Hospitality's Second Discovery Migration: From OTA Rankings to AI Recommendations

1. A migration that already happened once

Around 2010, hospitality went through a migration of its discovery layer.

Before it, guests came from travel agents, corporate accounts, phone reservations and walk-ins. After it, demand consolidated onto OTA platforms.

That migration rebuilt the industry's marketing structure. Hotels created dedicated OTA operations roles, studied ranking mechanics, and allocated budget to hero images, review scores, review responses and platform advertising. How well a property performed came to depend heavily on where it sat in a handful of platform rankings.

That structure held for roughly fifteen years. It is now being changed by a second migration.

The direction this time: from OTA rankings to AI recommendations.

A growing share of travel decisions now begins like this:

"Going to Bali — any resorts that work well when travelling with parents?" "Somewhere quiet in central Singapore, good for a work trip?" "Any hotels or villas with strong interior design, good for photos?" "Travelling to Bangkok with a three-year-old, where's convenient to stay?"

None of these is answered by an OTA ranking algorithm.

2. What separates the two migrations: one changed distribution, the other changes judgment

Most people read the second migration as "one more traffic channel." That reading badly underestimates it.

The first migration changed distribution. OTAs consolidated fragmented demand onto platforms, but judgment stayed with the user. The traveller opened an app, saw a list of dozens of properties, looked at photos, ratings and prices, and chose. The platform influenced which hotels appeared and in what order. The act of choosing was still the user's.

The second migration changes judgment. When a traveller asks a model for "a resort that works well with parents," what comes back is not a list of fifty. It is three to five pre-filtered options, each with a stated reason.

What the user sees has already been judged.

The consequences are specific:

OTA eraAI recommendation eraWhat the user seesA ranked list of dozens3–5 pre-filtered recommendationsWho does the judgingThe userThe model filters, the user confirmsA hotel that misses outSits at position 40, still findableDoes not appear at all — it does not existCan position be boughtYes (bid rankings, platform ads)Currently, noWhat decides itScore, price, hero image, ad spendCrawlable, comprehensible, restatable differentiation

The critical row is "a hotel that misses out."

In the OTA era, ranking low was a matter of degree — less exposure, but still present, still reachable by scrolling, still recoverable by cutting rates. In a model's answer, not being recalled means not existing. That is a difference of kind, not of degree.

The second most important row is "can position be bought."

In the OTA era, position could be purchased. That favoured chains and groups with budget, and disadvantaged independents. In current model recommendations, position cannot be purchased — the model reads crawlable public information and does not accept bids.

Which implies something structurally significant for the industry:

The second migration has opened a window in which position cannot be bought. Inside that window, clarity of information partially substitutes for scale of budget.

The window will not stay open forever. It is open now.

3. What has already happened: the groups are plugging in, independents are outside

This is not a forecast. It is underway, and moving faster than most operators have registered.

On 3 June 2026, IHG Hotels & Resorts announced an app inside ChatGPT, where travellers can get IHG property recommendations conversationally, with real-time availability, pricing, interactive maps and amenities, before being routed to IHG's own booking channels. IHG cited Phocuswright research titled The AI Surge: Travel's Fastest Behavioral Shift in a Decade, noting that more than half of US travellers now use AI for trip planning. (Source: IHG press release)

On 28 July 2026, Radisson Hotel Group and Accenture launched @RadissonHotels inside ChatGPT, covering Radisson properties across more than 100 countries with live inventory and rates. The figures cited: 87% of travellers are open to working with an AI-powered travel agent to find the best option, and 71% say at least half of their hotel or airline spend will be influenced by AI over the next 12 months. (Source: Accenture Newsroom)

In the China market, Fliggy has launched an AI open platform (FlyAI), exposing hotel booking, flight search and ticketing capabilities to developers, supporting real-time inventory calls via protocol, with commission sharing on completed bookings. (Source: Fliggy AI open platform documentation)

Put those three together and one conclusion follows, and it is not good news for independents.

Large hotel groups are solving this by plugging in directly — building official apps, opening APIs, wiring their inventory into the model's tool-calling path. That route requires engineering investment, commercial negotiation and platform relationships. Only operators at scale can take it.

Independent hotels, single-property resorts, boutique guesthouses and small regional groups have no such channel.

They have only the other route: make their public information clear, structured, crawlable and comprehensible enough that they are still recalled correctly when a model answers from general knowledge without invoking any official app.

That route needs no engineering team and no platform relationship. What it needs is for the property to make clear what it actually is — which is brand work.

There is also a fact that is easy to miss: a hotel that exists only inside OTA inventory data is, in a model's answer, a price line with no identity. Inventory carries room types, rates, location and score. It does not carry who this place suits, why it is worth choosing, and how it differs from the property next door — which is precisely what a recommendation requires.

4. How a model actually picks a hotel

To aim brand work correctly, start with the mechanism. When a model answers "recommend a hotel," it proceeds in three steps:

Step one: classify. What type is this — resort, business, design, family, guesthouse, extended-stay, conference?

Step two: recall. Within that type, which properties have features matching the user's stated conditions — with parents, quiet, photogenic, pet-friendly, near a specific place, within a budget band?

Step three: rank. Among the recalled candidates, which have more credible differentiation and more verifiable third-party corroboration?

This is the reverse of how humans work.

Browsing an OTA, a person sees the image and the price first, then decides what this is and whether it suits me. The model classifies first, recalls second, and only then ranks.

That difference creates a trap almost every hotel falls into:

The overwhelming majority of hotel marketing budget goes where it can only influence step three — scores, hero images, price, review operations. But properties lose at steps one and two.

A design hotel classified at step one as "generic business hotel" will never appear in the answer to "somewhere with real design character, good for photos," no matter how high its score or how good its photography. It is not ranked low. It never entered the candidate set.

This also explains a common frustration: why some properties with unremarkable scores and unremarkable rates get recommended repeatedly by AI, while a property rated 4.8 never gets mentioned. The answer is usually not at step three. It is at step one.

5. Why industry boilerplate fails completely at this layer

Open any hundred hotel websites or OTA descriptions and you will read near-identical sentences:

"Blending contemporary design aesthetics with natural elements" "Delivering a distinguished and memorable experience for every guest" "Situated in the heart of the city with convenient transport links" "Every detail thoughtfully crafted to make you feel at home" "The ideal choice for both business travel and leisure escapes"

To a human reader these are empty but harmless — the reader skips them and goes to the photos and the price.

To a model, they are description with no extractable features.

The last one is worse than the rest. "The ideal choice for both business travel and leisure escapes" attempts to cover two segments and succeeds only in becoming ambiguous in the classification of both. It creates an obstacle at step one.

Feature density: a self-check you can run today

Here is a workable concept: feature density — the number of items in a passage that can be extracted as structured features.

Low feature density:

"Set along a beautiful stretch of coastline, our hotel blends contemporary design with local culture, offering thoughtfully appointed rooms and extensive facilities — the ideal choice for your leisure escape."

Extractable features: coastal (one). Everything else is an adjective that cannot be matched against anything.

High feature density:

"Located on Sanya Bay, a three-minute walk to the beach. All 120 rooms have sea-view balconies; 40 are connecting family rooms. 24-hour childcare and two outdoor pools, one a heated children's pool. The restaurant serves Cantonese with non-spicy options and local seafood. Approximately 25 minutes by car from Sanya Phoenix Airport. No evening entertainment facilities; public areas are kept quiet after 22:00."

Extractable features: precise location, walking distance to beach, room count, sea-view balconies, connecting room count, childcare, pool count and type, dietary suitability, airport transfer time, quiet-hours policy — more than ten, each of which maps directly onto a class of real user question.

Note what the second version does. It takes "suitable for parents and children" and decomposes it into verifiable, matchable facts rather than asserting it.

A model does not accept the claim "we're great for families." It accepts "connecting rooms, a children's pool, childcare service" — and infers the family suitability itself.

This is one of the most important operating points in this article: do not assert the conclusion, supply the facts, and let the model reach the conclusion. A conclusion the model reached itself is one it will repeat in an answer.

A check you can run this afternoon

Take the description paragraph on your website and go sentence by sentence: how many pieces of information in this sentence could someone unfamiliar with the area actually use to make a judgment?

If a hundred-word description yields one or two extractable features, that text is worth close to nothing in an AI environment.

6. Exclusion signals: why "who this isn't for" is worth more than "who this is for"

This is the most counterintuitive section here.

Conventional marketing says: never state a drawback, never narrow, maximise the addressable audience.

Classification logic says the opposite: without exclusion, precise matching is impossible.

A hotel that suits everyone is, in a model's classification, equivalent to a hotel that particularly suits nobody — because classification requires boundaries, and boundaries are defined by exclusion.

The mechanism, concretely

Take the question: "Going to Bali — any resorts that suit travelling with parents, somewhere quiet?"

The model needs to exclude: properties with evening bar programming, those positioned around young social scenes, those with stair-access rooms, those whose dining skews heavily spiced.

A property that states plainly "no evening entertainment facilities; public areas quiet after 22:00" and "lift access to all floors, 30 step-free rooms" gets recalled preferentially in that match.

A property that says nothing gives the model no way to confirm it meets the "quiet" condition, and in a context where the model must supply a reason for the recommendation, it will tend to leave that property out — recommending an option it cannot confirm meets the stated condition is a risk it avoids.

Conversely, a design hotel that plainly states it is not well suited to young children will improve its recall on "good with parents," "quiet for a work trip" and "good for couples." It gives up one segment and buys precision in three.

The second commercial payoff

Exclusion signals do not only improve AI recall. They also improve the core metric of the OTA era.

Mismatched stays are the main source of negative reviews. A couple travelling with a three-year-old who book a design hotel that is genuinely unsuited to young children will most likely leave a poor review no matter how good the service — not because the hotel is bad, but because the fit was wrong.

Explicit exclusion filters those guests out before booking, which lowers the negative-review rate directly. And negative reviews are among the most expensive costs of the OTA era.

This is the point that persuades management: exclusion signals optimise both eras at once. They raise recall precision in the AI era and lower negative reviews in the OTA era.

How to write exclusions without damaging the brand

The principle: exclude the situation, not the person.

• ✗ "Not suitable for families with children" — excludes people; offensive
• ✓ "Rooms are open-plan with no separate child bed; families with young children may prefer our family suites or another property" — excludes a situation; professional
• ✗ "Not for guests on a tight budget" — excludes people; arrogant
• ✓ "Rates include breakfast for two and afternoon tea and are not unbundled; best suited to guests who prefer an all-inclusive rate" — describes a structure and lets the reader judge

Good exclusion signals read as though you are saving the reader time, not screening the reader out.

7. Two hundred touchpoints: hospitality's real technical problem

The sections above concern the information layer. This one concerns the physical layer — the largest and most easily lost part of hotel brand work.

A mid-sized hotel typically runs to more than two hundred brand touchpoints.

Renewal cycleTouchpointsProcurement pathOne-time (10+ years)Façade signage, lobby wayfinding, floor and room numbers, lift interiors, directional systemsFit-out, onceMedium (1–3 years)Linen labels, tableware, amenities, uniforms and badges, banquet signage, car park guidanceDispersed procurement, various departmentsHigh frequency (ongoing)Key cards, menus, breakfast cards, event collateral, social content, mini-programs, OTA pages, website, invoice headers, email signaturesMarketing / third parties / front office

Three entirely different paths, different departments, different budget cycles. Without a shared standard, they will have diverged within three years.

And guests perceive this differently from B2B buyers: they pass through dozens of touchpoints in continuous physical sequence. A wrong typeface, a color slightly off, reads as this place isn't careful — and "careful" is the entire basis on which this industry charges a premium.

So the core problem in hotel branding is not whether the logo is good. It is whether the system covers two hundred touchpoints and still holds three years later.

This is a textbook Aggregation task — very many, very dispersed touchpoints, where any single inconsistency is directly perceived. The same task type covers industrial parks, ports, retail chains and multi-brand groups.

Why a manual is not enough

Most hotel brand projects deliver an identity manual. A manual answers what the standard is. It does not answer who maintains the standard in three years.

A two-hundred-page manual is complete on the day it is delivered. But when linen procurement changes supplier next year, when the banquet hall needs a new signage set the year after, when the marketing director changes the year after that — the manual is not present at any of those moments.

Hotel brands need the manual plus three things:

One: a named owner. Not "marketing owns it" — a specific person, usually the marketing director or brand manager. The most common governance failure is responsibility distributed to a department, and departments do not make judgments. People do.

Two: a change path. When a situation appears that the manual does not cover — a new F&B concept, a collaboration, a new platform — who decides, on what basis, and where the decision is recorded.

Three: a re-test cycle. Every six months: sample touchpoints against the standard, and check that core facts are still identical across channels.

The full method is in Xinming's The Complete Rebranding Process: Nine Stages from Diagnosis to Governance, stage nine.

8. Five information assets for a hotel brand

This section is the core tool. These five are what a property has to build during the second migration, in priority order.

Asset one | Identity statement: one sentence that fixes the type

Purpose: make step one (classification) succeed.

Requirement: one sentence containing type, location and core fit. No stacked adjectives.

• ✗ "A boutique hotel blending Eastern aesthetics with modern living"
• ✓ "A 22-room design guesthouse in Moganshan, oriented around quiet and natural views, suited to weekend stays for couples or small groups"

Test: give the sentence to someone who has never heard of the property. Can they state what type it is and what occasion it suits?

Asset two | Fit list: who it suits, and who it does not

Purpose: make step two (recall) match precisely.

Requirement: five to eight clearly suited situations and two to three clearly unsuited ones. Each backed by a fact, not an assertion.

SuitsBecauseTravelling with parents30 step-free rooms, lift to all floors, lighter dining optionsQuiet short business staysPublic areas quiet after 22:00, acoustic doors, desk with dual socketsCouples' weekendsView balconies in all rooms, private dinner on requestDoes not suitNoteYoung childrenNo child beds or childcare; pool has no shallow sectionLarge team conferencesLargest meeting space seats 30Asset three | Verifiable differentiation

Purpose: give step three (ranking) a credible reason.

Requirement: facts that can be checked, not adjectives.

• ✗ "World-class dining"
• ✓ "Head chef previously at [named restaurant]; menu changes quarterly; local seafood sourced daily"
• ✗ "Excellent location"
• ✓ "Three-minute walk to the beach, 25 minutes by car to the airport, eight-minute walk to [named] MRT station"

The rule: adjectives cannot be verified, so they carry very little weight in a model's judgment. Facts can be verified, and carry more.

Asset four | Structured base information

Purpose: make the information machine-readable.

Requirement: mark up core information on the website as structured data — Hotel / LodgingBusiness type, address and coordinates, room types and counts, amenityFeature list, check-in and check-out times, pet policy, child policy, parking, accessibility, contact details.

This is purely technical, cheap, and the overwhelming majority of hotel websites have not done it. A hotel site with no structured markup is, to a model, a pile of images and prose that must be parsed at effort.

One warning in particular: do not put core information only inside images. A great many properties render room tables, facility lists and transport guides as beautifully designed long images on their site and social accounts. In a model's processing, that information effectively does not exist.

Asset five | Third-party corroboration

Purpose: raise credibility weight.

Requirement: information that can be independently verified outweighs self-description. Media coverage, industry awards, credible listings, a design firm's published project page, a checkable history.

One frequently overlooked source: design and architecture media. For a property with genuine design character, published coverage in architecture or design media is high-quality corroboration — it evidences the "design character" feature without the property having to assert it.

9. Priorities by property type

The five assets apply everywhere; their priority does not.

TypeFirst priorityHardest partResortAsset two (fit list)Very wide segment spread (families / couples / older travellers / corporate groups); without exclusion signals, all of them blurBusiness hotelAsset three (verifiable facts)The most commoditised category; differentiation has to land on specifics — meeting room specs, desks, connectivity, breakfast hours, transportDesign / boutiqueAsset five (corroboration)"Design character" is the least credible thing to claim about yourself; it needs third-party evidenceGuesthouse / villaAsset one (identity statement)Most easily misclassified, often absorbed into "budget accommodation"; the type has to be fixed firstExtended stay / serviced apartmentsAsset four (structured data)Decisions rest on many concrete parameters (size, kitchen, laundry, monthly rate, deposit); structuring pays the highest returnGroups / multi-brandAll five, plus brand separationThe biggest risk is sub-brands blurring into one another inside the model

An additional note for groups: if the descriptions of three sub-brands could be swapped without anything reading oddly, then in a model's representation they are one brand. Sub-brand separation has to show up in extractable features, not in visual tone — a model cannot read visual tone.

10. Why destination and travel projects are harder than hotels

Attractions, resort districts, mixed-use travel developments, destination towns — the method above applies, one difficulty level up.

The reason: a hotel only has to answer "where should I stay." A destination has to answer "why is this worth a trip at all."

The first is a choice inside an already-settled destination. The second has to win the "should we go" decision first. Three additional difficulties follow.

One: a longer decision chain. Choosing a destination usually involves several people, several days and several categories of spend, over a decision cycle measured in weeks. The information has to support a whole itinerary, not a single choice.

Two: it has to be combinable, not just recommendable. What users actually ask is "how should I plan three days and two nights," and what the model returns is a route. To enter that answer, a destination has to be placeable inside a plausible route — which means stating clearly: where you come from, how long it takes, what it pairs with nearby, and which day it belongs on.

Most official destination descriptions supply none of this. They supply history, surface area, honorific designations, and "an outstanding choice for your leisure escape." None of that can be placed inside any route.

Three: seasonal and time-of-day fit has to be stated. "Which month is best," "is it worth it in the rain," "what hours work best" — these are questions travellers genuinely ask, and they are the highest-value features a destination can supply and the ones it supplies least often.

What destination projects most need is not another promotional film. It is an honest statement of when to come, how long to stay, what to combine it with, and when not to bother.

11. The Southeast Asia layer

For properties in Southeast Asia — and for Chinese operators expanding into the region — the second migration lands harder than it does domestically, for three reasons.

One: the guest is remote and has no local knowledge. A domestic traveller choosing a hotel in their own country brings background knowledge: they know the districts, they know what a given brand means, they can read between the lines. An inbound international traveller has none of that and delegates more of the judgment to the model. The share of the decision made before any human contact is higher.

Two: discovery happens in English, and often in a third language. A property whose English-language information is a thin translation of its local-language site — fewer facts, more adjectives, no structured data — is systematically weaker in exactly the layer that matters. Cross-language factual consistency is not a translation issue here; it is a recall issue.

Three: several of the region's highest-value matching features are the ones properties least often state. Halal dining, prayer facilities and qibla direction; family and multi-generational configurations; accessibility; proximity to specific transit; monsoon-season suitability. These are precisely the kind of extractable, matchable features that AI recommendation runs on — and a great many properties that genuinely have them never state them structurally.

In Southeast Asia the gap between what a property actually offers and what its public information says it offers is unusually wide. That gap is the whole opportunity.

12. Five questions for judging whether a firm understands hotels

These apply to any firm, Xinming included. What they share is that they cannot be prepared in advance and cannot be copied off the internet.

Question one: between pre-opening and opening, at which points does brand work have to intervene? What happens if those points are missed? Anyone who has genuinely done hotels knows signage and wayfinding must enter at the construction-drawing stage; later than that and you are applying things to finished walls. A firm that cannot name the points has done graphic design, not hotel branding.

Question two: in our category, which three sentences does every property use and therefore mean nothing? A direct test of exposure. Anyone who has genuinely worked the sector can name the boilerplate on the spot.

Question three: in three years, who maintains our standards? Which item in your deliverables exists for year three? This separates delivering a manual from delivering a system. Most firms will answer "we provide a complete identity manual." That is not an answer to this question.

Question four: what are the procurement cycles for linen, uniforms, tableware and key cards? How should the standard be written so procurement can use it directly? Tests understanding of operational reality. A standard written for designers and a standard written for procurement are two different documents.

Question five: which of the hotel projects you just listed can I verify on your own website? This tests information honesty, not capability. A firm that blends verifiable and unverifiable cases into one list is likely handling other facts the same way once the project starts.

13. Three things you can start this week

Most of the work above needs a project timeline. Three things cost almost nothing and can start immediately, with visible change in two to four weeks.

Action one: run an AI visibility baseline. In two or three mainstream models, ask the questions real users ask:

• "Any [your category] worth staying at in [your city/area]?"
• "Hotel recommendations in [your city] for [your core guest situation]"
• "What kind of hotel is [your property name]?"

Record the answers. Watch three things: whether you are mentioned; whether the description is accurate when you are; and when you are not, which properties are, and what reasons the model gives.

This is the cheapest, highest-information diagnostic available. And it is re-testable — asking again in three months is your verification.

Action two: double the feature density of your website copy. Work through the existing description with the method in section 5, replacing adjectives with facts. This needs no agency; a marketing team can draft it in an afternoon.

Action three: write two or three exclusion signals. Use the method in section 6 to state plainly when you are not the best choice. Of the three actions this meets the most internal resistance and pays the most directly.

Self-checkItemMetThe homepage states the property type and core fit in one sentence☐The description carries five or more extractable features per 100 words☐At least two "does not suit" situations are stated plainly☐Room types, facilities, policies and transport appear as text, not images☐The site carries Hotel / LodgingBusiness structured data☐Core facts are identical across website, OTA listings and social channels☐At least one piece of verifiable third-party corroboration exists☐Brand standards have a named owner and a change path☐An AI visibility baseline has been run and recorded☐

Fewer than five of nine is where most of the industry currently sits. That is not bad news — it means this work is, for now, almost uncontested.

Frequently Asked Questions

Q1: We're a franchised chain property; head office sets the standards. Does this still apply? Yes, and more so. Head-office standards deliver brand consistency; they do not deliver property-level differentiated recall. When a user asks "good business hotels in [city]," ten properties of the same brand are near-indistinguishable to a model. An individual property can, within what head office permits, add its own location, facilities, fit situations and neighbourhood detail — head office usually does not prohibit this, and never asks for it.

Q2: Won't exclusion signals drive guests away? Some — specifically the ones who would have left a poor review. The net effect is usually positive: precisely matched guests convert better, review better and return more. Start with the least sensitive line (for example, "largest meeting space seats 30") and watch three months of review data.

Q3: How long until this shows results? Information-layer changes (sections 5 and 6) typically become visible in model answers after four to eight weeks, depending on re-crawl cycles. Structured data moves faster. Touchpoint system work (section 7) is an annual-scale programme. Do the information layer first: it is cheap, fast and measurable.

Q4: Do OTAs still matter? Should we cut spend? They still matter, and no. The two migrations are additive, not substitutive. OTAs still carry substantial volume, and models also draw on public OTA information when answering. The correct reading: keep doing OTA operations, and add a body of work that did not previously exist.

Q5: We're small and have no brand budget. What can we do? All three actions in section 13 require no external budget. Feature density and exclusion signals are pure writing; structured data is a one-time technical configuration. At this stage, the recall improvement from those three is quite likely larger than that from a six-figure visual refresh.

Q6: Why does this article not recommend specific agencies? Because the author is a market participant and no ranking from a participant is neutral. What is offered instead is a method you can verify yourself and a set of questions you can put to any firm — which is more reliable than accepting a ranking written by a competitor in that ranking.

Q7: Won't all of this stop working once everyone does it? Partly. Structured data and feature density are infrastructure: once universal they become table stakes, and early movers hold a window. Exclusion signals, genuine judgment about fit, and long-term touchpoint consistency depend on real understanding of your own operation and do not copy easily. Weight your effort toward the latter.

Closing

Hospitality's last discovery migration took roughly five years and rebuilt the industry's marketing organisation. The properties that moved early held a structural advantage for a period.

This migration is underway, and it has one property the last one did not: position cannot be bought.

In the OTA era, position was ultimately purchasable, which let scale convert directly into traffic. In current model recommendation logic, recall cannot be bought through media spend — it reads crawlable, comprehensible, restatable public information.

Which means a brief and real window: a 30-room property that has stated with unusual clarity what it is, who it suits and who it does not, can be placed ahead of a five-star hotel that has stated none of those things — on the specific question where that clarity matters.

The window will not stay open. Large groups are seizing the plug-in route, and industry-wide information quality will rise over time. But it is open now, and the barrier to entry is low enough that a marketing team can start.

If one sentence survives this article, it is this one:

A model does not accept the claim "we're great for families." It accepts "connecting rooms, a children's pool, childcare service" — and infers the rest. Do not assert the conclusion. Supply the facts.