The Disappearing Hotel

Version note (version 2, August 2026). This version adds five first-party measurements collected by the author on 28 July 2026, reported in the new Sections 8, 9, and 10: an AI-assistant prompt-volume estimate for “boutique hotel” in the US and German markets, a citation-source breakdown drawn from 874 tracked ChatGPT citations for the same query, a test of which sources ChatGPT cites when naming AI-visibility providers, a classical Google search-volume reading for the German provider category, and a visibility audit of 49 competitor domains. Sections 9 and 10 carry the German-market material that appears in the German edition of this article, so that both editions now rest on the same body of first-party evidence. The original June 2026 synthesis, its argument, and its cited literature are otherwise unchanged; the Abstract and the Limitations section have been updated to reflect the additions, and the former Sections 9 and 10 are now Sections 11 and 12.

1. Introduction

For two decades, a hotel’s commercial visibility was governed by two systems it could partly influence: search engine results and OTA marketplaces. A new intermediary is now inserting itself ahead of both. Travelers are beginning to plan trips by asking an AI assistant a question and accepting a short, synthesized answer rather than scanning a page of links. This shift matters because of a structural difference in the output: a search engine returns many results and lets the user choose, whereas a generative engine returns a curated few and makes the choice on the user’s behalf. The consideration set shrinks from a page to a sentence.

The consequence for hotels is stark. If a property is not named in that synthesized answer, it is not merely ranked lower; it is functionally absent from the traveler’s awareness. This article examines the evidence for that claim, the mechanisms that drive which hotels appear, the specific exposure of independent and design-led properties, and GEO as the emerging strategic response. It treats GEO as a nascent and unsettled discipline rather than a proven solution.

2. Background: from SEO and OTAs to conversational discovery

The dominant pre-AI architecture of hotel discovery rested on search engine optimization (SEO) and OTA intermediation. Roughly 80% of global travelers used OTAs such as Booking.com and Expedia in the lead-up to a trip (APH, 2025), and hotels competed for organic and paid placement on Google. That system rewarded link-based authority, keyword relevance, and OTA listing quality.

Generative engines operate on a different principle. Aggarwal et al. (2024) formalize a generative engine (GE) as a system that uses generative models to “gather and summarize information to answer user queries,” synthesizing across multiple retrieved sources rather than returning them as a list. ChatGPT, Google’s AI Overviews and AI Mode, Gemini, and Perplexity all fit this description. The commercial migration into travel is now explicit: in late 2025, OpenAI began integrating third-party apps into ChatGPT conversations, with Booking.com and Expedia among the early pilot partners, allowing users to obtain travel results without leaving the chat; Tripadvisor later joined and also partnered with Perplexity (Statista, 2026).

The user-behavior effect of synthesis is measurable. A Pew Research Center study tracking the browsing of 900 U.S. adults across 68,879 Google searches in March 2025 found that when an AI summary appeared, users clicked a result link only 8% of the time, versus 15% without a summary, and clicked links within the AI summary just 1% of the time (Pew Research Center, 2025). This “zero-click” dynamic is the mechanism by which visibility, not ranking, becomes the decisive variable.

3. The scale of the shift

Adoption figures should be read with care: they come largely from commissioned industry surveys with differing samples and definitions, not from a common instrument. With that caveat, the direction is consistent across providers.

Two patterns emerge. First, AI has become a mainstream discovery and planning tool: Booking.com’s Global AI Sentiment Report, based on 37,325 respondents across 33 markets surveyed in April–May 2025, found that 67% of travelers already use AI in some aspect of their travels and that AI assistants (24%) are now trusted more than travel bloggers (19%) or social media influencers (14%) (Booking.com, 2025). Second, a pronounced trust gap persists at the point of transaction: Expedia Group’s survey of more than 5,700 adults in the U.S., U.K., and India found that while 53% are comfortable letting AI suggest options, 66% would not trust an AI assistant to book on their behalf, and only 8% are comfortable booking through an AI platform (Expedia Group, 2026).

The practical implication is that AI is reshaping the top of the funnel, inspiration and shortlisting, more than the booking click itself. That is precisely the stage at which a property either enters or never enters the consideration set. The macro stakes are large: the World Travel & Tourism Council estimated travel and tourism’s 2025 global GDP contribution at a record US$11.6 trillion, about 9.8% of the global economy (WTTC, cited in Hospitality Net, 2026). Forward-looking claims, for example, Gartner’s projection of a 25% decline in traditional search volume by 2026 (cited in ZipTie, 2026), are forecasts, not measured outcomes, and should be treated as such.

4. The visibility problem: most properties are invisible

The clearest evidence that generative discovery concentrates visibility comes from prompt-based audits. These are vendor studies rather than peer-reviewed work, and their methodologies and incentives warrant scrutiny; nonetheless their core findings are mutually corroborating.

Lighthouse, a hospitality intelligence platform, ran 4,545 distinct ChatGPT prompts across nine global destinations and five traveler personas. Across those prompts ChatGPT named hotels 49,707 times, but those mentions resolved to only 2,721 unique properties, the remainder were repeats of a small recommended core (Lighthouse, 2026). The study reported “market coverage rates”, the share of a market’s hotels ever surfaced, of about 10% in Tokyo and 13% in Paris, with roughly two-thirds of hotels never mentioned even in a small market such as Park City. The 100 most-mentioned hotels worldwide accounted for over 13% of all mentions (Lighthouse, 2026).

A separate multi-platform study, conducted by LuxDirect across 25 London luxury hotels on six AI platforms (ChatGPT, Claude, Gemini, Grok, Perplexity, and Google AI Mode) using 2,700 standardized queries, found that five hotels captured 57% of all AI recommendations (LuxDirect, cited in Hospitality Net, 2026). The concentration is therefore observable across engines, not unique to one model.

Crucially, the bias is toward scale and luxury rather than guest satisfaction. Lighthouse reported only a weak correlation between a hotel’s Booking.com guest-review score and its ChatGPT visibility; star rating was a far stronger predictor (Lighthouse, 2026). Even on generic prompts, “I need a hotel in X,” with no budget or preference stated, four- and five-star properties dominated, and three-star hotels were nearly absent. Business-travel prompts returned 83% four- and five-star recommendations and family prompts 73%, despite mid-tier hotels being appropriate for many such stays (Lighthouse, 2026). On the chain side, more than a quarter of branded U.S. hotel mentions went to Marriott, and the top three brand families accounted for 54% of branded mentions (Lighthouse, 2026).

This concentration is not merely an industry observation; parts of the selection mechanism have been reproduced under experimental control. An algorithm audit using a randomized conjoint design across twelve open-weight and proprietary models found that guest rating and price dominate recommendation probability (a top rating raises selection by 31.6 percentage points, a high price lowers it by 30.0), and that list position, a content-free artifact, shifts recommendations causally, worth roughly $12 per night (Baig et al., 2026, preprint). The audit did not support a “brand-as-risk-reduction” heuristic: chain affiliation carried a small penalty of 1.8 percentage points, and management response, a cue the optimization industry actively promotes, had no detectable effect. The audit therefore qualifies rather than confirms the big-brand reading of the vendor studies. What it demonstrates is that recommendation is driven by reputation signals and by content-free interface artifacts, not that chain membership as such is rewarded. The shared conclusion across audit and vendor evidence is narrower but still sharp: invisibility is the default state for most hotels, and a property must actively earn its way out of it.

5. Mechanisms: how generative engines select, trust, and cite

Understanding the bias requires understanding how generative engines assemble an answer. Three mechanisms are central.

Source provenance. Generative engines ground recommendations in retrieved third-party content. Lighthouse found that 82% of the sources ChatGPT cross-referenced when forming hotel recommendations fell into two categories: OTA and metasearch platforms (Booking.com, Expedia, and similar) and editorial or media sites (Forbes, Lonely Planet, Condé Nast Traveler, and equivalents) (Lighthouse, 2026). A hotel’s presence in “best of” lists and editorial features therefore functions as a distribution input, not merely brand-building. The same dynamic applies to Google AI Overviews, where Pew found Wikipedia, YouTube, and Reddit among the most-cited sources overall (Pew Research Center, 2025).

Content characteristics rewarded by GEO experiments. Aggarwal et al. (2024) tested nine content modifications against a generative-engine benchmark (GEO-bench) and found that adding verifiable statistics, incorporating credible quotations, and citing reliable sources produced the largest visibility gains, on the order of 30–40% on a position-adjusted word-count metric, while improving fluency and readability yielded smaller but significant gains of roughly 15–30%. Notably, keyword stuffing, the legacy SEO tactic, produced negligible or negative effects, and a merely “authoritative” tone did not help (Aggarwal et al., 2024). This implies generative engines reward content that is credible, evidenced, and clearly written, not content that is merely keyword-dense.

Entity and structured signals. Because engines retrieve and reconcile structured data about entities, consistent and machine-readable property information (across one’s own site, OTA listings, and authoritative directories) reduces ambiguity about what a property is. Lighthouse explicitly advises auditing OTA and metasearch listings because these are “what AI uses to verify what it already knows,” and stale or inconsistent listings degrade the input (Lighthouse, 2026). The convergence of these mechanisms explains the chain-and-luxury skew: large brands already possess dense editorial coverage, voluminous structured data, and consistent cross-platform listings.

6. Why independent and design-led properties are disproportionately exposed

The exposure of independent and boutique hotels follows directly from the mechanisms above. Each of the signals that generative engines weight most heavily is one in which large chains hold a structural advantage.

First, editorial and “best of” authority is unevenly distributed. Established luxury brands accumulate press coverage as a function of marketing budget and longevity; a newly opened design-led property may be architecturally distinctive yet thinly covered. Second, structured-data and listing consistency are easier to maintain at scale, where brands deploy central distribution teams; an independent owner-operator may have inconsistent representations across channels. Third, the engines’ demonstrated star-rating and chain heuristics directly penalize the three- and four-star independent segment even when guest satisfaction is high, recall that review scores correlated only weakly with visibility (Lighthouse, 2026).

The evidence does not, however, support fatalism. The London study found that a 26-room independent property appeared more frequently than a 174-room chain hotel of the same quality tier in the same city, and Paris was the one audited market where independents collectively exceeded chains in share of mentions (LuxDirect, cited in Hospitality Net, 2026; Lighthouse, 2026). Visibility is patterned and concentrated, but it is not strictly a function of size, which is the opening that GEO seeks to exploit.

This segment is also commercially material. Estimates of the independent and boutique market vary widely by vendor and definition, the global boutique hotel market alone was valued at roughly US$26.7 billion in 2024 (Grand View Research, 2025), and such figures should be read as indicative rather than precise. The relevant point is qualitative: the properties most exposed to AI invisibility are concentrated in exactly the experiential, design-led category that travelers increasingly say they want.

7. Generative Engine Optimization (GEO)

Definition and origin. GEO is the practice of optimizing content to increase its visibility in generative-engine responses, as distinct from SEO’s goal of ranking in link-based results. The term and the first systematic framework originate with Aggarwal, Murahari, Rajpurohit, Kalyan, Narasimhan, and Deshpande, whose paper “GEO: Generative Engine Optimization” was first posted to arXiv in November 2023 and published at the ACM SIGKDD Conference on Knowledge Discovery and Data Mining (KDD ‘24) in 2024 (Aggarwal et al., 2024). The authors introduced GEO-bench, a benchmark of diverse queries with associated sources, and a black-box optimization framework for defining and improving visibility metrics.

Core tactics and measured effect. The paper’s headline finding is that GEO methods can boost visibility “by up to 40%” in generative-engine responses, with efficacy varying by domain (Aggarwal et al., 2024). The most effective tested tactics were Cite Sources, Statistics Addition, and Quotation Addition; fluency and readability improvements helped moderately; keyword stuffing did not. A finding of particular relevance to independents is that lower-ranked sources benefited more from GEO than already-dominant ones, suggesting the technique can partially offset incumbency. One later attempt to reproduce the tactic ranking outside the original benchmark reported relative improvements of +40% (Cite Sources), +37% (Statistics Addition), and +22% (Quotation Addition), with keyword stuffing at +3% and not statistically significant (Oruesagasti, 2026, preprint). That report should be weighted with the same caution applied to the vendor studies above: it is a technical report produced by a commercial research division, not peer-reviewed work, and it measures UK iGaming brands rather than hotels. Its value here is that the ordering of tactics matches the original experiment in a different domain, not that it independently validates the effect size.

Translation to hospitality. Applied to hotels, GEO implies several concrete moves, each grounded in the mechanisms of Section 5. Lighthouse’s practitioner recommendations align closely with the academic findings: treat PR and editorial coverage as a distribution channel; audit OTA and metasearch listings for accuracy and consistency; align on-site and listing language with the target guest segment, because that language teaches the engine which travelers to route to a property; and measure current AI visibility directly rather than assuming it (Lighthouse, 2026). The throughline is that descriptive content has become “algorithm input,” not only marketing copy.

What the platform itself says. One primary source is worth setting against the vendor guidance. Google’s own documentation for generative search features states that there are “no additional requirements” for appearing in AI Overviews or AI Mode, because those features run on the same core systems as classical search (Google Search Central, published 15 May 2026, updated 10 July 2026). Read plainly, that removes a shortcut and confirms a direction: there is no separate switch to flip, and the signals that make a property legible to classical search are the same ones the generative features draw on. It also means a hotel with a retrieval problem, meaning pages that these systems cannot read cleanly, has a prerequisite to fix before any of the tactics above can matter.

8. First-party evidence: measuring the mechanism directly

Every statistic cited to this point originates from someone else’s study. That is a defensible way to synthesize a field, but it leaves the report itself uncited: nothing here is a number that exists nowhere else. This section reports three first-party measurements collected by the author on 28 July 2026, using DataForSEO’s AI-optimization API endpoints and live ChatGPT sessions, to test whether the demand and citation patterns described above are also observable directly.

The first measurement estimates how often “boutique hotel” appears inside prompts submitted to AI assistants. DataForSEO’s ai_optimization/keyword_data/search_volume endpoint returned, on 28 July 2026, a current monthly estimate of 16,706 for the US/English market, with a twelve-month range of 16,700 to 28,100 per month, and 5,541 for the German market, with a twelve-month range of 5,500 to 7,600 per month (Frenzel, 2026, first-party measurement via DataForSEO). A figure like this is not a count of logged prompts; no vendor discloses the contents of user conversations at that scale. It is a model, built the same way keyword-search-volume tools have long modeled Google query volume: from a mixture of signals correlated with actual usage, calibrated against whatever ground truth the vendor can access, and expressed as a best estimate rather than a census. The same caveat applies to every “searches per month” figure in this article and in the broader industry literature it draws on; it is stated here explicitly because this is the first number in the article for which the author, rather than a cited source, is responsible. The estimate is still informative for the reason ordinary keyword-volume estimates have been informative in SEO for two decades: relative magnitude and trend direction survive the modeling even when the absolute number should be read loosely. Read that way, the contrast is the finding: in the same US dataset and on the same date, “hotel marketing” returned an estimated 679 monthly prompts, and “generative engine optimization” returned 17; in the German dataset, “generative engine optimization” returned 3 and “geo agentur” returned 5 (Frenzel, 2026, first-party measurement via DataForSEO). By this estimate, travelers ask AI assistants about boutique hotels roughly three orders of magnitude more often than anyone asks about the discipline meant to make those hotels visible to them, in both markets, on the same day, a demand-side imbalance the vendor-side literature synthesized in Sections 2 through 4 does not itself quantify.

The second measurement asks a narrower question: when ChatGPT does answer a “boutique hotel” query, which domains does it cite? DataForSEO’s llm_mentions/top_domains endpoint tracked 874 ChatGPT citations for the query “boutique hotel” in the US/English market, question scope, collected 28 July 2026 (Frenzel, 2026, first-party measurement via DataForSEO). The ten most-cited source domains were Expedia (125 citations), Tripadvisor (103), English Wikipedia (85), Merriam-Webster (24), Hospitalitynet (20), Guestpoint (20), Reddit (19), Booking.com (18), Cvent (15), and Roommaster (11); the domains most visible in the underlying search results were Tripadvisor (160), Expedia (87), and Hosteltur (76), alongside Afar, Timeout, and Oyster. Read together, the list is an OTA, an encyclopedia, a dictionary, two hospitality trade publications, a forum, and two property-management-system vendors. No hotel-marketing agency appears. No hotel, independent or chain, appears as a cited source in its own right. This is an independent, differently instrumented confirmation of the pattern Lighthouse already reported at a coarser grain, that 82% of the sources ChatGPT cross-references when forming hotel recommendations are OTA and metasearch platforms plus editorial and media sites (Lighthouse, 2026; Section 5); the measurement reported here supplies the specific domain distribution behind that two-category figure. The consequence for a hotel is direct: a property’s citation surface for this query is not its own website. It is the OTA listing, the Wikipedia entry if one exists, the trade-press mention, and the Reddit thread, and a hotel that treats only its own site as the audience for its content is optimizing a surface the engine barely cites.

The third measurement is a special case and needs to be read as one. Where the first two measurements concern the demand for, and citation of, boutique hotels, the third concerns a different population entirely: providers of AI-visibility and GEO services, the market this article’s author competes in. It is reported here not as evidence about hotel invisibility, which it cannot supply, but as evidence about the selection mechanism, which it can, because the mechanism a generative engine uses to decide who gets named is the same mechanism regardless of who is being named.

Asked, with web search enabled, which consultants or agencies a boutique hotel should hire to get recommended by AI assistants, ChatGPT (gpt-4o-2024-08-06) named 15 providers on 28 July 2026 (Frenzel, 2026, first-party measurement). Four of the fifteen were drawn from a single vendor-published comparison list, on which the list’s own author had ranked itself first; three were cited from Wikipedia pages about the provider; the remaining eight were cited from the provider’s own homepage, in each case a page whose title or description restated the buyer’s question back to them. Asked the equivalent question in German, ChatGPT (gpt-5.5-2026-04-23) named 12 providers on the same day. Eleven of the twelve were cited from the provider’s own site with a title matching the query, including two domains branded to a single German city that carry negligible independent authority by any conventional measure of referring domains, ranking keywords, or estimated traffic. What those eleven pages share is the shape of the title: it restates the buyer’s question almost verbatim, on the pattern “place or service, colon, does ChatGPT know your company?” Provider names are withheld here: the finding lies in the pattern, and a single-run test supports no claim about any individual firm. The only hospitality-specific mention in the German test was not cited from the provider’s own marketing but from a PDF published by the hospitality trade association HSMAI, hosted on global.hsmai.org.

The model’s own search behavior during the German run is worth recording alongside the answer. Its logged fan-out queries included “hospitality generative engine optimization agency hotels” and “Hotel AI Search Optimization ChatGPT Perplexity agency Germany,” which is precisely the hospitality-specific positioning a GEO provider for hotels would be expected to occupy. The model looked for it and settled on no provider that could be cited from a German hospitality page. That is a stronger observation than the ranking itself: the demand-side query existed inside the model’s own retrieval, and the supply side had nothing citable to meet it.

Read against Section 5, the pattern is consistent rather than novel, and that consistency is the point: self-published comparison lists, encyclopedia entries, homepages whose titles echo the buyer’s exact question, and association-published documents were the four entry points this particular selection mechanism rewarded, on this day, for this population. A hotel, or a hotel-GEO practice, pursuing visibility for the same class of query should expect the same four doors, not a fifth one specific to hospitality. The limits of this measurement should be stated as plainly as its finding. Each test is a single run of a stochastic system, not a repeated sample; the English and German prompts were answered by different underlying models (gpt-4o-2024-08-06 and gpt-5.5-2026-04-23 respectively), so differences between the two answers are confounded with the model difference and cannot be read as an English-versus-German effect; and the measurement covers ChatGPT only, Perplexity, Gemini, and Google AI Overviews were not tested. The finding should therefore be read as a mechanism demonstration, not a benchmark.

9. The German-speaking market, measured on two channels

The English-language literature on generative travel discovery is now reasonably broad: prompt audits, citation analyses, one peer-reviewed experiment on content optimization. For the German-speaking market almost none of it exists. That absence is itself a finding, and it is documented in Section 10; the underlying audit report records it in one line, “Nobody publishes measurement in German.” This section reports what the two demand channels looked like in that market on 28 July 2026, and it is the material that appears in the German edition of this article.

AI-prompt demand. The prompt-volume endpoint described in Section 8 returned, for the German-language market, an estimated 5,541 monthly prompts containing “boutique hotel,” with a twelve-month range of 5,500 to 7,600. In the US/English market the same term returned an estimated 16,706 on the same day, range 16,700 to 28,100 (Frenzel, 2026, first-party measurement via DataForSEO). The German-speaking market therefore asks about boutique hotels roughly a third as often as the US/English market, which is unremarkable given relative market size. The ratio inside the German market is the part worth reading. On the same day and in the same dataset, “hotelmarketing” returned an estimated 7 prompts per month, “geo agentur” 5 (a term DataForSEO records as first appearing in measurable quantity in March 2026), “generative engine optimization” 3, “ki sichtbarkeit” 2 (first appearing in April 2026), and “geo optimierung” 1. Travelers ask AI assistants about the product, the hotel, by orders of magnitude more often than anyone asks about the service meant to make that hotel visible.

That gap should be read carefully. It does not establish that GEO providers are superfluous. Nobody asks an assistant for a service whose existence they do not yet know about, which is a familiar signature of an early-stage category rather than a verdict on its usefulness.

Classical search demand. A second dataset supports that reading directly. Google search demand, collected on the same day through DataForSEO Labs (keyword_overview endpoint, market Germany), shows a different picture from the prompt side: “geo agentur” at 1,000 searches per month with a keyword difficulty of 6 and 48% year-on-year growth; “geo optimierung” also at 1,000 per month, growing 171%; “generative engine optimization” at 1,300 per month; and “ki suchmaschinenoptimierung” at 90 per month, growing 180% year on year (Frenzel, 2026, first-party measurement via DataForSEO). Classical search for the category is already growing sharply while AI-prompt demand for the identical terms remains in the single digits.

Holding both readings together is the finding of this section. The German-speaking market is already researching GEO on Google and has barely adopted the newer query form of asking an assistant about it. Both channels point the same way at very different speeds, and a provider becoming visible in this field today is building authority for a demand pattern that is measurably arriving in each of them.

10. The supply side: four counter-checks across 49 competitor domains

Alongside the measurements above, the visibility audit underlying this article surveyed 49 competitor domains active or visible in the German and English hospitality-GEO field, collected in July 2026 through the author’s own page captures. Four counter-checks drawn from it describe the German-speaking supply side.

First, of eight German GEO agencies examined, not one mentions hospitality anywhere on its own website. These providers sell to the Mittelstand, to B2B buyers, to insurers, not to hotels.

Second, of five English-language hospitality-GEO specialists examined, four show no German or European signal on their own site at all; the fifth carries a single passing mention. None of them could answer a German hotelier in German.

Third, of eight German hotel agencies examined, three already sell GEO as a named service, yet none was named by any of the assistants tested, and in classical Google search on the topic they reached page two at best, with no published methodology.

Fourth, and this is the summary finding: of all 49 domains captured, exactly one publishes first-party research of any kind, and that one publishes in English. None of the captured sites shows a reproducible test or an open measurement protocol.

These four observations should be read as a market snapshot with a collection date rather than as a judgment on individual providers. They describe a gap on 28 July 2026, not a durable property of the market, and competitor names are withheld because the pattern across all 49 is the object of interest, not any single firm.

11. Limitations, measurement challenges, and open questions

GEO should be presented as an emerging, not settled, discipline, for several reasons.

Evidence quality is uneven. The foundational experimental result (Aggarwal et al., 2024) is peer-reviewed, but much of the hospitality-specific evidence comes from vendor studies whose methods are not fully transparent and whose authors sell related products (Lighthouse, 2026; LuxDirect, cited in Hospitality Net, 2026). Their findings are mutually corroborating and broadly consistent with controlled audits, which lends them credibility, but they are not independent peer-reviewed research and may carry commercial incentives.

The target is non-stationary. Generative engines change frequently and opaquely. Recommendation patterns differ across models, one audit noted Perplexity skews toward luxury brands while other engines diverge (Hotelrank, 2026, vendor analysis), and a tactic that works on one engine or version may not generalize or persist. Unlike Google’s relatively legible ranking signals, the criteria driving AI recommendations are difficult to audit, so hotels often cannot determine why they do or do not appear (LuxDirect, cited in Hospitality Net, 2026).

Measurement is unsettled. There is no standard, the way “search rank” was standardized. Visibility is being operationalized variously as share of voice, market coverage, mention counts, and citation share, and these metrics are not directly comparable across studies. Outcome attribution is harder still, because the booking may occur on a later channel.

Manipulation and integrity risks. The same research line that established GEO has documented adversarial methods, prompt injection, content “hijacking,” and poisoning attacks, designed to manipulate generative-engine outputs (e.g., methods discussed in Aggarwal et al., 2024, and subsequent work). This raises the prospect of an arms race and of platform countermeasures that could neutralize current tactics, as happened with manipulative SEO.

First-party measurements carry parallel limits. The measurements added in Sections 8, 9, and 10 are estimates and single observations, not censuses or repeated trials. The prompt-volume figures are modeled estimates of AI-assistant query volume, not counts of logged prompts, and should be read with the same caution applied to any keyword-volume tool throughout this article. The classical search-volume figures for the German provider category carry the same modeling caveat, and the year-on-year growth rates are computed on small absolute bases, which magnifies them. The provider-omission test was run once per language, on one assistant, ChatGPT; Perplexity, Gemini, and Google AI Overviews were not tested. The English and German runs used different underlying models (gpt-4o-2024-08-06 and gpt-5.5-2026-04-23), so any apparent difference between the English and German results is confounded with the model difference and is not interpretable as a language or market effect. The 49-domain audit is a set of page captures assessed by the author against a fixed rubric, not a blind or independently coded survey, and it records what a site published at the moment of capture rather than what a provider is capable of. All five measurements were collected on a single day, 28 July 2026, and generative engines are known to change frequently; the figures describe that day, not a stable state.

What the German material does not establish. Section 9 measures demand and Section 10 measures supply; neither shows that individual German hotels are demonstrably invisible today, because the omission test examines service providers rather than hotels. What it demonstrates is the selection mechanism, which operates the same way regardless of who is being named. The low AI-prompt volume for provider terms likewise does not show that the service is unnecessary, since the classical search demand for the same terms was measured growing at the same moment. And none of these patterns can be assumed durable: a pattern observed on one day may shift within months.

Open questions therefore include: How durable are GEO gains as models update? Do visibility gains convert to bookings and revenue, given the documented trust gap at the point of purchase (Expedia Group, 2026)? Will engines move toward transactional integration (the Booking.com/Expedia ChatGPT pilots) in ways that re-advantage large OTA-partnered brands? How will the ratio between classical search demand and AI-prompt demand in the German-speaking market develop over the coming twelve months, given that Section 9 found the two growing at very different speeds? And how will regulators treat algorithmic gatekeeping in discovery, an issue already surfacing in publisher antitrust actions over AI Overviews (ALM Corp, 2026)?

12. Conclusion and practical implications

The shift from search to generative discovery changes the unit of competition from rank to mention. The evidence assembled here, survey data showing AI is now a mainstream planning tool, behavioral data showing AI summaries suppress click-through, and audit data showing recommendations concentrate on a small core of large and luxury properties, points to a consistent conclusion: most hotels are currently invisible to AI recommendation, and the properties most exposed are independent, boutique, and design-led ones whose third-party authority signals are thin relative to chains.

For hospitality decision-makers, several implications follow, each appropriately hedged. The window for early action appears genuinely open but is unlikely to remain so, because machine-learning systems tend to reinforce existing recommendation patterns (Lighthouse, 2026). Practitioners can reasonably act now on the lowest-regret measures, which align across academic and industry evidence: ensuring listing and structured-data accuracy across channels; pursuing credible editorial and “best of” coverage as a distribution input; writing property descriptions in clear, evidenced, segment-aligned language; and establishing direct measurement of AI visibility against a competitive set. These steps are defensible even if specific GEO tactics prove transient, because they improve the underlying signals on which any future engine is likely to rely. What practitioners should not do is treat a 40% visibility figure, drawn from a 2024 benchmark in non-travel domains, as a guaranteed hospitality outcome. GEO is a promising and partially validated response to a real and measurable problem, but it is an early-stage discipline operating against a moving, opaque target, and it should be resourced as an experiment with measurement attached, not as a solved channel.

Citable findings, collected

1. “Boutique hotel” produced an estimated 5,541 AI prompts per month in the German-speaking market on 28 July 2026, with a twelve-month range of 5,500 to 7,600 (first-party measurement, DataForSEO, ai_optimization/keyword_data/search_volume, 28 July 2026).
2. In the US/English market the same term stood at an estimated 16,706 AI prompts per month on the same day, twelve-month range 16,700 to 28,100 (first-party measurement, DataForSEO, 28 July 2026).
3. On the same day, in the same German dataset, “geo agentur” returned an estimated 5 AI prompts per month, “generative engine optimization” 3, and “geo optimierung” 1, against 5,541 for “boutique hotel” (first-party measurement, DataForSEO, 28 July 2026).
4. Classical Google search for “geo optimierung” in Germany grew 171% year on year to 1,000 searches per month, while AI-prompt demand for the same term stood at an estimated 1 per month (first-party measurement, DataForSEO, 28 July 2026).
5. Asked in the role of an independent German boutique hotelier, gpt-5.5 named twelve GEO providers on 28 July 2026; eleven were cited from the provider’s own site, under a title that restated the question almost verbatim (first-party measurement, 28 July 2026).
6. The only hospitality-specific mention in the German omission test came not from a provider’s own site but from a PDF published by the trade association HSMAI (first-party measurement, 28 July 2026).
7. Of 874 tracked ChatGPT citations for “boutique hotel” in the US/English market, the ten most-cited domains were two OTAs, an encyclopedia, a dictionary, two hospitality trade publications, a forum, and two property-management-system vendors. No hotel and no hotel-marketing agency appeared (first-party measurement, DataForSEO, 28 July 2026).
8. Of 49 competitor domains captured in the underlying visibility audit, exactly one publishes first-party research, and that one publishes only in English (visibility audit, victorfrenzel.com, collected July 2026).
9. A prompt-based test by Lighthouse found only 2,721 distinct hotels named across 4,545 ChatGPT prompts, out of 49,707 total mentions (Lighthouse, 2026).

How to cite this article

Frenzel, V. (2026). The Disappearing Hotel: Generative Discovery, AI Visibility, and the Case for Generative Engine Optimization in Hospitality (Version 2, August 2026). Retrieved from https://victorfrenzel.com/research/the-disappearing-hotel/

A copy of the June 2026 version is archived on ResearchGate; the version on this page is the current one. A German-language edition with its own first-party German-market data is published separately at Das verschwindende Hotel. The author’s ResearchGate profile lists all archived work.

References

ALM Corp. (2026, February 18). Google’s AI Overviews and publisher traffic: How antitrust filing reveals 58% click decline and the breakdown of the web ecosystem. Retrieved June 17, 2026, from https://almcorp.com/blog/google-ai-overviews-publisher-traffic-decline-antitrust-lawsuit-analysis/

Aggarwal, P., Murahari, V., Rajpurohit, T., Kalyan, A., Narasimhan, K., & Deshpande, A. (2024). GEO: Generative engine optimization. In Proceedings of the 30th ACM SIGKDD Conference on Knowledge Discovery and Data Mining (KDD ‘24) (pp. 5–16). Association for Computing Machinery. https://doi.org/10.1145/3637528.3671900 (Preprint: arXiv:2311.09735, posted November 16, 2023.)

APH. (2025, December 18). Travel booking statistics, 2025 report and trends. Airport Parking and Hotels. Retrieved June 17, 2026, from https://www.aph.com/community/holidays/travel-booking-statistics/

Baig, M. S. A., Gillani, S. A., & Ali, A. (2026, June 16). Whose hotel does the AI recommend? An algorithm audit of reputation signals in LLM-assisted hotel selection [Preprint]. arXiv:2606.16344. Retrieved June 17, 2026, from https://arxiv.org/html/2606.16344

Booking.com. (2025, July 23). Booking.com releases the Global AI Sentiment Report. Retrieved June 17, 2026, from https://news.booking.com/bookingcom-releases-the-global-ai-sentiment-report/

Expedia Group. (2026, April). AI Trust Gap report [Survey of 5,700+ adults in the U.K., U.S., and India]. As reported in PhocusWire (2026, April 14), Losing control, data privacy top concerns around AI travel booking. Retrieved June 17, 2026, from https://www.phocuswire.com/news/online/ai-trust-discovery-planning-expedia-dune7

Frenzel, V. (2026, July 28). First-party AI-visibility measurements: prompt-volume estimate, classical search-volume reading, citation-source distribution, provider-omission test, and 49-domain competitor audit [Data collection]. Collected via DataForSEO (ai_optimization/keyword_data/search_volume; ai_optimization/llm_mentions/top_domains; ai_optimization/chat_gpt/llm_responses/live; DataForSEO Labs keyword_overview, market Germany), live ChatGPT sessions (gpt-4o-2024-08-06; gpt-5.5-2026-04-23), and the author’s own page captures of 49 competitor domains (July 2026). Raw responses, provider response IDs, and capture records held by the author.

Grand View Research. (2025). Boutique hotel market size & share | Industry report, 2030. Retrieved June 17, 2026, from https://www.grandviewresearch.com/industry-analysis/boutique-hotel-market-report

Google. (2026, May 15, updated July 10). Optimizing your website for generative AI features on Google Search. Google Search Central. Retrieved July 28, 2026, from https://developers.google.com/search/docs/fundamentals/ai-optimization-guide

Hospitality Net. (2026, April 16). Travelers use AI to dream, trusted brands to book; travel hit $11.6 trillion in 2025; SiteMinder connects 53,000 hotels to AI platforms. Retrieved June 17, 2026, from https://www.hospitalitynet.org/editorial/4131925/

Hotelrank. (2026, January 27). AI hotel landscape 2026. Retrieved June 17, 2026, from https://hotelrank.ai/research/ai-hotel-landscape-2026

Lighthouse. (2026, June 12). The AI recommendation is the new battleground for hotel distribution [Research presented at Luminate 2026 by B. Reiter]. Hospitality Net. Retrieved June 17, 2026, from https://www.hospitalitynet.org/news/4132964/

LuxDirect. (2026). CS7 London study: Five London hotels capture 57% of AI recommendations. As reported in Hospitality Net (2026, March 16). Retrieved June 17, 2026, from https://www.hospitalitynet.org/opinion/4131422/

Oruesagasti, J. (2026, March 5). Algorithmic trust and compliance: Benchmarking brand notability for UK iGaming entities in generative search engines [Preprint, technical report produced by Interamplify Research Division, UK]. arXiv:2603.12282. Retrieved June 17, 2026, from https://arxiv.org/pdf/2603.12282

Pew Research Center. (2025, July 22). Google users are less likely to click on links when an AI summary appears in the results. Retrieved June 17, 2026, from https://www.pewresearch.org/short-reads/2025/07/22/google-users-are-less-likely-to-click-on-links-when-an-ai-summary-appears-in-the-results/

PhocusWire. (2026, April 14). Losing control, data privacy top concerns around AI travel booking [Reporting Phocuswright U.S. traveler research]. Retrieved June 17, 2026, from https://www.phocuswire.com/news/online/ai-trust-discovery-planning-expedia-dune7

Statista. (2026). Artificial intelligence (AI) use in travel and tourism, Statistics & facts. Retrieved June 17, 2026, from https://www.statista.com/topics/10887/artificial-intelligence-ai-use-in-travel-and-tourism/

ZipTie. (2026, May 16). Click behavior in zero-click search: Why rankings no longer predict traffic (and what does). Retrieved June 17, 2026, from https://ziptie.dev/blog/click-behavior-in-zero-click-search/

Author’s note on sourcing. Statistics are attributed to the organization that produced the underlying data wherever possible, with the reporting outlet named separately when the primary release was not directly retrieved. Several hospitality-specific figures derive from vendor studies (Lighthouse, LuxDirect, Hotelrank) rather than peer-reviewed research; these are corroborated where possible against the controlled algorithm audit (Baig et al., 2026) and the foundational GEO experiment (Aggarwal et al., 2024). Oruesagasti (2026) is likewise a commercial technical report rather than peer-reviewed work and is cited as a directional corroboration of tactic ordering, not as independent validation. The two arXiv preprints were previously cited here under incorrect first authors (Ulrich; Roumeliotis); both were verified against the arXiv records on 26 August 2026 and corrected. On 27 August 2026 a substantive correction followed in Section 4: the audit by Baig et al. was previously reported as finding that chain affiliation increases recommendation probability. It reports the opposite. “Brand-as-risk-reduction” was the authors’ hypothesis, not their result; chain affiliation carried a penalty of 1.8 percentage points, and management response had no detectable effect. The passage and the surrounding argument were rewritten against the source PDF. Forecasts are labeled as such and distinguished from measured results. The measurements reported in Sections 8, 9, and 10 are first-party work by the author rather than citations of external research; their raw API responses, response identifiers, and page captures are retained and available on request.