Written by Shamus Smith, Founder, LuxDirect
LuxDirect analysed more than 9,000 valid AI responses across four platforms on 12 and 13 April 2026: ChatGPT, Gemini, Grok, and Perplexity. The study tested live web-enabled AI responses against 25 London luxury boutique hotels, 20 of which were also covered by LuxDirect's earlier London training-data study, using the same 45-query set.
Live web search concentrated visibility rather than broadening it. The top five hotels captured 66.5% of all AI mentions across the full 25-hotel set. On the 20 hotels common to both studies, the top five held 69.9%, against 57.2% in training mode, an increase of 12.7 points.
The picture on booking sources is more nuanced than a single average suggests. Across all responses, an OTA brand appeared slightly more often than a hotel's own site. But that balance reverses sharply at booking intent, and it varies enormously by platform. Half of all responses named no booking source at all.
All figures are from structured API queries, not consumer chat apps, and measure what AI responses named, not clicks or bookings.
Methodology and scope
Findings are based on structured query testing across four AI platforms via direct API, with live web search enabled on each. For this study, "live web" means web-enabled API responses generated at query time, not static model memory and not consumer app behaviour. Real-user behaviour in the consumer apps may differ.
The comparison baseline is LuxDirect's earlier London study, which used the same 45 prompts in training-data mode. That study covered a different platform set, and 20 of this study's 25 hotels. Where a training-mode comparison is made below, it is computed on those 20 shared properties only, so the two sides of the comparison describe the same market.
A fifth platform was scanned but its per-hotel evaluation failed to write complete records, leaving coverage too partial to report. It is excluded from every figure in this article. This is not search ranking. It is how AI systems construct recommendations when drawing on the live web.
Live web search concentrated visibility rather than broadening it
The pre-study hypothesis was simple. Live web retrieval should surface fresher, more varied sources, and so broaden the field of hotels AI recommends. The data showed the opposite.
Across more than 9,000 valid responses, the four platforms produced 731 hotel mentions. The distribution is far from even. The top four hotels captured 59.2% of all AI mentions and the top five captured 66.5%. The bottom five collectively held 2.3%.
On the 20 hotels shared with the earlier training-data study, the top five held 69.9% under live web search against 57.2% in training mode, an increase of 12.7 points. Live web search did not dilute concentration; it intensified it.
The movement underneath that headline is the more useful finding. Concentration did not lift every leading property equally. The largest single gain was 4.7 points and the largest single loss was 5.2 points, and both belonged to hotels inside the original top ten. Live web retrieval appears to reward properties carrying dense, current reference material, and to penalise those whose visibility rested on older training signal.
How share of voice is calculated
Share of voice here is a hotel's count of mentions as a percentage of all target-hotel mentions in the study. A hotel counts once per response, however many times it is named in that response. Mention rate is the percentage of that hotel's own responses in which it was named.
The two measures answer different questions. Mention rate tells you how often a hotel appears. Share of voice tells you how much of the total available visibility it holds. A hotel can hold a large share of a small market or a small share of a large one, and the difference matters when planning where to invest.
Definitions are stated here because they change the result. A calculation that counted every occurrence rather than every response would produce a different concentration figure, and readers recomputing from the tables below should use the definition above.
Full visibility rankings: all 25 properties
Hotel identities are anonymised. Full named reports are available to clients.
| Rank | Property | Share of voice | Mention rate |
|---|---|---|---|
| 1 | Property A | 19.7% | 40.4% |
| 2 | Property B | 14.0% | 28.7% |
| 3 | Property C | 13.3% | 27.4% |
| 4 | Property D | 12.2% | 25.4% |
| 5 | Property E | 7.3% | 15.0% |
| 6 | Property F | 5.1% | 10.5% |
| 7 | Property G | 3.8% | 7.9% |
| 8 | Property H | 3.4% | 7.2% |
| 9 | Property I | 2.9% | 5.9% |
| 10 | Property J | 2.7% | 5.7% |
| 11 | Property K | 2.6% | 5.4% |
| 12 | Property L | 2.5% | 5.2% |
| 13 | Property M | 1.6% | 3.4% |
| 14 | Property N | 1.2% | 2.5% |
| 15 | Property O | 1.1% | 2.3% |
| 16 | Property P | 1.0% | 2.0% |
| 17 | Property Q | 1.0% | 2.0% |
| 18 | Property R | 0.8% | 1.7% |
| 19 | Property S | 0.8% | 1.7% |
| 20 | Property T | 0.8% | 1.7% |
| 21 | Property U | 0.7% | 1.4% |
| 22 | Property V | 0.7% | 1.4% |
| 23 | Property W | 0.5% | 1.1% |
| 24 | Property X | 0.4% | 0.9% |
| 25 | Property Y | 0.0% | 0.0% |
Top four share: 59.2% · Top five share: 66.5% · Bottom five share: 2.3%
One property registered zero mentions across all responses, on every platform and in every phase, despite being an established London luxury hotel. AI recognition of the property exists; AI recommendation does not.
Twelve of the 25 properties sit below 3% share of voice. That is the part of the table most hotels are actually in, and it is where the gap between recognition and recommendation is widest.
Mention rates are low, and vary sharply by platform
Averaged across the 25 properties, Gemini named a given target hotel in 10.6% of its responses, Grok in 8.4%, Perplexity in 7.6%, and ChatGPT in 6.4%.
That figure needs reading carefully. It is the average across the whole field, not the likelihood that any luxury boutique hotel appears. The leading property was named in 48% of Gemini responses, while most of the field sat in low single digits. The average is low because the tail is long, not because AI rarely recommends London hotels.
The commercial reading is direct. As more travellers move to AI tools that search the live web, the visibility gap between the top four and everyone else widens rather than narrows.
Platform divergence: the same hotel can be strong on one platform and absent on another
A hotel's AI visibility is not a single number. It varies substantially by platform. The table below shows the percentage of each platform's responses that mentioned each property, across all query phases combined.
| Property | ChatGPT | Gemini | Perplexity | Grok |
|---|---|---|---|---|
| Property A | 27% | 48% | 37% | 51% |
| Property B | 22% | 28% | 30% | 35% |
| Property C | 28% | 33% | 24% | 24% |
| Property D | 24% | 29% | 26% | 23% |
| Property E | 10% | 18% | 18% | 15% |
| Property G | 6% | 16% | 6% | 5% |
| Property K | 3% | 11% | 3% | 4% |
Selected properties shown. Figures are the percentage of that platform's responses naming each property across all phases.
Gemini surfaced the leading property in 48% of its responses against 27% on ChatGPT. Several mid-table hotels showed the same split: visible on Gemini, near-absent on ChatGPT or Grok.
The platform a traveller opens determines whether a hotel exists in their consideration set at all. Monitoring a single platform produces a misleading picture of where a hotel stands.
Which booking sources AI names, and when
The study classified every response by which booking sources it named: none, an OTA brand only, the hotel's own site only, or both. This measures what the AI's response text mentions, not where any booking goes. We did not track bookings, clicks, or transactions.
The distinction that matters here is between a cited source and a booking source. AI answers are dense with references, but a response that cites a hotel's rating from a review site is using that site as evidence, not sending a guest there to book. Review sites, ranking lists and rating citations are therefore not counted as booking sources, even when the site named also sells rooms. A manual audit of a 40-response sample found this classification correct in 35 cases, with the remainder shifting the headline figure by roughly one point.
| Booking source named | Share of responses |
|---|---|
| None | 49.1% |
| OTA brand only | 18.1% |
| Hotel direct only | 13.7% |
| Both | 17.0% |
| Unclassified | 2.1% |
Half of all responses named no booking source at all. That is the largest single finding in this section and the one most hotels can act on, because a response that names no source is a response where the guest is left to search again.
Across all responses, an OTA brand appeared in 35.0% and a hotel's own site in 30.7%. The overlap is large, so this describes which sources surface, not a measured split of business.
The booking-intent finding is the opposite of what we expected
We expected OTA presence to rise at booking intent and direct presence to fall. OTA presence does rise. Direct presence rises further.
| Query phase | Responses | OTA presence | Direct presence |
|---|---|---|---|
| Discovery | 5,688 | 35.0% | 23.7% |
| Comparison | 1,600 | 21.6% | 5.6% |
| Booking intent | 1,800 | 47.1% | 75.1% |
At discovery and comparison, OTA brands appear more often than hotel websites. At booking intent that reverses decisively: a hotel's own site appears in three quarters of responses, against an OTA brand in fewer than half.
One caveat belongs alongside that number. The booking-intent query set includes prompts that explicitly ask how to book directly, which will pull direct mentions upward. The honest reading is that when a guest signals clear transactional intent, AI systems are willing to name the hotel's own site, and frequently do.
That reframes the opportunity. The weakness is not that AI hands transactional guests to intermediaries. It is that at discovery, where the consideration set is formed, half of all responses name no source at all and OTA brands lead those that do.
Platform variance matters more than the market average
The market-wide averages conceal a tenfold spread. The table below shows how often each platform named an OTA brand and a hotel's own site.
| Platform | OTA presence | Direct presence | No source named |
|---|---|---|---|
| ChatGPT | 49.9% | 57.0% | 24.8% |
| Gemini | 7.1% | 15.6% | 83.0% |
| Perplexity | 17.4% | 15.8% | 73.5% |
| Grok | 71.6% | 38.1% | 16.5% |
Platform figures are computed on deduplicated records. The aggregate above includes unclassified responses, so the two sets are not directly comparable by averaging.
Gemini named a booking source of any kind in 17% of responses. Grok named an OTA brand in 71.6%. These are not variations around a mean; they are different systems behaving differently.
A hotel that measures its OTA exposure on one platform and generalises will reach the wrong conclusion. The market average of 35.0% describes no platform in the study particularly well.
The voice describing your hotel is increasingly a third party
When AI describes or points to a hotel, it leans on the sources it surfaces. On live web search those sources are often OTA listing pages, because they are structured, current, and dense with the signals retrieval systems favour. A hotel's own website often is not.
The result is an information hierarchy where the reseller's page is the one the retrieval layer cites, so the intermediary's description tends to shape how the hotel is presented.
There is one moment when this can be reversed. When a property first opens, resellers have not yet built out its listings, and the hotel's own site can briefly be the source AI cites most. That advantage erodes as resellers catch up, which makes a launch the right time to embed direct signals.
What hotels can influence
Retrieval systems reward clarity. A property that publishes accurate, structured, machine-readable facts gives AI something confident to anchor to.
- Confirm the website identifies the property as a hotel in its structured data.
- Ensure canonical references describe the current property accurately, including the name the property actually trades under.
- Treat the direct site as a primary source AI should cite, not an afterthought.
- Check discovery-phase visibility separately from booking-phase visibility. They behave differently.
None of this guarantees an outcome. AI systems evolve, and observations reflect current behaviour rather than permanent law. But the direction of influence is clear, and the launch window is finite.
Summary
CS10-London Live Web analysed more than 9,000 valid AI responses across four platforms, testing live web search against 25 London luxury boutique hotels using the same 45-query set as the earlier training-data study.
Live web concentrated visibility rather than broadening it. The top five captured 66.5% of all mentions across the 25-hotel set, and 69.9% on the 20 hotels shared with the training-data study, against 57.2% in training mode. Across the 25 properties, mean mention rates ranged from 10.6% on Gemini to 6.4% on ChatGPT.
Half of all responses named no booking source. OTA brands led at discovery, but a hotel's own site appeared in 75.1% of booking-intent responses against 47.1% for OTA brands. Platform variance was extreme, from 7.1% OTA presence on Gemini to 71.6% on Grok.
All figures are API-derived and measure what the AI's response named, not booking outcomes. The concentration pattern is established but not fixed, and properties that address their AI signals now will be in a materially stronger position than those that engage reactively.
Frequently asked questions
Did live web search broaden or concentrate AI hotel visibility in London?
It concentrated it. The pre-study hypothesis was that live web retrieval would surface fresher, more varied sources and broaden the field. The data showed the opposite. On the 20 hotels shared with the earlier training-data study, the top five captured 69.9% of AI mentions under live web search, against 57.2% in training mode.
Do AI platforms send booking-intent guests to OTAs?
Not in this study. At booking intent, a hotel's own site appeared in 75.1% of responses against an OTA brand in 47.1%. OTA brands led at discovery, where the consideration set is formed. Note that this measures what the response text named, not where any booking went, and that the booking-intent query set includes prompts explicitly asking how to book direct.
How much do AI platforms differ from one another?
Enormously. OTA presence ranged from 7.1% of responses on Gemini to 71.6% on Grok. The leading property appeared in 48% of Gemini responses and 27% of ChatGPT responses. Monitoring one platform and generalising will produce a misleading picture.
Why is a hotel's launch window important for AI visibility?
When a property first opens, resellers have not yet built out their listings, so the hotel's own site can briefly be the source AI cites most. This advantage tends to erode as resellers catch up. Embedding accurate, structured direct signals during the launch window helps a property hold that lead for longer.
How many London hotels had zero AI visibility in the study?
One of the 25 targeted properties registered zero mentions across all responses, on every platform and in every phase. It is an established London luxury hotel. AI recognition of the property exists, but AI recommendation does not.
What counts as a mention in this study?
A hotel counts as mentioned when its name appears in the AI's response text, once per response regardless of how many times it is named. Share of voice is that hotel's mentions as a percentage of all target-hotel mentions in the study. Mention rate is the percentage of that hotel's own responses in which it appeared.
LuxDirect sits between your hotel and the AI discovery layer. We align what AI says about you with the hotel you actually run.
Every week we systematically monitor how six leading AI platforms recommend your hotel across high-intent guest searches. We show you where AI is naming OTA brands instead of your own site, where competitors are outperforming you, and where your positioning is weak or underrepresented. Then we resolve the structural issues driving it and strengthen your direct booking position within the AI layer.
You do not need an internal technical team. LuxDirect operates as a visibility concierge. You approve. We execute.
Starting at £99 per month. The average luxury OTA commission runs between 18% and 22%. If LuxDirect recovers just one booking per month from an AI-mediated OTA referral back to your direct site, the service has paid for itself several times over.
Run AI Audit →