What an AI Visibility Audit Is
An AI Visibility Audit is a dated record of what AI answer engines return when a traveler asks for a hotel by category and intent rather than by name, together with an investigation of which sources are displayed or cited with those answers and which facts in the public record are wrong.
It has three parts. What the engines returned. What the accessible public record contains. What the gap means commercially.
It is not a score, not a dashboard, and not a ranking report. There is no single AI ranking to measure.
A recurring finding in AGR audits is that some properties are not invisible to AI systems at all. In documented captures, a system can describe a property accurately when it is named directly and still omit it when a traveler asks an unbranded category question. AGR calls those two observable outcomes branded retrieval and unbranded discovery. They have different diagnostic implications, and separating them is the work.
Audit method at a glance
- Current standard set: ChatGPT, Gemini, and Google AI Mode
- Fresh, logged-out sessions with no prior conversation
- Single-run captures, first answer recorded, no retries and no selection among attempts
- A property-specific query battery, not a template
- Public-record review conducted separately from the captures
- Source investigation wherever the answers displayed their sources
- Directional, not statistical. Every dated capture held on file and available on request
Table of Contents
Definitions
Five terms run through this page and through every audit. They are not interchangeable.
Branded retrieval. Finding and describing a property when a user names it directly. This establishes that a system can produce usable information about the entity when explicitly prompted.
Unbranded discovery. Whether a property appears when an AI system is asked to assemble recommendations for a category, an occasion, or a destination without the property being named. This is the stage at which a traveler’s shortlist forms.
The retrieval layer. The query-time process through which an AI system finds, selects, parses, grounds, cites, or otherwise uses available information to produce a response. Retrieval-layer optimization can improve eligibility, extraction, citation, and answer presentation.
The formation layer. Historical AGR practitioner shorthand for the public source environment in which entities, categories, definitions, relationships, and corroborating references exist and from which AI systems can retrieve, synthesize, and potentially learn. Current AGR doctrine does not treat “formation layer” as a directly observed proprietary model stage. The controllable object is the public source environment; the measurable object is observable AI behavior across repeated queries.
Machine-readable facts file. A structured plain-text record of an entity’s canonical facts, published for automated systems rather than for guests. Useful as a canonical machine-readable reference. Not a discovery mechanism, and not a guarantee of retrieval or citation.
A property can hold strong branded retrieval and no unbranded discovery at the same time. That combination is a recurring reason a well-credentialed hotel is absent from answers its own marketing is built to win.
What the audit answers
Six questions, in the order they matter to whoever owns the number.
Are you being recommended, or only recognized?
Branded retrieval is what happens when a traveler types your name. Unbranded discovery is what happens when a traveler describes a place and a purpose. A property can be described accurately, at length, on every system, and appear in none of the category answers that assemble a shortlist. It is a recurring finding in this work.
Which competitors are taking the recommendations?
The audit names them. Not a competitive set you supplied, the properties the systems actually returned for the queries your property should win. In the published Napa audit, a property with 4,400 square feet of event space took the private-events recommendation from a property with 55,000. The larger venue was not ranked lower. It was not in the answer. AGR classifies that observable pattern as competitive displacement.
What identity do the answers associate with you?
The answers reveal which categories and use cases a system associates with the property, stated in its own words. Business and convention. Golf. Value. Off-beach. Where the associated identity and the marketed identity diverge, the property can fail to enter the answers built around the position it is actively marketing. The audit quotes the system’s framing rather than characterizing it. AGR groups these outcomes as misclassification, where the category is wrong, and mispositioning, where the description is wrong.
Where do the platforms disagree with each other?
Three systems returning three different guest counts, or a ship name that does not exist, is not automatically three unrelated errors. Where the public record itself carries conflicting versions, cross-platform disagreement is one downstream symptom consistent with that. The audit checks whether those conflicts exist rather than assuming every platform failed for the same reason. It is also why the report never pools platforms into a combined figure.
Which sources are displayed or cited with the answers?
Where an answer displays or cites its sources, the audit records them. Correspondence between the answer and a displayed source is observable evidence. It does not establish that the source caused the output.
Where is the public record factually wrong?
Closed restaurants still listed. Lapsed investment terms quoted as current. A room count that disagrees across the property’s own pages and its own structured data. Each gets an entry stating the false claim, the property’s own correct figure, and where the false version appears. A figure is called incorrect only where the property’s own record establishes the correct value.
What we test
Queries are written to approximate realistic traveler and planner questions, framed by place and purpose rather than by property name, with one deliberate exception. The battery is built for the property, not applied from a template. Audits delivered to date have run from five queries on a single-product cruise brand to eighteen across a hotel and its branded residences.
| Query category | What it establishes |
|---|---|
| Broad luxury and tier | Whether the property enters the widest category answer at all |
| Occasion | Whether romance, honeymoon, and anniversary answers surface the property |
| Amenity and program | Whether a real, verifiable asset reaches the answers built around that asset |
| Culinary | Whether the property’s dining credentials surface, and against which competitors |
| Family | Whether family-oriented answers surface the same identity and credentials as luxury-oriented answers |
| Weddings and social events | Whether venue-specific credentials enter planner-oriented recommendations |
| Corporate, meetings, incentive | Whether meeting-space and group credentials surface accurately in planner-oriented answers |
| Geographic and sub-market | Whether the property surfaces for the micro-location it actually occupies |
| Traveler fit | Who should stay here. The one deliberate branded query |
| Competitive comparison | Whether the systems hold a coherent competitive set, and whether it matches yours |
The traveler-fit question has a distinct diagnostic role. If the systems answer it well and fail everything above it, the property holds branded retrieval without unbranded discovery, and the diagnosis changes completely.
Which AI systems we examine
ChatGPT, Gemini, and Google AI Mode form the current standard set. Current audits run all three. The published Estate Yountville audit is an earlier study that used ChatGPT and Gemini; historical audits retain the platform scope recorded with them. Supporting rate, review, and classification data is drawn from the booking and review platforms the answers cite, and from national tourism registries where one exists.
Why platforms are reported separately
Every result in the report belongs to one platform. There is no combined visibility figure, because different systems return different answer sets and source patterns, and a pooled number has no defined denominator. A property present on two systems and absent from the third is a finding. Averaging the three erases it.
Why every capture is a fresh, logged-out session
Each query is entered once, in a new session, with no prior conversation. Each capture records the first answer returned. No run is repeated to find a better or worse result, and no answer is selected from among attempts.
What a single run can and cannot establish
This is a single-run directional audit. It treats presence or absence as a dated observation and looks for patterns across the full battery. It does not measure statistical frequency, and it does not claim that identical text, or even identical inclusion, will reproduce on a subsequent run. Every report states this in its own methodology note. Every dated capture is held on file and available on request.
What we measure
| Measurement | What it establishes |
|---|---|
| Unbranded discovery presence | Whether the property surfaces at all, per query, per platform |
| Recommendation position | Whether it leads, sits mid-list, or appears only as a secondary aside |
| Competitive displacement | Which named properties took the recommendations |
| Category association | Which identity the answers associate with the property, in their words |
| Cross-platform disagreement | Where systems return different facts or different competitive sets |
| Factual accuracy | Every claim the systems make that the property’s own record contradicts |
| Tier and classification signals | Where star class, rating, and rate tier disagree with the actual credential |
| Source and citation pattern | Which domains and source types the displayed answers rest on |
| First-party versus intermediary presence | Whether the property’s own record appears in the source set at all |
| Machine-readable layer | Whether the structured data identifies the property as the entity it is |
Findings are numbered, seven to nine per report. Each leads with the verified data, names the competing properties, and states the source of the language the systems used.
How we diagnose the observed pattern
Measurement without diagnosis is a dashboard. The second half of the audit establishes what the observed pattern is consistent with. That phrasing is deliberate and it holds throughout the report.
The source environment
Where an AI surface displays or cites external sources, those sources become observable evidence. The recurring pattern is that a category answer cites a destination marketing site, a venue directory, or a review aggregator, and the audited property is absent from or weakly represented in that displayed source set. The report records that as a condition consistent with the observed exclusion, not as a proven cause.
OTA-defined identity
Booking platforms are heavily represented in the source environment. Cloudbeds’ vendor-produced 2025 research, based on 810 prompts across ChatGPT, Perplexity, and Gemini, found that online travel agencies accounted for 55.3 percent of AI-generated hotel citations and hotel websites for 13.6 percent. Those descriptions can also remain stale after the property record changes. The published Estate Yountville audit article records booking and directory descriptions that named on-site restaurants that had since closed.
Legacy records still in circulation
New and repositioned properties often carry a thin current public record while older descriptions and pre-opening specifications remain indexed and readable. The audit checks whether those legacy versions are appearing in current answers.
Contradictions inside the property’s own record
Some of the sharpest findings involve no third party at all. A length that differs between the deck plan, the press release, and the site’s own structured data. A room count that disagrees between the corporate page and the national registry. Where sources conflict, both figures are shown and the property’s own current position is stated as the reference.
The machine-readable layer
Structured data that types a yacht line as a generic local business. An image field pointing at a homepage. Alias handling that never connects the short-form name to the entity. A sitemap returning errors. These defects weaken the clarity and consistency of the property’s own machine-readable record. The audit documents them separately from the AI outputs, and does not assume that correcting any single defect will change a recommendation.
Which observable problem class the evidence supports
If a system describes the property accurately when named but repeatedly omits it from relevant category answers, the evidence points away from simple entity absence and toward an association, retrieval, or source-environment problem. If the systems repeatedly return facts contradicted by the property’s authoritative record, the observed representation is wrong. Conflating the two leads to the wrong remediation.
The audit states what the evidence supports and stops there. It does not claim to know how any model weighs a source internally.
What you receive
A PDF. Not a dashboard login, not a slide deck.
| Section | What it contains |
|---|---|
| Cover and scope | Platforms tested, dates, what the audit covers and what it does not |
| Executive summary | Verified credentials, and the sentence stating the core problem |
| AI visibility by query type | One row per query. Result, and the properties that appeared instead |
| Numbered findings | Seven to nine, each traced to a dated capture, each naming competitors and sources |
| Sources displayed with the answers | The source-type table, included whenever the captures displayed sources |
| What a corrected result looks like | Current answer against the position a correct answer would rest on |
| Why this is not an SEO, PR, website, or OTA problem | What search, public relations, the website, and the booking channels each influence, and where the remit of each one stops |
| What this means commercially | Segment exposure, with third-party research cited separately from the captures |
| Methodology note | Platforms, dates, single-run statement, and what was not tested |
Two structural rules govern the document.
Evidence from live captures and evidence from the public record are presented in separate parts and never mixed. One is what the systems said. The other is what the accessible public sources contain. Treating them as the same evidence is how visibility reports overclaim.
Positive and negative findings are reported under the same standard. The audit does not suppress a legitimate win, and it does not manufacture one to balance the report. In the published Estate Yountville audit the property appeared in one query of fifteen, and that one named the property directly.
What comes back depends on what we find
We look at the property before we write anything, and the scope of the report is set by the evidence rather than by a package tier.
The answers are clean. The property surfaces across the major query types, and the systems describe it accurately and consistently with its verified positioning. We tell you that and do not run a report. It costs one email, and it is the honest outcome.
The answers are fine, the record underneath is not. The systems return the property correctly, and the site carries conflicting data across its own pages, defective or absent structured data, contradictory machine-readable facts, or a broken crawl layer. That produces a shorter report focused on the structural findings, because there is no unbranded discovery finding to report.
Both. The property is absent or misdescribed in the answers, and the underlying public record contains conditions consistent with those findings. That is the full audit described on this page.
Report versus audit
A report records the symptom. An audit identifies the condition it is consistent with. The comparison below distinguishes measurement alone from AGR’s diagnostic audit. A product called a report or tracker can also include diagnostic work; its methods and deliverables determine what it establishes.
| AI Visibility Report | AI Visibility Audit | |
|---|---|---|
| Purpose | Measurement | Diagnosis |
| Presence and absence | Yes | Yes |
| Named competitive displacement | Limited | Yes |
| Category association analysis | No | Yes |
| Factual discrepancy register | No | Yes |
| Source investigation | Limited | Yes |
| Machine-readable layer review | No | Yes |
| Corrected-result reference | No | Yes |
| Remediation priorities | No | Yes |
The full distinction is set out at AI Visibility Report vs. AI Visibility Audit.
What this audit does not claim
We do not claim access to model internals
We observe outputs and investigate sources. We do not claim to know how any system weighs a document, and we do not sell a mechanism nobody outside the lab can see.
Google’s guidance on third-party SEO tools, services, and advice, published June 5, 2026, reinforces this boundary. It says good advice qualifies claims as opinion based on data or experience, or supports them with official Google guidance. Third-party tools cannot access Google’s internal ranking data or guarantee performance. AGR’s audit reports observed answers and investigates public sources, with inferences identified as inferences.
We do not claim there is one AI ranking
There is no single ranking to hold and no rate card that buys one. Any offer saying otherwise is describing something that does not exist.
We do not claim causation from correlation
An audit can establish that a property is absent, that a competitor is present, and that a given source was displayed with the answer. It cannot establish that the source caused the absence. The report says consistent with, and means it.
We do not treat one answer as permanent
Generative outputs are non-deterministic, and can also vary with platform, date, session state, location, account context, and retrieval conditions. The audit records the date, the platform, and the exact query, and states plainly that it is a single-run instrument.
We do not guarantee placement
Nobody can. What can be corrected are identifiable errors and inconsistencies in the portions of the public record that can be changed.
We do not publish an adoption statistic we cannot source
No percentage of travelers who now start with AI unless it is named, dated, and traceable to a primary study.
We separate observation from inference
What the systems returned is one body of evidence. What the public record contains is another. They appear in different parts of the report.
Nothing is modeled or projected
Every figure is a capture or a verified fact. No estimates, no forecasts, no illustrative numbers.
The evidence behind the method
This methodology came out of published fieldwork, not a service brainstorm.
The AGR Luxury Hotel AI Visibility Index. 824 ranked hotel recommendations across 180 answers from ChatGPT, Gemini, and Google AI Mode, in six United States luxury markets, captured by hand in a single day, logged out, in fresh private windows, on July 29, 2026. 152 properties were named at least once. 23 accounted for half of all 824 recommendations. In the average market, five properties took half of everything recommended, and in Chicago, Maui, and Napa Valley, four did. The same capture returned five recommendations for Mandarin Oriental, Miami, whose building had been demolished by controlled implosion 108 days earlier. The Index used ten-question submissions and disclosed three Gemini recaptures after clipboard truncation; its fieldwork protocol differs from the current property-audit method above. The Index establishes concentration in those markets on that day. It does not establish why the concentration formed or whether it persists. Read the Index
The Luxury Hotel AI Recommendation Study. AGR followed the Index with a September 8, 2026 study of 148 luxury hotels that had already been recommended at least once. The measured website AI-readiness variables, including lodging schema and llms.txt, showed no detectable association with recommendation frequency. A model containing Forbes Travel Guide rating, Michelin Key count, and market accounted for 54.7 percent of the variance in log recommendation slot count. The study does not test what gets a hotel into the recommendation set and does not establish that credentials cause recommendations. Read the study
A published property audit. The Estate Yountville audit is public in full, including all fifteen queries and every result. The property appeared in one. The fourteen misses covered luxury, boutique, romantic, anniversary, spa, food, weddings, corporate retreats, private events, incentive travel, and meeting venues. Read the audit
Independent research on where the answers come from. Cloudbeds, The Signals Behind Hotel AI Recommendations, 2025. Across 810 prompts on ChatGPT, Perplexity, and Gemini, online travel agencies accounted for 55.3 percent of AI-generated hotel citations and hotel websites for 13.6 percent. This is vendor-produced research, cited as external corroboration of the source-environment pattern.
Independent research on traveler behavior. Cornell Center for Hospitality Research, An Examination of AI in Travel Planning Across Traveler Spending Segments, Young Jang and Christopher Anderson, 2026. The study surveyed 1,029 active U.S. travelers across four spending segments. AI Chatbots and Assistants ranked fourth in the aggregate planning-tool table, behind Search, Reviews and Review Websites, and Official Hotel Websites. Stated comfort was generally highest for discovery and information tasks and lowest for booking activities and tours. The report states that Curacity supported data collection for the study; it does not specify the structure or amount of that support. The same study found concerns about accuracy, transparency, and generic recommendations among the leading reported barriers to AI use in travel planning. These are self-reported attitudes and stated comfort, not observed behavioral telemetry.
Further audits have been conducted for hotels, resorts, branded residences, and ultra-luxury yacht lines. Those are confidential to the addressee and are not published.
What happens after the audit
The audit is the diagnosis. It is not a work order, and not every finding should become a billable intervention.
Correctable directly. Factual errors in the property’s own record. Contradictions between its pages, its structured data, and its corporate parent. Stale booking-platform descriptions. Absent or defective machine-readable identity. Specific, fixable, largely under the property’s control. Retrieval-layer issues are generally the first corrective target when the problem is data availability, schema, listing accuracy, or citation eligibility. If that resolves the absence, the work is done.
Requires source-environment work. Persistent absence from unbranded discovery, an associated identity that contradicts the marketed one, and a thin or contradictory third-party record can justify source-environment remediation. These conditions may sit partly outside the property’s direct control. The relevant discipline is Knowledge Formation Optimization. KFO structures, sequences, distributes, corroborates, and corrects intellectual frameworks and entity definitions across the public information environment and measures whether AI systems reproduce them accurately across relevant queries and over time. It does not edit model parameters and cannot guarantee recommendation inclusion.
Should simply be monitored. Some findings reflect a credential the property does not hold or a position it is not trying to win. The audit says so rather than manufacturing a project.
Who should commission one
Independent luxury hotels and resorts. Ultra-luxury cruise and yacht lines. New openings and pre-launch properties, where the record the systems will read is forming now and can be set rather than corrected. Repositioned properties, where the previous identity is still indexed. Branded residences, where a finite inventory sells once and discovery is a sell-through problem rather than a branding one. Properties carrying heavy OTA dependence. Ownership and asset-management groups assessing discovery exposure across a portfolio.
AGR works with independent luxury hotels, resorts, and cruise lines. The method applies wherever a property has a distinct identity and real direct-booking value. Where a property competes primarily on price, correcting the record has little to work with.
Questions we are asked
Can you guarantee my hotel appears in ChatGPT or Gemini?
No, and nobody can. What we can do is document where the property is absent, misdescribed, or described from sources it does not control, and correct the ones that are correctable.
Is an AI visibility audit just SEO with a new name?
No. SEO primarily addresses discoverability and performance in search systems. The AI Visibility Audit examines observable AI outputs, query-time and retrieval evidence, displayed or cited sources where available, factual conflicts in the public record, and machine-readable entity conditions. These are overlapping but different objects, and the audit does not assume one hidden causal mechanism.
How is this different from a GEO or AEO audit?
GEO and AEO labels do not define a fixed audit scope. AGR’s AI Visibility Audit examines observable answers, displayed or cited sources, and broader public source-environment conditions, then reports which observable problem class the evidence supports. Other services may overlap with that work. Where the finding concerns data availability, extraction, or citation eligibility, retrieval-layer work is the right fix and we say so.
Do I need special AI schema or an llms.txt file?
Structured data matters because it must accurately identify the entity and match what is visible on the page, not because there is a special AI markup. There is none. A machine-readable facts file is useful and is not a discovery mechanism. Anyone selling either as the mechanism is selling markup.
Why does the same question give a different answer each time?
Because generative outputs are non-deterministic, and can also vary with platform, date, session state, location, account context, and retrieval conditions. That is why every capture is a fresh logged-out session with the platform, date, and exact query recorded, and why the report states it is directional rather than statistical.
Do you have results you can show me?
The Estate Yountville audit is published in full with every query and result. The AGR Luxury Hotel AI Visibility Index publishes its findings and method; the underlying capture dataset is available on request. Subsequent property audits are confidential to the addressee.
How is this different from an AI visibility tracker?
A mention count alone does not explain which identity the answers associate with you, which competitor took the category, which sources were displayed, or which facts in the public record are wrong. Those are the questions this audit was built around. Some trackers include that analysis; compare the work delivered rather than the product label.
What do you need to start?
The property name and official website, the market, the guest segments that matter commercially, and the two or three competitors worth benchmarking against.
Request an AI Visibility Audit
Americas Great Resorts. Luxury hospitality demand infrastructure since 1993.
