Does Technical GEO Really Matter for AI Visibility? 5,058 Hotel Websites Studied.
We tested schema, llms.txt, page speed and crawler settings against how often five AI engines named 5,058 hotels. The checklist explained very little, so we looked at what else might distinguish the hotels AI recommends most often.
Hotels are being sold a checklist for AI visibility. Add schema markup. Publish an llms.txt file. Make the site faster. Let the AI crawlers in. Most of this is reasonable web hygiene, and they now often come with a promise attached: do this and ChatGPT, Google AI Mode and the rest will recommend you.
Do these technical checklists actually make hotels more likely to be recommended by AI?
To answer that we turned to our ongoing Boutique Hotel AI Index: 5,058 single-hotel websites across 100 destinations worldwide, named by ChatGPT, Google AI Mode, Google AI Overviews, Gemini and Microsoft Copilot.
For every destination we set the third of hotels named most often against the third named least often, and asked what the first group has that the second does not.
We went in with some reservations. Recent studies outside hospitality have found that these checklist items do little to move AI citations: Ahrefs tracked 1,885 pages that added schema and saw their citations barely move, and SE Ranking found no measurable link between llms.txt and AI citations across nearly 300,000 domains. So we were not expecting miracles. We wanted to see whether that held for hotels, and a comparison of this size is where any difference should show up.
1. The technical checklist, tested
If someone tells you that adding schema or an llms.txt file will get your hotel recommended by ChatGPT, the fair question is what evidence they can show you. We went looking for it across 5,058 hotel websites, and found very little.
Does the website have any schema markup?
Yes on 71.1% of the most recommended hotels’ websites, compared to 69.5% of the least recommended.
Does the website have hotel schema markup?
Yes on 29.4% of the most recommended hotels’ websites, compared to 25.0% of the least recommended.
Is an XML sitemap available?
Yes on 74.2% of the most recommended hotels and on 74.2% of the least recommended. A real tie!
Does a fast website make a difference?
Speed scores did not separate the two groups, and the most recommended sites had heavier pages: a median homepage of 5.5 MB against 4.6 MB. The median Lighthouse lab score was 54 for the most recommended and 56 for the least recommended, out of 100. Real-user Core Web Vitals, which exist for 2,060 of 4,599 sites, were passed by 34.4% and 33.1%.
How about Javascript heavy pages with little content?
321 of 4,599 homepages (7.0%) show fewer than 100 words when JavaScript is switched off. We found no large penalty, but near-empty homepages were recommended somewhat less often, and only 57 sites were confirmed as depending on JavaScript for their content. That is too few to settle the question.
Almost every item on the checklist we measured differs by less than five points between the two groups; the full table, with intervals and the items too rare to judge, is in the method section at the end.
llms.txt
A special mention is reserved for llms.txt, a plain-text file intended to tell AI systems about the hotel and the website in a format specifically developed to be AI friendly.
In July 2026, 20.5% of the hotel homepages we could reach had one (1,033 of 5,038). By 12 November it was 26.0% (1,286 of 4,946).
About half of the files appeared to be made automatically by a website builder or SEO plugin. Taken as a whole, hotels recommended most often were about as likely to have an llms.txt file as hotels recommended least, and the files made by a builder or plugin showed no link with being recommended at all.
All of it together
Finally, we asked a straight question. If you knew everything we measured about a hotel’s website, all twelve items, how well could you predict how often AI would recommend it compared with other hotels in its destination? The simple answer is barely at all.
Together, all twelve technical measures explained just 1.4% of the differences in how often hotels were recommended within their destination. On this evidence, the technical checklist tells you very little about which hotels AI recommends.
Table 1. The technical checklist: most recommended and least recommended hotels.
| Website measure | Most recommended third | Least recommended third | Gap in points |
|---|---|---|---|
| llms.txt file | 21.8% | 18.7% | +3.2 |
| Any schema markup | 71.1% | 69.5% | +1.6 |
| Hotel-type schema | 29.4% | 25.0% | +4.4 |
| XML sitemap | 74.2% | 74.2% | +0.1 |
| Passes Core Web Vitals (real-user data) | 34.4% | 33.1% | +1.3 |
| Server response slower than one second | 17.8% | 20.0% | -2.2 |
| Blocks at least one AI crawler * | 4.0% | 4.1% | -0.1 |
| Blocks GPTBot * | 3.6% | 3.4% | +0.1 |
| Near-empty homepage without JavaScript * | 6.4% | 7.9% | -1.5 |
| FAQ schema * | 2.6% | 3.2% | -0.6 |
| Review-rating schema * | 5.7% | 4.9% | +0.8 |
| Sitemap shows an update in the last six months † | 81.5% | 74.5% | +6.9 |
| All twelve together: share of the difference in how often hotels were recommended | 1.4% | ||
2. If not the website, then what?
Since the technical checklist items showed little connection with which hotels AI recommends, we turned our attention to other things we can measure at scale.
We ran the same comparison, on the same hotels, using facts about the hotel instead of its website: its Google guest rating, its nightly price, its star class and how many reviews it has. These are not purely facts of a brick and mortar kind. They also reflect how a hotel is positioned and how guests have received it.
Together, these hotel facts explained 23.5% of the differences in how often hotels were recommended within their destination, against 1.4% for the technical checklist.
The technical checklist versus hotel facts
This chart needs a qualification. Our research prompts ask for luxury, boutique and romantic hotel recommendations, so star class and price are to a degree conceptually baked into what we are asking for.
Among the 3,125 hotels never recommended for a luxury question, the hotel facts explained 7.4% of the differences in how often hotels were recommended. That share is smaller once luxury questions are set aside, but the comparison points the same way: the hotel itself, its ratings, its reviews, its category and its pricing explain far more than the website’s technical checklist, which stays close to nothing.
The most useful of these findings is guest rating. Price and star class are partly what our questions asked for. Rating is not, and it held wherever we looked: for every type of question (Table 2), among hotels of the same star class (+21.8 points among four-star hotels alone), and among hotels never recommended for a luxury question (+12.8). Of the facts we measured, it is the one a hotel’s own operation shapes most directly.
A controlled experiment by Baig, Gillani and Ali, in which twelve AI models chose between hotels with randomly varied details, also found guest rating had the largest effect.
Table 2. Most recommended and least recommended hotels, by type of question.
Each cell: the most recommended third against the least recommended third, with the gap in points.
| Boutique questions | Romantic questions | Luxury questions | |
|---|---|---|---|
| Hotel identified as five-star | 35.1% vs 20.9% (+14.2) | 40.3% vs 20.9% (+19.4) | 64.7% vs 31.9% (+32.9) |
| Hotel identified as four-star | 48.7% vs 48.8% (-0.1) * | 45.4% vs 45.9% (-0.5) * | 33.4% vs 62.1% (-28.7) |
| Hotel identified as under four-star | 16.1% vs 30.3% (-14.1) | 14.4% vs 33.2% (-18.9) | 1.9% vs 6.0% (-4.2) |
| Guest rating 4.5 or higher | 85.7% vs 76.7% (+9.0) | 82.1% vs 66.9% (+15.2) | 85.7% vs 72.7% (+12.9) |
| More than 425 Google reviews | 49.0% vs 40.2% (+8.8) | 64.0% vs 49.9% (+14.1) | 64.5% vs 52.2% (+12.3) |
| Median Google reviews | 418 vs 310 | 617 vs 424 | 640 vs 454 |
| Price above typical for the destination | 67.5% vs 33.8% (+33.7) | 68.6% vs 35.9% (+32.7) | 73.0% vs 31.9% (+41.1) |
| Median nightly price (US$) | $296 vs $209 | $290 vs $192 | $364 vs $234 |
Price needs care. The median nightly price was $288 in the most recommended group and $178 in the least recommended. None of our questions is a budget question, so a higher price is partly what was asked for, and this does not show that AI prefers expensive hotels, or that raising rates would get a hotel named more often. The price gap does hold within a star class (+32.5 among four-star hotels), and about half of it remains among hotels never named for a luxury question (+19.4).
Star class mostly reflects the luxury question. For luxury questions the five-star gap is +32.9 points, but almost every hotel named there is four-star or higher, so it separates five-star from four-star hotels that are already at the top of their market. Among hotels never named for a luxury question, star class did not separate the groups (+2.7), while rating and price still did.
Review count shows a gap too, but a less clean one. On every type of question, the most recommended hotels were more likely to have more than 425 Google reviews (+8.8, +14.1 and +12.3 points), and their median review count was higher (Table 2).
Comparing hotels of the same star class, the difference mostly disappears, with one exception: among five-star hotels, reviews still separated the groups on luxury and romantic questions (+21.4 and +15.7 points). Because reviews partly travel with star class, we treat review count as a supporting signal rather than a finding of its own.
One checklist item looks like a hotel fact in disguise. Hotel-type schema has the one visible raw gap among the schema items. Once star class and chain membership are accounted for, the difference in naming between sites with and without hotel schema falls from +2.3 to +0.5 percentile points, with an interval that spans zero.
Put plainly, what goes with being named often is the kind of hotel it is and how its guests have received it, far more than how its website is built. That is close to how a well-informed friend would recommend a hotel, and often a long way from how the technical checklists are being presented.
3. What this analysis means for hotels
We are not arguing that the technical side of a website does not matter. Schema, a fast site and a clean sitemap are the basics of keeping a website in order, and an llms.txt file can sit alongside them with little additional effort. But there is a difference between good technical work and selling false hope that a handful of technical changes will make AI recommend a hotel more often. Our findings give us very little reason to believe they will.
These are systems that can read a website, set it beside the hotel’s reviews on Google and Tripadvisor, its OTA listings and every other mention of it, and form a view. Wondering whether they can understand a hotel without a few lines of structured markup takes us back to a much earlier internet. If AI cannot work out what a hotel is about, the cause is rarely a missing file.
SEO stopped being a checklist exercise years ago, and there is little reason to expect AI visibility to be different.
Nor should anyone expect a new formula to replace the old one. Most of the difference between hotels is explained neither by the website nor by the facts we measured, and AI recommendations move: asked the same questions twice on the same day, the engines named the same top hotel 60.2% of the time (62.8% without Microsoft Copilot; see the method section).
The industry has put a great deal of effort into making hotel websites easier for AI to read. The better question may be whether the hotel itself gives AI a reason to recommend it: how its guests rate it, and how clearly it stands out in its market.
What this analysis does not include
Hotels that no engine ever named are not in the data. We compare often-named with rarely-named hotels, so this cannot show whether a technical problem keeps a hotel from being named at all. We believe the dataset is large enough to show whether the commonly recommended technical checklist items go with being recommended more often by AI.
We are not suggesting that hoteliers ignore a sound technical base, simply that we found little measurable association between these technical checklist items and how often hotels were recommended. This observation is association only. Nothing here shows what would happen if a hotel added or removed any item.
We did not measure the quality of a hotel’s content, including such things as how the homepage is optimised, what subjects its blog articles cover or the quality and breadth of any FAQs answering travellers’ questions. We also did not investigate any hotel’s reputation beyond Google ratings and reviews, including OTA ratings, and any additional press and editorial coverage. These findings challenge the technical checklist, not the wider value of a hotel’s website or marketing.
Full methodology, statistical results and limitations
Exploratory. This analysis was not in the study’s pre-registered list of confirmatory claims.
Data: The Boutique Hotel AI Index, runs 01-15 and weekly runs to 9 November 2026; website crawls of July and October 2026; robots.txt and llms.txt checks of 20 July and 12 November 2026; Google hotel listing data of 19 July 2026. Every figure is produced by script from the study files.
Same-day yardstick: when the same prompts were run twice on the same day, the top-named hotel was the same in 60.2% of cases and the sets of hotels named overlapped by 52.0%. Without Microsoft Copilot the figures are 62.8% and 54.5%. We print both because Copilot returned usable lists for only 181 of 300 prompts in the same-day test while its run-to-run data is intact, so including it makes the engines look less consistent than the stricter figure does.
Corrections: we have corrected our own figures before and will correct this piece if we find an error. Study hub: https://kollective.agency/boutique-hotel-ai-index/
How we compared
Unit. A hotel with its own website: 5,058 single-hotel domains across 100 destinations, drawn from the 9,280 hotels the engines named. Chain corporate sites, where one domain serves hundreds of hotels, are left out.
Being named. For each hotel, the share of answers about its destination that named it. Runs 01-15 (18 June to 19 July 2026, 21,592 answered prompts) pair with the July crawl. Three four-week windows from the weekly series pair with later crawls, the last ending 9 November 2026; that window uses four engines, because Microsoft Copilot left the panel in November.
Comparison. Each hotel is ranked only against hotels in the same destination. Gaps are the share of the most recommended third with an item minus the share of the least recommended third. Intervals are 95%, from resampling the 100 destinations 2,000 times. We call a gap small enough to ignore when its 90% interval sits inside plus or minus five points; for items on fewer than one site in ten that band means little, so we do not use it. Where we give percentile points, they are the average position (0 to 100, within the destination) of hotels with an item minus hotels without it.
The technical checklist in full
Table A1. Checklist items on the websites of the most recommended and least recommended hotels (July 2026 crawl, runs 01-15).
| Checklist item | Sites checked | most recommended third | least recommended third | Gap in points (95% interval) | Reading |
|---|---|---|---|---|---|
| llms.txt file | 4,532 | 21.8% | 18.7% | +3.2 (+0.2 to +6.2) | Small at most; see the llms.txt section |
| Any schema markup | 4,580 | 71.1% | 69.5% | +1.6 (-1.7 to +5.1) | Within five points |
| Hotel-type schema | 4,580 | 29.4% | 25.0% | +4.4 (+1.4 to +7.6) | Follows hotel class; see below |
| XML sitemap | 4,532 | 74.2% | 74.2% | +0.1 (-2.8 to +3.0) | Within five points |
| Passes Core Web Vitals (real-user data) | 2,060 | 34.4% | 33.1% | +1.3 (-4.0 to +6.6) | No gap seen; fewer sites, wider interval |
| Server response slower than one second | 4,599 | 17.8% | 20.0% | -2.2 (-4.8 to +0.5) | Within five points |
| Blocks at least one AI crawler * | 4,532 | 4.0% | 4.1% | -0.1 (-1.7 to +1.3) | Too rare to rule out small effects |
| Blocks GPTBot * | 4,532 | 3.6% | 3.4% | +0.1 (-1.3 to +1.4) | Too rare to rule out small effects |
| Near-empty homepage without JavaScript * | 4,599 | 6.4% | 7.9% | -1.5 (-3.4 to +0.4) | Too rare to rule out small effects |
| FAQ schema * | 4,580 | 2.6% | 3.2% | -0.6 (-1.8 to +0.6) | Too rare to rule out small effects |
| Review-rating schema * | 4,580 | 5.7% | 4.9% | +0.8 (-0.8 to +2.4) | Too rare to rule out small effects |
| Sitemap shows an update in the last six months | 2,506 | 81.5% | 74.5% | +6.9 (+3.5 to +10.2) | Reflects an actively maintained website; see below |
* Fewer than 10% of sites have this item. With so few, the data can rule out large differences only. Schema is read from the page as the server delivers it, which is what a crawler that does not run JavaScript sees. “Gap” is the most recommended share minus the least recommended share.
llms.txt in full
We read the files to see where they came from. Many sign themselves: a website builder’s template, or a line from an SEO plugin. Others share an identical set of headings with ten or more other sites.
Table A2. Who made the llms.txt files.
| Where the file came from | July 2026 | November 2026 |
|---|---|---|
| llms.txt files read | 1,073 | 1,317 |
| Signed by a website builder or SEO plugin (Wix, Yoast, All in One SEO, Rank Math, Hostinger and others) | 407 (37.9%) | 378 (28.7%) |
| Shared template (identical headings on ten or more sites) | 193 (18.0%) | 260 (19.7%) |
| Made by a tool or template, at least | 600 (55.9%) | 638 (48.4%) |
| Not from a common builder or plugin, or not identified | 473 (44.1%) | 679 (51.6%) |
These shares are a floor: they count only the files our signatures caught. A wider set of signatures, used when we stress-tested this analysis, attributed more of the remaining files to tools, which is why we say about half, at least.
Table A3. The llms.txt gap, split by where the file came from.
| llms.txt group, compared with sites that have no file | Sites in group | most recommended third | least recommended third | Gap in points (95% interval) |
|---|---|---|---|---|
| July: made by a tool or template | 537 | 12.8% | 13.4% | -0.6 (-3.3 to +2.3) |
| July: not from a common builder or plugin | 385 | 11.7% | 7.0% | +4.7 (+2.8 to +6.8) |
| November: made by a tool or template | 596 | 15.4% | 17.5% | -2.1 (-4.8 to +0.8) |
| November: not from a common builder or plugin | 556 | 16.2% | 11.8% | +4.3 (+1.8 to +6.8) |
The July llms.txt gap. In July the gap was +3.2 (+0.2 to +6.3) points, an interval that sits just clear of zero on its own. It does not survive a correction for the 27 comparisons we ran, and it shrinks to between +1.5 and +2.1 when three runs with unusually long Google AI Mode answers, and Microsoft Copilot, are set aside. In November the gap was +1.6 (-1.7 to +4.9), an interval that spans zero.
Facts about the hotel in full
Table A4. Facts about the hotel, beside the checklist (runs 01-15).
| Hotels with data | most recommended third | least recommended third | Gap in points (95% interval) | |
|---|---|---|---|---|
| Facts about the hotel | ||||
| Five-star class | 4,100 | 40.3% | 17.0% | +23.4 (+19.9 to +27.1) |
| Four-star class or higher | 4,100 | 87.1% | 62.7% | +24.3 (+20.5 to +28.2) |
| Google guest rating of 4.5 or higher | 5,054 | 82.7% | 65.4% | +17.4 (+14.0 to +20.8) |
| Nightly price above the typical price for the destination | 4,607 | 72.1% | 31.5% | +40.6 (+36.8 to +44.5) |
| More Google reviews than is typical for the destination † | 5,054 | 56.7% | 43.5% | +13.3 (+9.7 to +16.7) |
| Checklist items, for comparison | ||||
| Sitemap shows an update in the last six months | 2,506 | 81.5% | 74.5% | +6.9 (+3.5 to +10.2) |
| Hotel-type schema | 4,580 | 29.4% | 25.0% | +4.4 (+1.4 to +7.6) |
| llms.txt file | 4,532 | 21.8% | 18.7% | +3.2 (+0.2 to +6.2) |
| Any schema markup | 4,580 | 71.1% | 69.5% | +1.6 (-1.7 to +5.1) |
| XML sitemap | 4,532 | 74.2% | 74.2% | +0.1 (-2.8 to +3.0) |
Star class, rating, review count and price are from Google’s hotel listings, collected 19 July 2026. Price is the median of Google’s one-night rates for two adults, in US dollars, across check-in dates from 25 August to 30 September 2026, the same dates for every hotel. “Typical” means the median among the hotels compared in the same destination. † The review-count gap does not hold up within a star class; see below.
Do they hold up against each other? Rating and price do. In one model with all four facts together (3,804 hotels, 95% intervals), a hotel priced above the typical rate sits 17.2 (15.1 to 19.2) percentile points higher in its destination’s naming order, one rated 4.5 or above 9.1 (7.3 to 11.0) points higher, and a five-star hotel 6.5 (4.3 to 8.9) points higher than all other hotels, four-star included. Looking only inside one star class, rating and price still separate the groups: among four-star hotels the rating gap is +21.8 (+16.5 to +26.9) points and the price gap +32.5 (+26.7 to +38.5). Review count does not: inside a star class its gap is +3.0, -0.6 and +7.2 points for three-, four- and five-star hotels, each with an interval that spans zero. We read review count as a marker of a hotel’s class and size, not as a finding of its own.
Head to head. How much of the difference between hotels, in how often they were named, does each set of measures account for? We put all 12 technical measures in this article into one model and the hotel facts into another, on all 5,058 hotels. The technical measures account for 1.4% (1.0 to 2.4), or 4.5% if we also count whether a measurement existed at all (whether a site has real-user speed data, for example, is closer to a sign of traffic than a checklist item). The hotel facts, with price, rating and reviews measured against the other hotels in the same destination, account for 23.5% (21.0 to 26.4): about a quarter. Adding the technical measures to the hotel facts adds 1.1 percentage points. Tested on destinations the models had not seen, the figures are 0.8% (technical), 23.0% (hotel facts) and 23.6% (both). Of the hotel facts, price does the most work: on its own it accounts for 17.7%, against 8.3% for star class.
Outside the luxury question. Among the 3,125 hotels never named for a luxury question (ranked on their boutique and romantic answers), the gaps were: five-star +2.7 (-0.3 to +5.7) points, rating +12.8 (+8.4 to +17.4), price +19.4 (+14.6 to +24.2); median nightly price $214 against $178. Price among independent hotels alone: +40.5 (+36.5 to +44.4).
Head to head outside the luxury question. The same two models on the 3,125 hotels never named for a luxury question, ranked on their boutique and romantic answers: technical measures 0.6% (0.5 to 1.7), or 1.2% with markers for missing values; hotel facts 7.4% (6.1 to 9.3); both 8.2%. On destinations the models had not seen: -0.2%, 6.7% and 6.4%. Of the hotel facts, price alone accounts for 4.2%, rating alone 2.9% and star class alone 0.5%.
The second number matters as much as the first. Even the hotel facts leave about three-quarters of the difference unexplained. Some of that will be things we did not measure: location, size, press coverage, how well known a hotel is. Some will be the answers themselves. The Index has shown from the start that the same question, asked twice, often returns different hotels, and splitting our 15 runs in two, the naming order repeats with a reliability of 0.84 to 0.90, so roughly a tenth to a sixth of the variation is run-to-run noise that no attribute could account for. We cannot say how the rest divides.
Head-to-head models. The outcome is each hotel’s position within its destination (0 to 100), so differences between destinations are already removed. The figures are plain R squared from linear models. Technical measures: the eleven items in Table 1 plus the sitemap date; the lower figure uses their values only, the higher one adds a marker for each missing value. Hotel facts: star class, plus guest rating and the logarithms of review count and price, each expressed relative to the other hotels in the same destination; a missing value is its own category, so all 5,058 hotels are used. Intervals come from resampling destinations 1,000 times. The held-out test fits on nine-tenths of the destinations and predicts the remaining tenth, ten times over, averaged across 20 random divisions. The reliability figure is the Spearman-Brown estimate from splitting the 15 runs into two halves (odd and even, first and second half, and a random split) and correlating each hotel’s position in the two halves.
Table A5. The hotel facts by type of question, among hotels named for that type.
| Gap in points (95% interval) | Boutique questions (3,347 hotels) | Romantic questions (3,332 hotels) | Luxury questions (1,933 hotels) |
|---|---|---|---|
| Five-star class | +14.2 (+10.5 to +17.8) | +19.4 (+15.0 to +23.7) | +32.9 (+27.7 to +38.5) |
| Guest rating 4.5 or higher | +9.0 (+5.3 to +12.7) | +15.2 (+11.4 to +19.1) | +12.9 (+8.3 to +17.7) |
| Price above typical | +33.7 (+28.8 to +38.7) | +32.7 (+28.1 to +37.4) | +41.1 (+34.9 to +47.3) |
Experimental evidence and one more checklist item
How this sits with the experimental evidence. Baig, Gillani and Ali ran a controlled experiment in which twelve language models chose among five invented hotels with randomised attributes. A top guest rating raised the chance of being chosen by 31.6 points, and a high price lowered it by 30.0 points. Their hotels were all described as four-star, so star class was not tested, and their three travellers were a business guest, a family watching its budget and an eco-minded couple: none asked for luxury. Our observational data points the same way on rating. On price it points the other way, and the two results are not in conflict: they measured which of five supplied hotels a model picks, we measured which real hotels come up when someone asks for luxury, boutique or romantic places to stay.
Two checklist items look like hotel facts in disguise. Hotel-type schema has the one visible raw gap among the schema items. Once star class and chain membership are accounted for, the difference in naming between sites with and without it falls from +2.3 to +0.5 percentile points, with an interval that spans zero. And the sitemap date is best read as a sign of an actively run website. About a third of dated sitemaps were stamped within a week of our check (978 of 2,789 in July), which is what software does on its own, and those carry most of the gap. Simply having a dated sitemap, against having none, made no difference. It is not evidence that updating a sitemap gets a hotel named.
A note on blocking GPTBot
190 hotel websites in the study blocked GPTBot, OpenAI’s training crawler, in July. 143 of them carry the exact list of eight crawlers in Cloudflare’s managed robots.txt, a setting that blocks GPTBot but not OAI-SearchBot, the crawler OpenAI uses for search. OpenAI’s documentation says the two are independent, and that a site can “allow OAI-SearchBot in order to appear in search results while disallowing GPTBot”. ChatGPT named 102 of the 190 at least once (53.7%), against 51.3% of all 5,637 hotel websites in the study: about the same, within a wide margin (a difference of about 2 points, interval -7.1 to +9.9). This is an observation, and the why is an open question. Sitter’s robots.txt study found 3.3% of hotel sites blocking any AI crawler, and also found blocking to be largely platform-driven.
Engine by engine
Table A6. Gaps by engine (points, 95% interval).
| Engine (runs 01-15) | Hotels it named | Five-star class | Guest rating 4.5+ | Price above typical | llms.txt | Any schema | Hotel-type schema |
|---|---|---|---|---|---|---|---|
| ChatGPT | 2,438 | +20.2 (+15.6 to +25.0) | +13.0 (+9.0 to +17.2) | +37.4 (+31.4 to +43.1) | +0.6 (-3.8 to +4.8) | -0.0 (-5.2 to +5.1) | -1.6 (-6.6 to +3.6) |
| Google AI Mode | 4,042 | +17.1 (+13.3 to +21.0) | +14.2 (+10.9 to +17.6) | +26.3 (+21.0 to +31.6) | +1.7 (-1.9 to +5.2) | +1.0 (-2.8 to +4.9) | +1.8 (-1.6 to +5.3) |
| Google AI Overviews | 1,714 | +22.3 (+16.6 to +28.3) | +9.8 (+5.5 to +14.6) | +39.8 (+33.9 to +45.7) | -1.3 (-6.4 to +3.8) | -1.6 (-6.5 to +3.6) | +0.8 (-4.9 to +6.8) |
| Gemini | 1,403 | +18.7 (+12.1 to +25.4) | +12.0 (+6.8 to +17.3) | +38.5 (+30.9 to +45.8) | +2.1 (-2.9 to +7.3) | -6.8 (-12.8 to -0.8) | +0.4 (-6.2 to +7.0) |
| Microsoft Copilot | 1,313 | +17.4 (+10.7 to +24.3) | +11.5 (+6.2 to +16.9) | +39.0 (+31.6 to +46.4) | -2.2 (-7.5 to +2.9) | -6.5 (-12.4 to -1.0) | +0.6 (-5.6 to +6.9) |
Each engine is compared only among the hotels it named: between 52% and 74% of hotels were never named by ChatGPT, Google AI Overviews, Gemini or Microsoft Copilot, and 20% by Google AI Mode. On that basis every hotel fact separates the two groups on every engine: five-star +17.1 to +22.3, rating +9.8 to +14.2, price +26.3 to +39.8, smallest on Google AI Mode. The checklist gaps sit close to zero on every engine. Any schema is slightly less common among the most recommended hotels on Gemini (-6.8) and Microsoft Copilot (-6.5), on smaller sets of hotels and before any correction for the number of comparisons.
Limitations
Limitations. Schema is measured as delivered by the server, before any JavaScript runs; across 4,906 sites crawled in both July and October, server-delivered Hotel schema rose from 26.5% to 30.5%. Real-user speed data covers 2,060 of 4,599 sites. Speed and JavaScript figures are from July only. Server timing comes from a single location in Europe. Hotel facts exist only where a hotel was matched to its Google listing: that worked for 97.9% of the most recommended hotels across all 9,280 named and 77.5% of the least recommended, so the least recommended are under-represented before any comparison; inside the set we compare, match quality does not track how often a hotel was named. Star class is on file for 4,100 of 5,058 hotels and a price for 4,607; prices cover one stretch of about five weeks of dates, collected on 19 July 2026. Star class is on file for more of the most recommended hotels (86.0%) than the least recommended (76.7%). Counting every hotel without one as not five-star gives a gap of +21.7; only the extreme assumption that every such hotel outside the most recommended group is five-star removes it (-1.7 (-7.0 to +3.5)). Under the same extreme assumption the price gap is still +27.8. The match between a hotel and its website was wrong in 3 of 40 sampled cases (about 7.5%); repeating the technical analysis on a stricter set of 4,677 sites left every conclusion unchanged. Names in the weekly runs were matched to the runs 01-15 hotel list with 96.6% to 98.1% success. The November check of llms.txt is not an independent repeat of July, since most of the same sites had the same kind of file both times. Answers were collected through Rankscale’s automated sessions, in English, from three fixed question wordings per destination (boutique, romantic and luxury); what a signed-in user sees, on ChatGPT in particular, can differ.
