Let me describe a moment that happens in basically every independent hotel I work with. The owner opens the email platform, sees a soft week three Tuesdays out, and fires off a 15-percent-off flash sale to the whole list. Bookings come in. Everyone high-fives. The sale “worked.”
Here is the uncomfortable question I always ask next: how many of those people were already going to book?
Because if a guest had your dates in a cart, your hotel bookmarked, and a calendar reminder set, and then you emailed them 15 percent off the rate they were about to pay full price for, you did not win a booking. You bought a booking you already had, and handed back margin for the privilege. Multiply that across a year of “it worked” sales and you have quietly trained your best, most loyal guests to wait for a coupon.
This post is about the fix. It is called uplift modeling, and it is the most underused idea in hotel marketing. The whole point is simple: stop discounting guests who would have booked anyway, and aim your discount budget only at the people whose decision actually changes because of the offer.
The four kinds of guests every promo hits
Whenever you send an offer, every person on your list falls into one of four buckets. This framing is the entire mental model, so it is worth slowing down on.
| Guest type | Books without offer? | Books with offer? | What your discount does |
|---|---|---|---|
| Sure Things | Yes | Yes | Pure margin given away |
| Persuadables | No | Yes | The only group worth paying for |
| Lost Causes | No | No | Wasted send, no harm |
| Sleeping Dogs | Yes | No | Offer actively backfires |
The Sure Things were booking regardless. Every dollar of discount you give them is margin set on fire.
The Persuadables are the gold. They were on the fence, and the offer tipped them over. This is the only group where a discount creates a booking that did not otherwise exist.
The Lost Causes are not booking no matter what you do this cycle. Emailing them costs little, but they pad your “reach” numbers and fool you into thinking the campaign was bigger than it was.
And then the dangerous one, the Sleeping Dogs. These are people who were going to book at full rate, but an aggressive discount email makes them feel like a sucker for almost paying full price, so they stall, shop around, or wait for the next sale. Over-promoting your loyal base can genuinely suppress bookings. I have watched it happen.
A normal “who is likely to book” model, the kind every email tool nudges you toward, happily targets the Sure Things, because they look fantastic on paper. They open everything, they have booked before, they are “engaged.” Uplift modeling does the opposite. It hunts for Persuadables and tells you to leave the Sure Things and Sleeping Dogs alone.
The question that breaks most hotel promo strategy is not “who is likely to book?” It is “who is likely to book BECAUSE of this offer?” Those are different people, and the gap between them is where your margin leaks out.
Why this matters more for hotels than almost anyone
Two reasons uplift thinking pays off harder in hospitality.
First, your margin structure is already under siege. When an OTA takes its cut, you are giving up roughly 15 to 25 percent of that booking. The whole reason you push direct is to claw that margin back and build a healthier OTA mix. So when you then turn around and reflexively hand a 15-percent discount to people who would have booked direct at full rate anyway, you have recreated the OTA tax with your own two hands, except now nobody else even gets the guest. If you want the deeper version of that math, I wrote it out in the book-direct commission breakdown, and the book-direct CRO service is built around plugging exactly these leaks.
Second, hotel demand is lumpy and date-specific. A retail brand can blast a discount and let it ride. You cannot, because a guest who would have paid rack rate for your one suite on a Saturday in peak season is the most expensive person on earth to discount. The cost of mis-targeting a Sure Thing in hospitality is brutal precisely because your inventory is scarce and time-bound.
The version you can run this quarter (no data scientist)
I am going to give you the practical, do-it-in-a-spreadsheet version first, because the machine-learning stuff is an optimization, not a prerequisite. The foundation of all uplift work is the same humble thing: a holdout group.
Here is the play.
Step 1: Pick one real promotion and one real audience
Say it is your shoulder-season midweek offer going to your past-guest and abandoned-booking list. Do not send it to everyone yet.
Step 2: Randomly hold back a control group
Before you send, randomly carve out, say, 20 percent of that audience and send them nothing (or your normal non-discount newsletter). The keyword is randomly, not “the people who seem least likely to book.” Random assignment is the only thing that makes the comparison honest. Most email platforms have an A/B or holdout split built in; if not, a spreadsheet and a random-number sort does the job.
Step 3: Wait the full booking window, then compare
Look at the booking rate in the treated group versus the control group over a window that matches your real booking lead time. The difference between those two numbers is your uplift. Not the raw booking rate of the people who got the offer, the difference. That gap is the only thing the discount actually created.
A clearly illustrative example so the numbers feel real: imagine the discounted group books at 6 percent and the held-back group books at 4 percent. Your uplift is 2 percentage points. That means for every 100 emails, the offer generated 2 incremental bookings. The other 4 booked anyway, and you gave those 4 a discount for nothing. Suddenly the “true” cost of each incremental booking is way higher than the sticker discount suggested, because you are also eating the margin you gave the Sure Things. These figures are made up to show the shape of the math, not a result I am promising you.
Step 4: Slice the uplift by segment
This is where it gets useful. Run that same treated-versus-control comparison within segments you already have: past guests versus never-stayed, opened-last-3-emails versus dormant, direct-bookers versus OTA-sourced, lead time, room type interest. You are looking for the segments where the gap between treated and control is biggest. Those segments are your Persuadables. The segments where treated and control book at nearly the same rate are full of Sure Things, and you should stop discounting them.
That is uplift modeling. The whole thing. Buckets, a clean holdout, and the discipline to read the difference instead of the headline number.
Graduating to a real uplift model
Once you trust the holdout habit and you have enough volume, you can let an actual model find Persuadables at the individual level instead of the segment level. The cleanest approach for most hotels is the two-model method:
- Build one model predicting booking probability for people who got the offer.
- Build a second model predicting booking probability for the holdout group who did not.
- For each guest, subtract: predicted-with-offer minus predicted-without. That difference is their individual uplift score.
Rank everyone by that score. Send the offer to the top slice, leave the bottom and the negatives (your Sleeping Dogs) alone. There are tidier single-model techniques too, but the two-model version is the one I can explain to a GM in ninety seconds, and explainable beats clever in a hotel that has to actually run this.
The features that tend to carry uplift signal for hotels: recency and frequency of past stays, how the guest originally found you, lead time behavior, rate sensitivity in past bookings, email engagement depth, and whether they have ever bounced from a booking page. You feed those in. The model does not tell you who will book; it tells you whose decision moves.
The most expensive guest you can discount is the one who already had their credit card out. Uplift modeling exists to keep you from emailing that person a coupon.
Where to point this first: retargeting and email
Two channels give you the fastest payback.
Retargeting ads. The default retargeting setup chases everyone who hit your booking page with the same discount creative for two weeks. A lot of those people booked already, or booked through an OTA, or were always going to come back. Uplift logic says: suppress the Sure Things, and reserve the discounted creative for the genuinely on-the-fence segment, while showing everyone else a no-discount reminder. This pairs directly with fixing the search side of the funnel, because if guests are finding OTAs before they find you, you are paying to retarget demand you should have captured for free. That is the whole argument in how OTAs intercept your search traffic and why your hotel ranks below OTAs for its own name.
Email promotions. Stop the all-list blast reflex. Segment by uplift, send the discount to Persuadables, send Sure Things a value-add or experience email at full rate, and let Sleeping Dogs simply stay asleep with normal content. Your discount budget shrinks and your incremental bookings hold or rise. The work of building those value-add, non-discount touchpoints is exactly what the content and reputation service supports.
The honest caveats
I am not going to oversell this. Uplift modeling needs volume; if you send a few hundred emails a month, your holdout groups will be too small to read confidently, and you should run simpler tests and revisit this when you have grown. It needs patience, because you must wait out the full booking window before judging anything. And it needs you to resist the dopamine of the raw booking count and stare at the uplift instead, which is genuinely harder than it sounds when bookings are rolling in.
It is also not a ranking or visibility tactic. It optimizes the demand you already have. The job of creating that demand, getting found in Google and in AI answers, lives elsewhere; if you are starting from scratch, the 2026 hotel SEO starter guide and the AI visibility service are the front of that funnel, and hotel SEO is the engine underneath. Uplift modeling just makes sure that once the demand shows up, you stop paying retail to convert people who were already sold.
There are no guaranteed results here, and anyone promising you a fixed lift is guessing. What I can tell you is that the math of giving discounts to Sure Things is always negative, and that holdout testing is the only honest way to find out who your Persuadables actually are.
Want help wiring this up?
If you are tired of running flash sales and never knowing whether they made you money or just gave it away, this is exactly the kind of measurement work I love. Book a free intro call over at the booking page and we will look at your actual list, your booking windows, and where your discount budget is leaking, then design a holdout test you can run next cycle. No coupon required.