Hedonic Intelligence
Studies

Two national-scale investigations. Read once, adversarially tested, published once.

617 hotels combined, 1.86M+ reviews, across two independent studies built the same way: generate every hypothesis, attack it, publish only what survives.

1.86M+ combined guest reviews, 617 hotels, two independent studies

One study is finished at 117 hotels. The other just published its first release across Portugal's 500 best hotels: four times the scale, read the same way.

The four numbers to take away
Sample details & methodology
Luxury/Midscale churn: rating-8 reviews with churn signal populated, n=24,005 of 47,666, Wilson 95% CI 14.5–18.0% / 3.6–4.4%. Early recovery: matched-severity medium friction events, n=5,373 of 67,709, Newcombe CI 40.8–45.1pp. Fought for the fix: holding severity and resolution status fixed, high guest-effort runs +12.84pp [95% CI 10.75–14.93] more detractors than low-effort within fully-resolved cases; replicates on churn-likelihood at +17.65pp. Coordinated-push flags: 32 of 1,258 tested hotel-platform series flag as statistically implausible volume bursts (29/32 survive a seasonal placebo); 6 of those 32 additionally clear a four-signal compound coordinated-push pattern; never validated against a confirmed-manipulation ground-truth case. For reference, this band's earlier headline numbers stay live elsewhere on this page: 89.9% of friction events nationally never get a guest-confirmed response (PT500-M9, cited in Study 02's own Finding 01), and 14.6% of the 500 hotels flag RED on the 6-axis pre-LOI screen (PT500-M20, cited in Finding 04).
16.2% vs 4.0% Luxury hotels lose four times more guests at the identical 8-out-of-10 score than midscale hotels do.Feeds: which tier your screen quietly stops checking.
43pp Catching a problem before the guest complains works nearly twice as well as fixing it after the fact.Feeds: whether a service-recovery fix or a bigger renovation spend is the higher-return move.
+12.84pp Even a "resolved" complaint costs extra if the guest had to fight for the fix, on top of whatever the fix itself bought back.Feeds: whether your own reporting stack tracks how hard a guest had to push, or only whether the ticket closed.
6 of 32 A screen already exists for whether a hotel's reviews look like a coordinated push, and it can clear a hotel just as easily as flag one.Feeds: the specific question a skeptical buyer asks in a first meeting, before any question about a repair bill.
Study 01 · Finished, adversarially verified, ready to read
259,647 reviews · 117 hotels · 8 countries · 4 platforms · 25 to 26 languages

The 117-Hotel Cohort: where the guests you already lost are sitting in a portfolio

Published

The kind of finding a due-diligence memo is usually written without: which guests are already gone, which tier and price posture concentrates the loss, and what early recovery is worth in points of return intent. Eight chapters, each closing in a decision, not a description.

86% of 81 qualifying hotels show a 40-point swing in who comes back, between guests who called the stay overpriced and guests who called it worth it, a split most acquisition screens never run
16.2% Luxury hotels lose guests at four times the Midscale rate, at the identical 8-out-of-10 score, the gap a capex reposition case has to clear
43pp more guests recommend the place when recovery happens before they complain rather than after, comparing like-for-like problems, and it's used in under 4% of the cases that call for it
Sample details & methodology
DD Gap: 81 of 117 cohort hotels clearing n≥39 in both price-verdict cells, mean gap 53.8pp (95% CI 51.3–56.4), 86% of qualifying hotels clear the 40pp threshold. Tier churn: rating-8 reviews with churn signal populated, n=24,005 of 47,666 (50% coverage), Wilson 95% CI 14.5–18.0% (Luxury) / 3.6–4.4% (Midscale). Early recovery: matched-severity medium friction events, n=5,373 of 67,709, Newcombe CI 40.8–45.1pp on the 43-point gap.
How the risk is sized

Churn and non-return figures above are read from what guests write, not from reservation data. They signal intent, never confirm a booking.

What this means for a buyer

The gap concentrates where the underwriting case is hardest to reverse: once capital is committed to a Luxury reposition, a won't-return rate four times the Midscale baseline is the kind of number that erodes the return case quietly, deal by deal, long after the bid is signed.

Has a won't-return screen run on the assets already in your portfolio, or only on the ones you're bidding on?

Read what the score isn't telling you →
Study 02 · Published
500 hotels · 1,596,417 reviews · 4 platforms · national index

The Portugal 500

Published

The kind of read a listing page can't give you: which hotel looks risky but isn't, which one looks safe but is quietly losing its own guests, and what a "resolved" ticket still costs when the guest had to fight for it. Every finding closes in a decision, not a description.

3x An underrated hotel, one its own guests already rank higher than its blended score shows, flags real risk about three times more often than average. A hotel whose rating looks a little too good flags at close to zero: the opposite of the instinctive read.
+12.8pp More guests turn against a hotel when they had to push, follow up, or argue for a fix, even though the problem gets fully resolved either way. The gap is bigger, not smaller, once a fix actually lands.
11% is all a hotel's own star rating explains about whether its guests are loyal to the property itself or just to the destination around it, tested three separate ways, same answer each time.

First national release, 2026-07-10. Adversarially tested the same way as Study 01.

1.6M+ reviews / 500 hotels, national index / 4 platforms at once / Booking, Expedia, TripAdvisor, Google
Sample details & methodology
Flattered vs. underrated: underrated hotels (rank_gap ≥70 positions) flag RED 42.3% [Wilson 25.5–61.1], k=11/26, vs 14.8% [11.4–19.1] base rate, k=48/324; flattered hotels flag RED 0.0% [0.0–13.3], k=0/25. Effort scar: holding severity and resolution status fixed, high guest-effort runs +12.84pp [95% CI 10.75–14.93] more detractors than low-effort within fully-resolved cases; replicates on churn likelihood at +17.65pp. Property-Pull Index: the published rating explains just 11% of which kind of asset you're looking at (R²=0.112, Spearman ρ=0.319, Pearson r=0.334, n=493). 145,674 friction events detected nationally across 500 hotels and four platforms; 130,912 of those, 89.9% (Wilson 95% CI 89.7–90.0%), are coded never-addressed by the guest's own account. Read from what guests write, not an operations log: signal, not a resolution record. Confirmed platform- and scale-invariant to within 3pp on every cut tested. Corpus: 1,596,417 reviews extracted of 1,665,388 collected.
How the risk is sized

Read from what guests write, not from an operations log: signal, not a resolution record.

What this means for a buyer

The instinct is to worry about the hotel whose rating looks a little too good. This inverts it: the hotel that should worry you is the one its own guests already rank higher than the world does, and a complaint that shows as "resolved" in an ops log can still be quietly costing you the guest who had to fight for the fix.

The index exists now. Have you checked where the hotel you're deciding on sits in it?

See the national index →
Before you name a property (product demo, not a study)

See what you would know before the first call

Walk through exactly what lands in your inbox after a paid, one-hotel engagement, built from data on hotels we've already studied, so the shape is real even though the property isn't yours yet. Not a finding on your asset. A blueprint for what one would look like.

Want to see the shape of the answer before you risk naming the asset? This is that.

See what you'd know before the first call →
How an investigation works

The discipline that separates a finding from a label. Three stages, every property.

One that survived, from the Portugal 500

Two of the country's largest hotel groups, and the guests who never came back.

Self-reported loyalty is easy to overstate, so this run tested it behaviorally instead: 109,829 consecutive guest journeys, tracked by the same reviewer moving from one national hotel to the next. Large domestic multi-property groups showed no revealed lift in same-brand rebooking, at or below what national chance alone would predict, and the result held up on an independent guest-half split. International full-service chains cleared the identical test at up to eight times chance. One domestic group is the exception: a genuine, replicated lift of its own, and it stays out of the pattern below on purpose.

Large domestic groups
≤ chance
at or below the national base rate, 2 of the 3 largest domestic full-service groups tested
International chains
up to 8.0x
FDR-significant, split-half replicated, across the two clearest cases

The pattern held across 29 brand-level tests, corrected for testing 29 at once, not cherry-picked. Property, not brand, is what a repeat guest is actually rewarding, and this is what happens when that claim is checked against what guests actually did, not what they said.

Sample details & methodology
109,829 consecutive, unambiguously-ordered same-author visit pairs, TripAdvisor only (the sole platform with person-stable author identity in this corpus). 29 brand-level tests, Benjamini-Hochberg FDR-corrected across all 29. Class-level restatement: no domestic or international group is named here; the one real domestic exception is never folded into the "no domestic lift" pattern.
01
Generate: many hypotheses
The full guest record is read. Every pattern, friction, and signal becomes a candidate hypothesis.
02
Test adversarially: every one is attacked
Each hypothesis is tested against alternative explanations, edge cases, and statistical floors. Only those that hold up under adversarial pressure advance.
03
Close in a decision
Of every crop of hypotheses, only 6 to 10 survive to close in a named action, carrying a stated sample size, a margin of error, and an explicit limit.
Every chapter on Study 01 exists because it survived this.
From investigation to decision

A 43-point swing in return intent when recovery lands early, tested against a peer set your committee can defend. The same investigation, on the property you name.

One hotel, a bespoke peer set, every finding closing in a decision your committee can act on. Built from the outside, from public reviews alone. No operator cooperation, no PMS data, no NDA.

Scope Any hotel you name: a target before a bid, a rival you will never get inside, an asset under an operator you are deciding whether to keep. Built from the outside, from public reviews alone.
Peer set A bespoke comparison group for your property, on segment, geography, and competitive positioning. Defensible to an investment committee. Not an average across every hotel we've already studied.
Output Each finding closes in a decision, with confidence intervals and an explicit perimeter.
Cadence One-off, no subscription, no onboarding. 14 days from named hotel to finished read.

Every read is done by the person who built the method. This is what that looks like.

This is an independent practice in its first engagements: commission a read and you are among the first, and the work is priced and structured accordingly.

01
See the national index
500 hotels, published, and every finding closes in a decision.
02
See the proof, not the pitch
Chapter 1: two guests, one hotel, and only one of them is coming back.
03
Name a hotel
Book a 20-min call. Bring a property: is this worth commissioning on yours?
Book a 20-min read intro →