# An agent's view of hotels

Published by AuraScope on 8 October 2026. Last updated 8 October 2026.

Canonical page: https://research.aurascope.co/research/an-agents-view-of-hotels

We asked AI agents eight stay questions about eight hotels in New Zealand and Australia, then re-fetched the pages the agents used. Agents mostly answered, often without reading the hotel's own page, and apart from one hotel that refused every client, the failures AgentReady isolated turned up in booking engines.

About this study

[Builds on Vercel and Ora.ai's AgentReady study](https://vercel.com/kb/guide/make-your-site-readable-by-ai-agents)

AuraScope. 381 agent runs and controlled fetch probes on eight hotel sites in New Zealand and Australia, run 24 to 28 Sep 2026 with corrections to 30 Sep.

## Abstract

Vercel and Ora.ai's AgentReady study isolated two ways AI agents miss answers, content that appears only after JavaScript runs and servers that refuse automated visitors, but it sampled no hotel. To see whether those failures occur where agents go on hotel sites, and what agents do once refused, we put eight stay questions to three AI agents about each of eight hotels in New Zealand and Australia. Agents succeeded in 306 of 381 runs, but for 91 of their 294 correct answers no page the agent opened stated the answer, most of them from agents that used web search instead of reading the hotel's site. Where agents did read, a facts page such as an FAQ answered most often, as documentation did on software sites. Re-fetching the pages agents used with controlled clients showed both failures: one hotel challenged every client on every page, yet agents succeeded in 34 of its 48 runs, mostly without reading the answer on any page. Elsewhere the failures sat at booking engines, which a plain fetch could not read at seven of eight hotels, three of them sending most of their challenges with a success code. The hand-picked hotels show which failures exist, not how common they are, and a check of one hotel's site would start at the booking entry, then ask whether its facts page states each stay fact.

1 · Why we ran this experiment

## What a software-site study left open for hotels

### Where this started

A hotel's website answers what guests ask before booking: when they can check in, whether the dog can come. An AI agent, software that reads websites and acts for a person, can use those answers only if it can reach and read the page.

In [Cloudbeds' research](https://www.cloudbeds.com/hotel-ai-recommendations/data/), across 810 prompts to ChatGPT, Perplexity and Gemini, official hotel websites received 13.6% of citations against 55.3% for online travel agencies. The cause is unmeasured, but the gap shows what is at stake.

Vercel and Ora.ai's [AgentReady study](https://vercel.com/kb/guide/make-your-site-readable-by-ai-agents) planted five answers on a documentation site they built. Where the answers appeared only after the page's JavaScript, its code for the browser, had run, a plain fetch that only downloads the page found none. A JavaScript-running client found all five, and a server refusing agent traffic with a 403, the code for access refused, kept both clients out. Unrepeated by others, these [probes](https://github.com/agentready-org/standard/blob/main/data/README.md) are Signal, our label for one credible source with a stated method.

The study also watched agents on 25 software sites. A documentation page was the usual next stop after the homepage and the commonest source of answers, a route the study turned into advice for site owners. Observed rather than controlled, this pattern is Signal too, and the study sampled no hotel.

Our earlier paper, [A hotel's take on agent readiness](https://research.aurascope.co/research/hotels-take-on-agent-readiness), adopted the retrieval lessons for hotels, since failures come from how a page is delivered. Hotel sites have no documentation tree, so that paper left the route pattern aside and named the next test: rerun the probes on hotel sites.

### Our questions

So three questions stayed open. Do AgentReady's two failures exist on hotel pages agents use, including the booking engine, the system that shows prices and takes bookings? Do agents answer from the hotel's own pages? Do they move through hotel sites as through software sites? The answers show whether checking a hotel's site as agents read it is worthwhile, and what to check first.

### The hypothesis

We believe AgentReady's two failures, content needing JavaScript and servers refusing or challenging automated visitors, also occur on hotel pages agents use for stay questions. We also believe agents answer mostly from a few of the hotel's own pages.

We further expected runs to fail where pages could not be delivered, and the answering page to vary by question and site, a bet against the software-site pattern. It may seem obvious that agents will fail when they cannot read the content of the page. The interesting part is whether the operator knows the agent has failed. We are seeking clarity of what actions the agent takes once it has been denied and how effective lifting agent blockers can be.

Eight sites cannot show how often a failure occurs, so under our existence bar one site showing it on a page an agent used counts as reproducing it. Four outcomes, fixed in advance, would count against the hypothesis:

1. Plain fetches routinely read stay facts on booking-engine and JavaScript-built pages.
2. No site blocks or challenges a request announcing itself as an AI agent.
3. Agents answer every question from the homepage alone.
4. Runs succeed as often on blocked or empty pages as on others.

2 · How we ran it

## Fixing the answers first, then watching agents

An agent can miss a hotel's answer by failing to reach the page, failing to read the fact, or reading it and answering wrongly. One run mixes all three, so the experiment records each separately.

### The eight hotel sites

Asking whether a failure exists, not how often, calls for varied sites and booking engines rather than a random draw. We picked four independent and four brand hotels, with engines from at least four vendors. At brand sites agents started from the brand's root, which lists many hotels.

| Ownership | New Zealand: booking engines | Australia: booking engines |
| --- | --- | --- |
| Independent | Two hotels: Cloudbeds, Mews | Two hotels: Amadeus, SynXis |
| Brand | Three hotels: Amadeus, Amadeus, the brand's own engine | One hotel: SynXis |

Table 1. The eight hotels by ownership and country, with each hotel's booking engine.

### Stage 1: fixing the answer key

Every hotel got eight questions: check-in, check-out, cancellation, parking, pets, street address, room types, and a room's price for the test stay. A run can be scored only against the hotel's own answer, so Stage 1 fixed all 64 answers before any agent ran.

Firecrawl's crawler mapped each property into a manifest of 117 to 759 pages, the only pages where a fact could count as published. An extraction agent found each fact there, an AI judge described below confirmed it on the live page, and we reviewed every one. The final key held 60 published facts, each verified on a page we call its reference page, and four blanks where the hotel publishes nothing.

### Stage 2: the agent runs

Each agent got only the hotel's name, its web address and one question, and was observed rather than steered. We used AgentReady's two harnesses, software wrapping a model with tools: Anthropic's [Claude Agent SDK](https://code.claude.com/docs/en/agent-sdk/overview) and Vercel's [eve](https://github.com/vercel/eve). Claude Sonnet 5 ran under both and GPT-5.6 Luna under eve, making three pairs of harness and model. Neither model is AgentReady's, so comparisons hold for harnesses and pages, not models.

How an agent reads decides what it sees, so each pair asked every question with two tool sets. The fetch-only set was the harness's own fetch and web search; the browser set was [agent-browser](https://www.npmjs.com/package/agent-browser), a real Chrome browser without search. Runs were capped at 20 tool calls, or 40 with a browser, and three of the 384 hit a token budget, leaving 381 scored. All ran from a Sydney datacentre on [exe.dev](https://exe.dev/), so results describe a datacentre visitor, though the SDK's fetch and the providers' web search may originate elsewhere.

### Stage 2: the fetch probes

A failed run mixes model, harness and site, so it cannot show whether the page would give any client the fact. So we repeated AgentReady's controlled test on those pages with three probe clients, each fetching every page once. Two were plain fetches announcing different names, Anthropic's Claude-User fetcher or desktop Chrome, and the third was a headless browser.

By design the probes read 249 pages: per site, the homepage, every fact's reference page, up to 40 pages agents used most, and the booking entry. Each reference page served as a positive control, a check where the fact is known to be present.

### The judge

A hotel may write 2pm where an agent writes 14:00, so a judge decided whether a text states a fact: a TypeSafe [System One](https://docs.typesafe.ai/concepts/system-one.md) model that answers narrow questions with a probability. It scored each answer: a run succeeded when the answer was correct, or the agent said it could not find a fact the hotel does not publish. The judge also checked every page text the agent received: an answer counts as read on a page when one of those texts states the fact. Search results, whose text is not logged, never count.

To test the judge, three Claude agents labelled samples of its calls blind and overruled 14 answer verdicts it was unsure of. Across five kinds of judgment on 101 items, agreement ran 85% to 99%. The judge was weak on hedged answers, which give a value while saying the hotel does not state it, so all answers on unpublished facts were classified by hand. As a check on those Claude labellers, a person also labelled 39 items without seeing the judge's verdicts or the labellers' answers, and agreed with the labellers on 29 of the 34 items both had labelled. That check matters because the labellers ran on Claude, the same family of model as two of the three agent pairs, so they could share those agents' blind spots, and 39 items check them only loosely. The 20 wrong answers gave one more test. If the judge were marking correct answers as wrong, some of those 20 would be answers the agent had read on the hotel's page, and the judge's page check would show the hotel's fact on a page that agent received. It found the fact on a received page for none of the 20, so nothing suggests the judge was failing right answers. Together these checks make the verdicts sound enough to count, as this paper does, but not precise enough to quote as exact rates, which is why we rate them Signal.

### Strengths and limits

The key was frozen before any agent ran, and the four blanks act as a negative control, a check where any value given came from elsewhere. A hotel that challenged every probe was re-fetched on a second date, so that block counts as Confirmed.

The same choices set the limits. Three Amadeus and two SynXis engines, and Cloudflare, a service that sits in front of websites, at all eight, mean one vendor's behaviour can look like a pattern. One run per combination ranks no pair, and the Claude Agent SDK's fetch returns a page summary, so its fetch-only runs may understate reading. Eight answer-key corrections and some scoring rules came after the verdicts. No rate here says how common a mechanism is among hotels, ranks models, or describes the consumer assistants travellers use.

3 · What we found

## Answers, pages, routes and failures

### How often agents answered

Of the 64 hotel-and-question pairs, 60 have a published fact, 11 of them on the booking entry. Of the 381 runs, 306 succeeded, though room price was scored correct in only 6 of 47. Setting price aside, facts on the hotel's own site were answered correctly in 272 of 287 runs and read on a page in 198. Facts only at a booking engine's own web address, at three hotels, were correct in 17 of 30 and read in 4. Agents mostly answered, less often where the answer lived in the booking engine.

- **306 of 381** Agent runs that succeeded across all eight hotels and all eight questions.
- **272 of 287** Runs answering correctly on facts published on the hotel's own site, room price aside.
- **17 of 30** Runs answering correctly on facts published only at a booking engine's own web address.

Figure 1. Runs that succeeded, and runs answering correctly by where the hotel publishes the fact. 381 agent runs, 24 to 28 Sep 2026.

### Reading the hotel's page

A correct answer can come from the hotel's page, a search summary, another site's listing or the model's memory, and only the first shows the hotel's site working. Of the 294 answers scored correct, excluding right not-found replies, 203 were read on a page the agent opened and 91 on none, a Signal that splits by how agents looked. The table's last column adds 5 browser runs that read the fact, then hit the tool-call cap.

| How the agent looked | Runs that succeeded | Runs on published facts reading the answer on a page |
| --- | --- | --- |
| Fetch-only, did not search | 64 of 67 | 60 of 66 |
| Fetch-only, searched | 94 of 125 | 25 of 114 |
| Browser | 148 of 189 | 123 of 178 |

Table 2. Success and reading by how the agent looked, across 381 runs.

Fetch-only agents that searched gave 66 of the 91 unread answers: 34 cite a hotel page carrying the fact, 4 one not stating it, 3 another site and 25 nothing. A citation is the agent's claim, but most that name a source point to the hotel, so search replaced reading the hotel's page, not the hotel as a source.

### The answering pages

We sorted each fetched page by its address and whether it held a verified fact, so routes could be compared with AgentReady's. A front page is where an agent starts: a homepage, or a brand site's property page or root. A facts page, the hotel counterpart of documentation, is an FAQ or terms page or another page holding a verified fact other than rooms or contact.

The hypothesis expected the answering page to vary by question and site. At six of eight hotels, one page answered three to five of the eight questions and was where the most runs there read their answer. At the three independent hotels among them that page was a facts page: an FAQ, a stay-information page and a terms page. At the three brand hotels it was the hotel's own page on the brand site, which counts as a front page. At two of the six, another page tied for the most runs: a facilities page at one and an FAQ at the other. A shared answering page at six hotels runs against the hypothesis and toward AgentReady's pattern, a Signal.

### Routes compared with software sites

The table sets hotel routes against AgentReady's; its answering-page row counts the same 208 runs that read the fact on a page. Fetch-only runs compare most closely, since AgentReady does not say whether its agents ran JavaScript.

| Measure | AgentReady, software sites | Hotels, this experiment | Comparability |
| --- | --- | --- | --- |
| First page fetched | Homepage, 92% of 708 reaching it | Front page, 310 of 326 reaching it (95%) | Closely |
| Next page | Docs, 59% of 708 | Facts page, 96 of 326 (29%), 66 going no further; fetch-only, 55 of the 142 reaching it | Partly: classes differ |
| Answering page | Docs, 394 of 844 grounded answers (47%) | Facts page, 112 of 208 read on a page (54%); front page 52 | Direction only |

Table 3. Agent routes on AgentReady's 25 software sites and on the eight hotel sites. AgentReady's figures count runs.

AgentReady's figures, from its [specification](https://github.com/agentready-org/standard/blob/main/spec/spec.md), count runs, and six of its 25 sites supplied 691 of its 1,033. Its grounded answer, one traced to a page the agent fetched, is a looser test than ours, and no model is shared, so the table shows direction and rough size, not exact percentages.

Routes were short, a median of 3 pages, and most often ended on a facts page, as on software sites. The step between differed: hotel agents went next to a facts page far less often than software-site agents went to documentation, even fetch-only ones. The answering page transfers and the route does not, a Signal comparison that cannot rule out the models.

### Blocks and JavaScript

From the datacentre, most blocks the probes met, fetches refused or answered with a challenge page, came from one hotel. That hotel, a brand hotel in New Zealand, challenged all three probe clients on every page, 117 of 117 fetches, on two dates, so from the datacentre a hotel refusing every controlled client on every page is Confirmed. From a residential New Zealand connection, the same hotel let the plain fetches through but still challenged the headless browser, 39 of 39. Swapping the announced name changed the outcome on one page only, so blocks depended on network and kind of client, almost never the name, a Signal.

A price question leads to the booking entry. To the plain fetch announcing Claude-User, the booking entry was readable only at the independent hotel on Cloudbeds, a challenge page at four hotels and an empty shell, a success with almost no visible text, at three. No client read a stay total, the whole stay's price with currency, from any booking entry. Failing the plain fetch at seven of eight hotels, across four booking systems, is Confirmed. The hotel that challenged every client sent its challenges with a 403, the code for access refused. Yet the three other hotels whose booking engines challenged sent 19 of their 24 challenges with a 200, the code for success, so a check trusting the code would pass them.

A page can be empty for other reasons, so JavaScript-only content counted only where the plain fetch got an empty shell and the headless browser got the fact. Four pages met that test, all on one brand hotel's Amadeus booking engine, where the fact was the street address. On reference pages all three clients read the same 42 of the 59 probed facts, the browser no more than plain fetches. The failure exists, Confirmed on one hotel's booking pages and one fact, but on content pages JavaScript was not the obstacle.

The hypothesis also expected runs to fail where pages were not delivered. Runs succeeded in 249 of 268 cases where the Claude-User fetch got the fact's reference page normally, against 37 of 60 on a challenge and 8 of 24 on an empty shell. Of the 75 failed runs, 39 were on challenged or empty pages, about half, tangled with topic: empty shells were mostly room prices, challenges mostly at the hotel that challenged every client.

### Room price

Scoring marks an accurate total for another room wrong, so we hand-checked all 47 room-price runs for a stay total from the hotel's engine. None of the 24 fetch-only runs read one. All eight engines took dates in the address, but agents guessed the parameter names wrongly, and dated entries showed no total until a room was chosen. Six of the 23 browser runs read one, each after clicking into a cart or checkout, a Signal that only clicking through reaches it. The scored six differ, also counting a correct “no price published” reply at the hotel that publishes no price and an estimate built from another hotel's “from” rates, so only four runs are in both.

### Blocked or unpublished

Although one hotel challenged every probe client, 34 of its 48 runs succeeded. Of those, 24 were read on no page, 6 on the hotel's own pages, all by the one browser pair that got through, and 4 on Google Maps or a travel-club listing. The same hotel also had 5 of the 20 wrong answers, the most of any hotel. With one hotel, whether the block caused these unread answers and errors is a Hypothesis.

An unpublished fact takes the hotel's pages away in another way, since no value can come from the hotel. Of 21 answers on the four unpublished facts, 13 gave a value, and 10 of those said the hotel's site does not state it or could not be read. Most agents filled the silence from elsewhere, usually saying so.

### The findings at a glance

| Finding | Key figures |
| --- | --- |
| Agents mostly answered, often without reading the hotel's page | 306 of 381; 91 of 294 unread |
| A facts page answered most, but came next less than docs | 112 of 208; next 96 of 326 against docs' 59% |
| Both failures exist; outside the hotel that challenged every client, at booking engines | Booking entry, 7 of 8 hotels; one hotel, 117 of 117 |
| Only browser agents clicking through read a stay total | 6 of 23 browser, 0 of 24 fetch-only |
| Agents still answered at the fully challenged hotel, mostly reading no page | 34 of 48; 24 unread |

4 · What the results mean

## What a check of one hotel's site should look at

We expected both failures on hotel pages agents use, and agents answering mostly from a few of the hotel's own pages. The first holds, but narrowly: both failures exist, Confirmed, yet apart from the hotel that challenged every client, they sat at the booking engine. The second holds for answering and short routes, though the answering page varied less than we bet: at six hotels one page answered most often. The expectation that runs fail where pages were not delivered fits about half the failures, 39 of 75, tangled with topic. None of the four outcomes named in advance occurred.

Among the beliefs retired is that a correct answer shows the hotel's site worked, since search often stood in. A Signal comparison with no model shared also retires the belief that software-site paths describe hotels.

A check of one hotel's site would start where failures concentrated: the booking entry, read with a plain fetch and a headless browser. It would count any challenge page whatever its status code. Next, on Signal findings, it would ask whether the facts page states each stay fact in visible text, where agents that read the hotel most often found the answer. Booking-engine failures are the vendor's to fix, so advice to hotels covers their own site.

Where a check would start

The booking entry, then the facts page

Read the booking entry with a plain fetch and a headless browser, counting any challenge page whatever its status code. Then ask whether the facts page states each stay fact in visible text.

What stays unknown is mainly how common these failures are. A check needs only a failure that exists and recurs, as the full block at one hotel did, so the unknowns limit what it can claim: it reports what that hotel's pages delivered, never what agents will answer.

We will continue to pursue clarity into this space and are looking to widen the scope of research. Token budgets are affecting the entire industry and we must make do with what we have. In the near future we will have more compute and resources to thoroughly question our own conclusions and other avenues of enquiry. Stay tuned.

5 · References

## 10 sources

1. Vercel with the Ora.ai research lab. Make your site readable by AI agents. 28 Aug 2026. Accessed 8 Sep 2026. [vercel.com/kb/guide/make-your-site-readable-by-ai-agents](https://vercel.com/kb/guide/make-your-site-readable-by-ai-agents)
2. AgentReady, Vercel with the Ora.ai research lab. Agent runs and fetchability dataset README, CC BY 4.0, data from June to August 2026. Date not shown. Accessed 9 Sep 2026. [github.com/agentready-org/standard/blob/main/data/README.md](https://github.com/agentready-org/standard/blob/main/data/README.md)
3. AgentReady, Vercel with the Ora.ai research lab. Specification v1.0. Undated living page. Accessed 9 Sep 2026; re-read 1 Oct 2026. [github.com/agentready-org/standard/blob/main/spec/spec.md](https://github.com/agentready-org/standard/blob/main/spec/spec.md)
4. AuraScope. A hotel's take on agent readiness. Concluded 24 Sep 2026. Read 1 Oct 2026. [research.aurascope.co/research/hotels-take-on-agent-readiness](https://research.aurascope.co/research/hotels-take-on-agent-readiness)
5. Cloudbeds. How AI recommends hotels. Publication date not shown. Accessed 9 Sep 2026. [cloudbeds.com/hotel-ai-recommendations/data/](https://www.cloudbeds.com/hotel-ai-recommendations/data/)
6. TypeSafe. System One. Undated. Accessed 22 Sep 2026. [docs.typesafe.ai/concepts/system-one.md](https://docs.typesafe.ai/concepts/system-one.md)
7. Vercel. eve agent framework, source of the published package version 0.52.5; the run was pinned at 0.51.1. Undated. Read in September 2026; exact date not recorded. [github.com/vercel/eve](https://github.com/vercel/eve)
8. Vercel Labs. agent-browser, npm package; the run was pinned at 0.31.1. Undated. Read in September 2026; exact date not recorded. [npmjs.com/package/agent-browser](https://www.npmjs.com/package/agent-browser)
9. Anthropic. Agent SDK overview, Claude Code Docs; the run used the npm package @anthropic-ai/claude-agent-sdk pinned at 0.3.268. Undated living page. Accessed 2 Oct 2026. [code.claude.com/docs/en/agent-sdk/overview](https://code.claude.com/docs/en/agent-sdk/overview)
10. exe.dev. Hosted worker machines, Sydney region. Undated. Used 24 to 28 Sep 2026. [exe.dev](https://exe.dev/)

Key terms

## In plain language

- **AgentReady** Vercel and Ora.ai's study: controlled probes on one documentation site they built, plus 1,033 agent runs on 25 software sites.
- **Answer key** The 64 answers fixed from each hotel's own pages before agents ran; eight were corrected after verdicts were seen.
- **Booking entry** The booking engine's search page with the test stay in its address: two adults, two nights from 10 Nov 2026.
- **Challenge page** A page security software serves instead of the content, asking the visitor to prove it is a person.
- **Empty shell** A page served as a success with under 200 characters of visible text: the sign of content needing JavaScript.
- **Plain fetch, headless browser** A plain fetch downloads a page's HTML without running its JavaScript; a headless browser, a screenless browser, runs it.
- **Fetch-only and browser tool sets** Fetch-only: each harness's own fetch and web-search tools. Browser: a real Chrome browser the agent drives, without search.
- **Front page** Where an agent starts: an independent hotel's homepage, or a brand site's property page or brand root.
- **Facts page** An FAQ, policies, terms or stay page, plus other pages holding a verified fact except rooms or contact.
- **Read on a page** The judge found the fact in the text of a page the agent received; search results never count.
- **The judge** TypeSafe's System One model, which answers one narrow question with a probability and never fetches pages.
- **Strength labels** Signal: a pattern at 3 or more of the 8 hotels, or one credible outside source. Hypothesis: fewer. Confirmed: existence bar met, a case seen twice, or two sources agreeing.
