Reinventing Hotel Loyalty with AI: Moving from Points-Based Programs to Predictive Personalization
Marriott Bonvoy ended 2025 with 271 million members, after adding 43 million in a single year. Ask most general managers whether their own loyalty database is driving incremental, profitable repeat stays and you'll get a pause. CBRE counted more than 675 million members across the five largest hotel programs in 2024, up 14.5%, while program revenue grew 8.3% to $1.2 billion. Members accounted for 52.8% of occupied rooms at those brands, up two points. Membership is growing faster than the money, and Skift Research's November 2025 loyalty report says it plainly: membership is growing, but "true engagement and frequency per member are declining."
The problem is structural. A traditional points program treats every member as a transactional account: stay, earn, redeem, repeat. It says nothing about why a guest is choosing your property, when they are about to disengage, what they would pay for in the gap between bookings, or how their lifetime value is trending. The program rewards stays that would have happened anyway and ignores almost everything else that determines real loyalty: the bad service recovery you never made, the upsell they would have paid for, the cross-property trip they took with a competitor because you never knew they were planning it.
What AI changes is the layer that sits above the points ledger. Predictive personalization swaps demographic segments for a per-guest probability model that keeps updating: likelihood to churn, to upgrade, to respond to a particular offer at a particular hour. Guests say they'll pay for it. Deloitte's survey of 5,564 US loyalty members across seven consumer industries, hotels among them, found 51% of Gen Z and 53% of millennials would spend more with a brand that personalizes, against 19% of baby boomers. Most companies are still early, though. McKinsey's 2025 State of AI found 88% of organizations use AI in at least one function, but only 39% say it has touched enterprise-wide EBIT.
This is the operator's version, for independent properties, boutique brands and regional groups. They can't match Marriott's R&D budget. They can move faster and know their guests better. What follows covers the data layer that makes predictive loyalty possible, the four models that do most of the work, an implementation sequence, and the governance that keeps the program on the right side of the privacy line.
Why Points Stopped Working
Points programs solved a pre-internet problem: recognizing the guest who came back. Punch cards, frequent flyer accounts and Holiday Inn's Priority Club let a business spot its repeat customers, and the points made that recognition feel like a reward. For a long time that was enough, because the alternative was nothing.
The alternative is no longer nothing. Every guest who walks into your hotel is already known to a half-dozen tech companies that know more about her preferences, her travel patterns, and her purchasing intent than your loyalty system does. Google Maps knows which restaurants she ate at on her last trip. Expedia knows what she searched and abandoned. Her credit card issuer knows where she stayed for the last decade and at what rate. In that environment, a generic 10% member discount and a path to free nights at properties she may never visit is not a competitive moat - it is a commodity. Skift's finding that engagement per member is falling is the data version of the same point.
Meanwhile the program isn't getting cheaper. CBRE puts loyalty cost per occupied room at $5.46 in 2024, about 1.6% of total revenue, and program fees grew 4.4% that year, ahead of total revenue growth of 2.7%. And the currency itself isn't landing: in Deloitte's survey, 40% of loyalty members admitted they sometimes forget to redeem. The next dollar of loyalty spend has to buy intelligence, not more currency.
"The size of the loyalty list has stopped predicting the strength of the business. The next dollar of loyalty investment has to buy intelligence, not currency."
The Four AI Models That Drive Modern Loyalty
Strip out the marketing language around "AI loyalty" and four predictive models do most of the work. Each solves a problem the points ledger can't solve on its own.
1. Churn prediction
Build this one first. Churn prediction reads behavioral signals (falling stay frequency, points piling up unredeemed, lower email open rates, fewer app sessions, shrinking ancillary spend) and gives each member a near-term probability of drifting away. The intervention doesn't have to be a discount. Often the better ones are personal: a note from the GM of the guest's usual property, an experience offer in the destination they were last searching, a tier-status grace extension before it lapses.
2. Customer lifetime value modeling
CLV models look forward. They estimate what a guest will spend across all stays, channels and ancillary categories over a set horizon, and they revise that estimate every time the guest interacts. The uncomfortable thing CLV shows: most programs spend the same on every member, and members aren't worth the same. CLV lets you move spend toward the minority of members who will generate most of the forward revenue.
3. Next-best-action engines
A next-best-action (NBA) engine sits on top of the churn and CLV models and answers the operational question: given everything I know about this guest right now, what is the single most valuable thing I can offer her in the next 24 hours? The answer might be an upgrade at her favorite property, a spa pre-book for next month's stay, a delayed-checkout email, a points bonus on a competitor-comparable rate, or content designed purely to re-engage. The critical point is that the action is selected from a portfolio of possibilities, not a single channel - the model is choosing between email, app push, on-property in-room display, front-desk script, and concierge handoff in real time.
4. Emotional and experiential scoring
The newest layer, and the one most differentiated for boutique and luxury operators, is emotional loyalty scoring. These models combine sentiment from reviews, NPS responses, social signals, and on-property feedback to estimate the affective relationship a guest has with the brand - not just their transactional frequency. Two members can have identical stay counts and wildly different emotional loyalty scores; only one will survive a service failure or a price comparison. Skift Research's November 2025 report on the future of hotel loyalty argues that experiences are becoming a loyalty engine as programs move past points. For independents who cannot compete on global redemption breadth, this is the asymmetric advantage.
Where does your property actually stand? The AI Readiness Score takes two minutes, asks no one for a call, and emails you a report sized to your rooms and ADR.
Get Your ScoreArchitecture: The Data Layer That Makes It Work
None of the four models above can outperform the data they are trained on. A common reason hotel loyalty AI projects stall is that they launch on top of fragmented, unreconciled, low-quality data - PMS data that does not match CRM data, loyalty IDs that do not match folio email addresses, channel manager records that lose the loyalty number when a booking comes in through an OTA. A model trained on dirty data will confidently produce dirty predictions.
The fix is a unified guest profile, usually delivered through a customer data platform (CDP) that sits between the source systems (PMS, CRM, RMS, POS, booking engine, loyalty database, app, marketing automation) and everything downstream: the models, the marketing tools, the front-desk screen. Travel-focused CDPs stitch records together with deterministic matching on loyalty IDs, email addresses and phone numbers, plus probabilistic matching on name and travel patterns, so each guest ends up with one record carrying stay history, preferences, sentiment and scores. The big brands have been doing this for a while: Wyndham deployed Amperity's CDP in 2020 to unify customer databases and centralize online and offline transactions.
The technology stack at a glance
The table below is the simplified reference architecture for a property or group launching AI-driven loyalty. The exact vendor in each row is less important than the function it provides - the stack only works if every layer is present and the data flows are bidirectional.
| Layer | Function | Representative vendors |
|---|---|---|
| Source systems | PMS, CRM, RMS, POS, booking engine, loyalty database, app, email | Opera, Mews, Cloudbeds, Salesforce, IDeaS |
| Identity resolution | Stitch records into a single canonical guest profile | Amperity, mParticle, native CDP modules |
| Customer data platform | Unified profile + activation API + segmentation | Revinate, dailypoint, Ireckonu, Snowflake CDP |
| Predictive models | Churn, CLV, NBA, sentiment / emotional scoring | In-house ML, Klaviyo predictive, Optimove |
| Activation channels | Email, app push, in-room display, front-desk script, concierge handoff | Revinate, Cendyn, Braze, native PMS modules |
| Measurement and attribution | Incrementality tests, holdout groups, CLV-weighted ROAS | In-house BI, Tableau, Looker, native CDP reporting |
The Economic Case: Where the Money Comes From
There's no credible published benchmark for what a predictive loyalty program returns at an independent hotel, and vendor case studies rarely travel. Build the case on your own numbers, lever by lever: revenue retained from members who would otherwise have lapsed, commission saved on stays that move from an OTA to direct, ancillary revenue per stay, and revenue per marketing send. Measure each against a holdout group rather than against last year, because last year had a different market in it. On the cost side, count CDP licensing, integration work, model tuning and staff time.
Data quality on day one decides most of the outcome. Properties that clean up before the first model is trained get usable predictions sooner. The ones that skip it usually end up doing the cleanup later anyway, after the first model has already misfired.
Tier Strategy in an AI World
One of the most counterintuitive consequences of predictive personalization is that classic loyalty tiers - Silver, Gold, Platinum, Diamond - become less central, not more. The tiers were always a coarse proxy for "how much should we care about this guest?", and the proxy was based almost entirely on stay frequency. CLV modeling gives a sharper answer. Say a Silver member has a $48K forward CLV: she deserves more attention than a Gold member at $14K who is two reservations away from churn. Most operators are not ready to abandon the tier branding (Bonvoy Platinum is itself a marketing asset), so the right move is to keep the public-facing tiers and operate an internal score that overrides them when the two disagree.
Here is how the two systems interact for four hypothetical members.
| Member | Public tier | Forward CLV | Churn risk | Operational priority |
|---|---|---|---|---|
| Member A - biz traveler, 32 stays/yr | Platinum | $96K | Low (4%) | Standard premium handling |
| Member B - anniversary couple, 2 stays/yr | Silver | $48K (high spend, slow growth) | Medium (18%) | Personal outreach, upgrade reserved |
| Member C - declining traveler, 6 stays/yr | Gold | $14K | High (54%) | Save-the-guest sequence triggered |
| Member D - new high-value signup | Member | $71K (predicted) | Low (8%) | Onboarding NBA, fast-track tier ladder |
Channel Economics: Why This Drives Direct Bookings
Predictive loyalty is also a direct booking program, because nearly every personalization signal only works on the direct channel. The OTAs don't pass through enough behavioral context to feed the models, and points lose their pull when the booking comes through Booking.com or Expedia. The table below shows SiteMinder's 2025 average booking value by channel.
| Channel | Avg. booking value (2025) |
|---|---|
| Hotel websites (direct) | $516 |
| Wholesalers, DMCs and tour operators | $445 |
| GDS | $392 |
| OTAs | $312 |
SiteMinder's 2025 data has hotel websites averaging $516 per booking against $312 on OTAs. At the global level, Hospitality.today, citing BCG's 2026 AI-First Hotels report, put digital direct bookings at $262 billion against $266 billion in OTA transactions, a gap of about $4 billion, and BCG credits loyalty perks, member-only rates and better brand-site content for closing it. Independents have followed more slowly and with less infrastructure. Every booking that comes back direct also skips an OTA commission, which The Percentage Company puts at 15% to 25%. A predictive loyalty program is a direct booking machine wearing different clothes.
"A Silver member with a $48K forward CLV deserves more attention than a Gold member with a $14K forward CLV who is two reservations away from churn. The tier was always a proxy. CLV is the answer."
A 90-Day Implementation Sequence
Sequence matters more than vendor choice. Operators who launch every model at once usually launch none of them well. The order below works the same for a single boutique property and a portfolio; the milestones stretch, the order doesn't.
Days 1 to 30: Data foundation. Identify all source systems and document the data each one holds. Stand up the CDP and complete identity resolution - every guest must have a single canonical profile, with stay history, loyalty status, preferences, and consent flags resolved. Pull a clean six-month export from the PMS, CRM, and booking engine to use as the training corpus. Establish baselines for the four KPIs that will define success: loyalty churn rate, direct booking mix, ancillary capture per stay, and email-driven revenue per send.
Days 31 to 60: First model live (churn). Deploy the churn prediction model in shadow mode against historical data and validate accuracy against actual outcomes. Once its predictions hold up against what actually happened, move to production but route interventions to a small holdout group first. Build the save-the-guest playbook - three to five distinct interventions with clear ownership (GM personal outreach, concierge offer, tier extension, surprise upgrade, F&B credit). Measure incrementality with a 10 to 15% holdout.
Hotels starting this work often benefit from a structured guest experience and personalization assessment before standing up the CDP - getting the data architecture, consent framework, and intervention playbook right at the start is what separates the deployments that pay back fast from the ones that stall. Explore our AI-Powered Guest Experience Systems service → for how that engagement works.
Days 61 to 75: Second model live (CLV). Train and deploy the CLV model with the same shadow-then-holdout pattern. Begin reallocating marketing spend and front-desk attention based on the forward value scores rather than the public tiers. This is the change-management moment that requires the most internal communication - front-desk teams will treat members differently than the tier badge says they should, and that needs to be explicitly authorized at the GM level.
Days 76 to 90: NBA engine and measurement layer. Layer the next-best-action engine on top of the first two models. Begin with three channels (email, app push, front-desk script) and a constrained action portfolio (six to eight offer types). Stand up the measurement dashboard that the GM and revenue manager will use weekly. By day 90, you should have all three models in production, a clean holdout-based attribution view, and the first month of revenue data showing the lift.
Governance, Consent, and the Privacy Line
The same techniques that make predictive loyalty so effective also make it one of the riskier personalization projects a hotel can run. The models work because they consume behavioral data - and behavioral data is among the most sensitive data guests hand over. A program that crosses into creepy or non-compliant territory will damage trust faster than the points ledger ever delivered it.
Three governance practices matter most. First, explicit consent and clear preference centers: every member should know what data is being used, see how it is being used, and be able to revoke or modify consent in one click. Programs that hide behavioral usage in T&Cs face real regulatory risk under GDPR, CCPA, and the new state privacy laws. Second, model auditability: every prediction that triggers a guest-facing action should be explainable in plain language to the GM, to the guest if she asks, and to a regulator if required. Black-box recommendations that nobody can defend are a liability in waiting. Third, asymmetric value: the guest should always receive obvious value in exchange for the data. If a member can articulate "they remembered I prefer a quiet floor" without being able to articulate "they are using my browsing data to predict my next trip," the program has crossed the line.
The practical implementation is a consent framework that distinguishes between operational personalization (always allowed within scope of stay), marketing personalization (opt-in), and behavioral profiling (explicit opt-in with renewal). Combined with a clear privacy statement and an internal review process for any new model that touches guest data, the framework covers most of the regulatory and reputational risk.
The Independent Operator's Advantage
It is tempting to read all of this and conclude that only the global chains can play. The opposite is closer to the truth. The chains have scale on the loyalty side and a long tail of legacy data problems on the technology side. Independent and boutique operators have smaller, cleaner datasets, faster decision cycles, and more direct guest relationships - and packaged CDP and model products now put much of what Marriott builds in-house within reach of a single property. Deloitte's outlook on the future of hospitality starts from the same place: "The days when scale alone guaranteed success are fading." Treat the loyalty database as something that learns, not a ledger. The mechanics are no longer the moat. The intelligence on top of the mechanics is.
Frequently Asked Questions
Do we still need a points program at all?
For most properties, yes - but its role changes. Points remain useful as a transparent recognition mechanic and a redemption-tracking system, and they anchor the public tier branding. What they no longer do well is drive incremental behavior. The right architecture keeps the points ledger as table-stakes infrastructure and layers the predictive intelligence on top of it. Independent operators sometimes succeed by replacing points entirely with a recognition-plus-value program, but that requires confidence in the predictive layer and a willingness to retrain guest expectations. Most groups should start with the layer-on approach and revisit the question after 12 months of data.
How much guest data do we actually need before the models work?
Quality matters more than volume. A churn model needs enough history to have seen members lapse; a CLV model needs longer stay histories before it settles; emotional scoring needs review and NPS volume, which public OTA reviews can supplement. Properties with two years of fragmented, duplicate-heavy data will usually get worse results than properties with one year of clean, identity-resolved data.
What is the relationship between predictive loyalty and our revenue management system?
They are complementary, not competing, layers. RMS optimizes pricing at the rate-and-channel level for the inventory you have today. Predictive loyalty optimizes which guests you want to attract and retain across the lifetime of the relationship. The two converge at the booking - a great RMS will dynamically discount a high-CLV at-risk member when their churn risk crosses a threshold, and a great loyalty model will tell the RMS which segments to protect rate integrity for. The integration is usually a webhook or an API call from the loyalty layer into the RMS, not a wholesale platform decision.
How do we handle GDPR, CCPA, and the patchwork of U.S. state privacy laws?
The core architectural answer is consent management at the profile level - every member has an explicit consent record covering operational, marketing, and behavioral personalization, and any model that consumes data checks consent at runtime. Geographic policy differences are handled by tagging profiles with jurisdiction at sign-up and applying the strictest applicable rules. A frequent compliance gap is not the model but stale consent on legacy records; running a consent refresh campaign before launching predictive features prevents this. Engaging counsel on the consent language and on cross-border data transfers is non-negotiable.
What does this look like at a single-property independent vs. a portfolio?
The architecture is the same; the operating model is different. A single independent typically deploys a packaged CDP+models combination from a hospitality-native vendor (Revinate, Cendyn, dailypoint), runs the program through the existing GM and revenue manager, and reports monthly to ownership. A portfolio adds a central data team, a portfolio-wide CLV model that recognizes cross-property stays, and a shared playbook that each property executes locally. The portfolio gains the ability to identify members worth winning across multiple destinations; the independent gains speed and intimacy. Both can win at this - the loser is the mid-size group that tries to centralize without investing in the data team to support it.
Related Research
Get two full chapters of The 2026 Official Guide to Hotel AI
The AI Readiness Scorecard + the P&L Opportunity Heat Map - see exactly where your property stands and what each gap is worth.
Plus one tactical AI play each week. Unsubscribe anytime. Prefer to start interactive? Take the 2-minute AI Score.
Reading about it is the easy part.
The gap between an article like this and an actual result is a ranked list of what is worth your money at your property. There are two ways to get one, and neither of them costs anything.
- Peter Mack, founder, HospitalityOS