference between a problem you can influence this quarter and one that takes years.

Read that way, the split is genuinely good news for independents. It says the platforms where you can win are the ones that reward exactly what a focused hotel can build: accurate structured information on your own site, and credible editorial coverage across the web. The platforms where incumbency dominates are real, but they are not the whole board.

The same hotel, measured six ways

This is not a theory. Across our case studies, built on structured query testing across six AI systems via API, we run the same hotels, the same queries, in the same window, against each platform, and record how each one answers. The results diverge more than most operators expect.

In our Manchester study, the average rate at which a hotel was mentioned at all differed markedly by platform: one platform named the panel hotels 27.7 percent of the time, another only 17.8 percent, on identical queries in the same week. In our Edinburgh study, a single property captured 14.6 percent of mentions on one platform and just 2.0 percent on another. Same hotel, same week, a sevenfold difference depending only on which assistant a guest happened to open.

The starkest version is a single property in our study that sat inside the top eight on two platforms yet recorded zero mentions on a third. A General Manager watching only the first two would conclude the hotel was performing well. A General Manager watching only the third would conclude it was invisible. Both would be reading a real number. Neither would be reading the whole picture.

A note on what these figures measure, because it matters for a piece about measurement. Our studies record how the AI models describe and cite hotels in their answers: which properties they name, and which sources they draw on. They are a rigorous measure of what the model says. They are not a substitute for tracking a live booking button inside a consumer app, which is a different surface with its own commercial layer. We are careful to claim only what we measure. Even held to that stricter standard, the platform-to-platform spread is large, and it is the spread that makes a single blended number unsafe to act on.

What the concentration data shows

Our own live-web study of the London luxury market found visibility already concentrated at the very top. Across the six platforms, the leading four hotels captured 64.3 percent of all visibility in our live-web query set, up from 48.3 percent in an earlier training-data study of the same market. The shift appears to be a platform-mechanism story: moving from remembered patterns to live retrieval did not broaden the field in this sample, it concentrated it further. The properties outside the top group were not absent because they were lesser hotels. They were absent because the signals the platforms rely on were thinner for them.

That concentration is the incumbency effect made visible. It is steepest where replies lean on model memory and looser where they draw on the live web, which is precisely why the fix differs by platform. A hotel missing from a live-web reply has a citation and structured-data problem it can address. A hotel missing from a memory-led reply is fighting a slower battle for recognition. Same absence, two different causes, two different responses. A single score cannot tell them apart.

What a General Manager should take from this

The practical instruction is simple. Do not ask whether your hotel is visible in AI. Ask which platforms it is visible in, and why the others are not. Invisibility on a live-web reply points to work you can start now, on your own website and through editorial coverage. Invisibility on a memory-led reply is a longer game of building the kind of consistent, cited presence that eventually earns a place in the model's default recall.

A blended number hides that distinction, and the distinction is the entire actionable part. This is why LuxDirect treats visibility as a per-platform diagnostic rather than a single verdict. The Luxury Visibility Index may summarise performance at the top level, but the headline number is never the answer. The platform breakdown is. It is built from all six platforms measured separately, so a General Manager can see not just that a gap exists, but which gap it is and what closes it.

If you want to see where your hotel appears across ChatGPT, Perplexity, Gemini, Claude, Google AI Mode and Grok, run a free LuxDirect AI Visibility Scan. It shows which platforms recommend you, which competitors they recommend instead, whether guests are being routed to your direct site or to an OTA, and the first fix most likely to improve your position.