Hedonic Intelligence
The Portugal 500

Same 8.90 to 9.10 rating.

One hotel owns its guests. The other rents them from the coastline.

Portugal's first national index of hotel guest experience: 500 hotels, four platforms, read once and adversarially tested, published 2026-07-10.

Book a 20-min intro → See the headline findings ↓

This is the first release. Findings refresh on a cadence, not continuously (see the perimeter below for what that cadence is, and is not).

Somewhere in this index, two hotels sit at the identical rating, close enough that a guest scanning scores alone would call them the same hotel. They are not: the Property-Pull Index, how much of a hotel's return demand it owns outright versus borrows from the destination, is what actually separates them. A curated resort captures the credit for demand even when the underlying draw is the coast; a plain hotel on the same coastline does not. Format decides it, not location.

Property-Pull Index · same rating, still this far apart
Both bars show the middle 80% (p10–p90), the same statistic each time, not a full range next to a narrower one. Shrink the ratings down to a sliver, 8.90–9.10, eighty-six hotels a guest would call interchangeable, and the spread barely moves: 12.6 points instead of 13.0. The rating didn’t do the sorting. It just came along for the ride.
What the rating explains
Barely a tenth of it.
The star score can’t tell you which twin you’re looking at. The rest comes from somewhere the rating never measures.
Sample details & methodology
Three separate statistical tests agree on this: R²=0.112, Spearman ρ=0.319, Pearson r=0.334 (n=493 hotels, within this curated 500, not the market at large). A per-hotel, empirical-Bayes-shrunk index, not a raw rate: hotels with thin attribution volume are pulled toward the national mean (91.4% property-attributed) rather than allowed to sit on a noisy extreme, which is why the floor lands at 67.2% and not lower. Denominator: 91,339 property-or-destination-attributed return-intent reviews nationally; 493 of 500 hotels clear the ≥15-attributed-mention floor to receive a score, 7 do not. Every number here is a lower bound on the true destination-captured share: 12.0% of return-intent reviews never reach a substantive attribution at all, so the true split skews further toward destination-borrowed than the index alone shows. Counter-intuitively, resorts score HIGHEST on this index (94.3% mean), not lowest: a curated, all-inclusive format absorbs the destination experience inside its own perimeter and gets credited for it; a plain hotel on the same coastline does not capture that same credit.

Before you underwrite, renew, or walk away from either one, which twin would your own reporting already have told you it was?

The Portugal 500 · Self-locator

Is your hotel already in this set? Search and find out.

500 celebrated Portuguese hotels already sit in this index, read guest by guest across four platforms. Type a name below, a competitor's, a target's, or your own, to see if it is one of them.

500 hotels indexed · 1,665,388 reviews collected across them as of the July 2026 index cut · 494 of 500 read on all four platforms at once, the other six on three

Try a name you know. It fills in as you type; pick one to see it.

1.6M+ reviews
1,665,388 collected, 1,596,417 extracted, the full public guest record the investigation reads
Every great hotel
in Portugal, every great hotel, nothing else
4 platforms
Booking, Expedia, TripAdvisor, Google, all at once
First release, 2026-07-10
periodic re-reads planned, not a live feed
The Portugal 500 · First release

Four things the national index already found. Seven more below.

Every angle run against the full 500-hotel record: Portugal's own curated best, not a random sample of the market. Every candidate finding attacked before it shipped. This is what survived the attack, ordered by who it's for.

Finding 01 · Independence

Even when Portugal's own curated best hotels do fix a guest's problem, how hard the guest had to fight for that fix costs nearly as much as the problem itself.

"Resolved" in an ops log can hide two very different guests: the one who asked once and got an easy fix, and the one who had to push, follow up, or argue their way there. Hold the exact same problem and the exact same severity fixed, and the guest who had to fight still leaves less happy, on top of whatever the fix itself bought back.

Fully resolved, but the guest had to fight for it
+12.84pp
more detractors than an easy fix · CI 10.75–14.93pp · severity-matched
Never resolved at all
+6.08pp
the same effort penalty, smaller once nothing gets fixed · CI 5.25–6.91pp
Sample details & methodology
Holding friction severity and resolution status fixed nationally: high guest-effort runs +12.84pp [95% CI 10.75–14.93] more detractors than low-effort within resolved='yes' cases, and +6.08pp [5.25–6.91] within resolved='no' cases, i.e. the effort penalty is larger, not smaller, once a fix actually lands. Placebo-tested at z≈10.7–14.5. Replicates on a second outcome (churn likelihood): +17.65pp / +5.18pp on the identical design. A shrinkage-stabilized per-hotel Guest Effort Score (421/500 hotels) correlates moderately (r=0.434) with a hotel's overall detractor rate, a convergent-validity signal, not proof on its own.

This sits on top of an already-real gap. Nationally, close to nine in ten friction events Portugal's own curated best hotels' guests describe are never addressed at all, by the guest's own account, not an internal ticket, and even a full, guest-confirmed fix, no fight required, only closes most of the distance back to a guest who never had a problem, not all of it.

Sample details & methodology
Never addressed: 89.9% at the event level (CI 89.7–90.0%, k=130,912/n=145,674), 87.9% at the review level (CI 87.7–88.1%, k=92,749/n=105,513); guest-perceived resolution, read from the guest's own account, not a hotel-side ticketing log. Full-resolution swing: 54.1% to 14.6% detractor rate, RD −24.66pp [95% CI −25.43 to −23.89], hotel×severity-matched, denominator 6,076/86,292, placebo-tested 45+ SDs from noise. Residual scar: even after a full, guest-confirmed fix, 20.4% still won't return (CI 19.4–21.5%) and 14.6% still turn into detractors (CI 13.7–15.5%), against a 5.5%/1.9% no-friction baseline, a 14.9pp/12.7pp gap that resolution narrows sharply but never closes.
Who this is for

Universal. This line belongs in every operator report, every acquisition file, every renewal review. If a reporting stack tracks that a case closed but not how much the guest had to push to get there, this is the gap it is hiding.

Does your own reporting stack track how hard a guest had to fight for a fix, or only whether the ticket eventually closed?

Finding 02 · Capital-allocation verdict

Whether a hotel owns its own demand and whether it needs urgent repair are two completely unrelated questions. About one in ten of Portugal's own curated best hotels answers both of them badly at once.

Cross a hotel's demand-ownership, does it earn its own return guests or rent them from the destination, against how urgently its building needs repair, and the two are statistically independent: knowing one tells you almost nothing about the other. That independence is exactly what makes the overlap worth naming: a hotel that doesn't control its own demand AND needs money spent on it right now is a specific, sizeable, and currently unscreened archetype.

The fragile quadrant
9.8%
renting its demand AND needing urgent repair · k=48/490 · CI 7.5–12.7%
Carries a recognized badge
10.4% vs 64.6%
fragile quadrant vs. the mirror-opposite, safest quadrant · Newcombe delta 54.2pp [35.8–67.5]
Sample details & methodology
Demand-ownership (Property-Pull Index, empirical-Bayes-shrunk) and capex-urgency rate: Spearman rho=−0.081 (p=0.072, n=490), statistically independent axes. Fragile quadrant (bottom-tercile property-pull, top-tercile capex-urgency): 48/490=9.8% [Wilson 7.5,12.7]. Authority-tier badge composition: 10.4% [4.5,22.2] of the fragile quadrant (5/48) vs 64.6% [50.4,76.6] of the mirror-opposite robust quadrant (31/48), Newcombe delta 54.2pp [35.8,67.5]. A third corner, "owned demand, worn building" (47 hotels), carries authority-tier recognition near the population base rate (29.8%), the reminder that a badge buys out of demand risk more reliably than it buys out of capex risk.
Supporting: the capex-sequencing rule this finding used to lead with
Read the guest's own words for urgency rather than aesthetics: reviews flagging an urgent, functional problem turn guests against a hotel at 78.7% (CI 78.1–79.4%, k=11,799/n=14,988), reviews flagging only a cosmetic issue at 4.2% (CI 3.7–4.8%, k=208/n=4,953), an 18.7x gap confirmed on every platform independently, floor gap 63.4pp (Booking) to 83.6pp (Google). Correlational, not a causal claim that a cosmetic fix has zero effect. Still the right capex-sequencing rule (fund urgent-functional before cosmetic); demoted here because the finding is not news to a PIP-literate reader, not because it stopped being true.
Who this is for · owner, PE, asset manager

A second, independent risk screen alongside Finding 04's triage flag: this one asks specifically whether an asset's demand base and its capex needs are compounding each other, not whether either alone crosses a threshold. An unrecognized, unbadged hotel sitting in the fragile corner is a materially different underwriting case than a badged asset in the same corner, and the badged case is six times rarer.

Of the two assets on your own desk, which one is carrying both a demand base it doesn't control and a building that needs money now, and would your current diligence process have caught the combination?

Finding 03 · Loyalty structure

Ask a guest why they'll come back and the destination gets some of the credit. Ask why they won't, and it barely gets blamed at all.

Guests who explicitly say why they won't return blame the quality of the stay itself, not the price, not a competitor, not the location, at every single one of 384 hotels tested nationally, no exceptions. If a guest is walking away, the review almost never says "too expensive."

Quality
85.99%
the stated reason guests won't return · CI 85.74–86.24% · k=63,246/n=73,550
Price
5.34%
a distant second, ahead only of competitor (3.78%) and distance (2.07%)
Sample details & methodology
Quality is the plurality-or-majority stated non-return reason at every one of 384 qualifying hotels (hotel-level minimum share 62.7%), replicating on all four platforms independently (84.5–88.9%). Disclosed methodology risk: "quality" could partly be an easy catch-all bucket relative to eleven narrower ones; per-hotel-minimum and four-platform-replication checks argue against a national-pooling artifact but can't fully rule out a prompt-design one.

Loyalty status itself doesn't change the deal either way. Fixing a returning guest's problem is worth exactly the same, in percentage points, as fixing a first-timer's, not more, not less: two opposite management theories (loyal guests forgive faster; loyal guests judge harder) both turn out wrong. What is real: a guest who says they've stayed before starts a little more forgiving before anything even breaks, which argues for making sure a first-timer's problem gets the identical fix, since they have less goodwill banked if it doesn't land.

Supporting: who gets the credit when a guest does come back
Ask a returning guest at a branded hotel why they will come back, and the property itself gets credited 78.2–78.3% of the time; the brand loyalty program specifically gets credited 0.9% of the time (property beats it 88.3:1, brand-program-alone isolate). Checked a second way, in revealed behavior not stated preference: tracking guests across two consecutive stays at the country's biggest domestic hotel groups (n=109,829 visit pairs), guests aren't rebooking the same brand any more often than chance would predict at two of the three largest domestic full-service groups; international chains show the opposite. A stated preference and a tracked behavior, from two different guest populations, land on the same answer: guests are rebooking the building, not the badge. Two ratios coexist depending on scope: 88.3:1 isolating the brand program alone, 29:1/29.5:1 pooling every brand-attributed cause; state which. Holds on all four platforms, ratio range 20–52:1.
Who this is for · owner, PE · brand-affiliation and retention decisions

Relevant to any underwrite carrying a brand flag, any brand-affiliation renewal decision, and any service-recovery program deciding where to spend effort first.

If a guest of yours is walking away right now, does your own exit data actually say why, or has everyone already assumed it was the price?

Finding 04 · Pre-LOI triage screen

The hotel to worry about on this screen usually isn't the one whose rating looks a little too good. It's the one its own repeat guests, head-to-head, would rank higher than the world does.

Companion capex read, same screen: hold the rating flat at 9.00 and the share of a hotel's complaints that's actually a capex problem, not a training issue, still spans a 5.7x range between the two hotels closest to that exact score.

Run every hotel through the same six-axis risk screen and cross it against a second, independent measure: what a hotel's own returning guests say about it, guest against guest, not the blended public rating. Hotels the public rating flatters show zero elevated risk on the screen. Hotels the rating quietly underrates flag red roughly three times more often than average.

Two separate methods, a same-guest tournament and an unrelated cross-platform percentile spread, built independently, agree closely on which hotels these are.

Underrated hotels
42.3%
flag RED on the screen · k=11/26 · vs 14.8% base rate (k=48/324)
Flattered hotels
0.0%
flag RED · k=0/25 · the opposite of the intuitive "too-good rating hides risk" read
Sample details & methodology
Underrated hotels (same-guest tournament ranks the hotel ≥70 positions better than its blended rating credits it): RED 42.3% [Wilson 25.5,61.1], k=11/26, vs 14.8% [11.4,19.1] among the rest, Newcombe delta +27.5pp [10.2,46.6]. Flattered hotels (rating ranks the hotel ≥70 positions better than its own guests' head-to-head record supports): RED 0.0% [0.0,13.3], k=0/25. Cross-validated against an unrelated method: Spearman(rank_gap, cross-platform percentile delta) = 0.577 (p=1.0e-34, n=375). Killed explicitly: no enrichment of review-manipulation flags in the flattered set. Companion capex read: within the certified 8.90–9.10 rating band, capex-fixable share of friction ranges 4.9–62.7% across 75 hotels; two hotels within 0.03 points of an identical 9.00 span 10.9–62.7%, a 5.7x range at essentially zero rating gap.
Supporting: the base rate and the deal-tracker gap
14.6% of the 500 hotels flag RED on the screen (CI 11.8–18.0%, k=73/500, v1.2-canonical, E3-corrected roster). Of the 72 trade-press-verified RED hotels on the pre-correction roster, 91.7% (66/72) show no discoverable 2024-26 transaction event, the complement of an 8.3% documented-window cut (k=6/72); six hotels sit inside a transaction, renewal, or concession window independently verified in trade press, which is the tell that the screen catches something real, not noise; the other sixty-six aren't on anyone's deal tracker, which doesn't prove something's wrong, only that whatever the guests are already saying hasn't reached a deal desk yet. One additional RED hotel from a later roster correction not yet cross-referenced. The screen checks six things at once: complaints going unaddressed, urgent-repair language, safety mentions, a rating that's masking the real picture, guests quietly leaving for good, and a worsening trend in how many guests turn against the hotel. A relative, within-elite-set percentile screen, not an absolute distress threshold. A competing 76/500=15.2% figure is a compilation error, never reproducible, never cite it.
Who this is for · PE, asset manager, advisory / DD

A pre-screen for a pipeline of acquisition or renewal candidates, and a differentiation point against a transaction-tracker subscription: this surfaces a population that kind of tool cannot see, and specifically redirects attention away from the instinctive "that rating looks suspiciously good" read toward the hotel its own guests are quietly telling you more about than the blended score shows.

The last time a rating looked too good to be true on your own shortlist, did you flag the flattered hotel or the underrated one, and would this screen have told you which one actually carries the risk?

The rest of the release, ten seconds

Seven more findings, one line each.

01
78.4% Named-rival comparisons: the rival wins about three times out of four. Food is the worst dimension for it (12.5% win rate), location the only near-parity one (48.4%). One brand-specific edge inside this same pattern: name a hotel's own sister property under the same flag, not an outside rival, and it loses even more often, 86.85% of the time versus 79.14% for an unrelated rival.
02
goes together Hotels that own more of their own demand also tend to have guests writing about real emotional highs, and naming staff by name, more often. One thing about that pattern doesn't fit yet, and we're not smoothing it over: resolved complaints go down slightly as ownership goes up. We don't know why. We're saying so anyway.
03
10.1–12.5% One named staff member is behind a median share that large of a hotel's warmest reviews, at 389 of 500 hotels.
04
69.8% Guests complaining about an unanswered message or a vague confirmation turn against the hotel more often, and call the whole stay overpriced more often, than guests complaining about WiFi, the AC, or breakfast. It's the largest gap of the five cheap-fix categories tested; WiFi is the smallest.
05
44.1% At the priciest quarter of PT-500 hotels, 44.1% of guests say they'd recommend the place, against 24.2% at the cheapest quarter, but overpriced complaints climb right alongside it (40.8%→61.7%). Raising the price doesn't go unnoticed, and it doesn't only backfire.
06
1 in 19.3 A guest who reports being genuinely hurt at a hotel, a fall, a cut, food poisoning, still hands out a literal perfect 10 about one time in nineteen, at 78 hotels nationally. The rating system works almost everywhere. It just goes quiet at the very top.
07
0 of 111 When a Portuguese hotel guest actually commits to leaving, a same-trip walkout or a stated future booking elsewhere, they never once name Airbnb as the destination, and nearly one in five name a specific rival hotel next door.
How the investigation was scoped

Built to read the whole country. Not a sample of it.

Seeded from recognition programs the market already uses to identify quality. Filtered to hotels only. Collected across all four platforms simultaneously, so no property's guest record is read from only the platform where it happens to have the most reviews.

01Seed

Recognition, not review count

Hotels were selected from recognition programs that identify quality: Michelin Keys, Forbes Travel Guide, Leading Hotels of the World, Relais & Chateaux, Small Luxury Hotels, and equivalents. Guest volume was not a criterion; distinction was.

02Filter

Hotels only, entire country

Apartments, hostels, serviced residences, and non-hotel accommodations were excluded. The working set covers Portugal north to south: coastal resorts, urban luxury, wine-country estates, historic palaces, filtered to hotels only.

03Collect

All four platforms simultaneously

Booking, Expedia, TripAdvisor, and Google. Every property matched and collected across all platforms at once, so the guest record is complete, not a sample of wherever most reviews landed.

04Investigate

The investigation, not a dashboard

Structured extraction across 40+ fields in 25 to 26 languages is the infrastructure. On top of it: every angle run, every non-obvious hypothesis generated, every candidate finding tested adversarially before it ships. Findings that survive that process are the only ones that ever get published.

If a hotel you are watching sits in this set, how much of its guest record would you be missing by reading only the one platform it happens to have the most reviews on?

How the investigation works

Every angle run. Every finding earned. Nothing here made the cut by accident.

A fixed pipeline aggregates what it was told to label. This investigation starts from the full public record and finds what no label had a bucket for. That is why it surfaces findings no rating, no review-monitoring tool, and no operator dashboard would have queued.

01Generate

Every angle run against the full record

Not just the angles the pipeline was told to check. Every non-obvious hypothesis the data could support is generated: value signals, loyalty patterns, friction mechanisms, quiet-decline markers, rival-leak. The record is the starting point, not the filter.

02Attack

Every candidate finding tested adversarially

Before a finding is closed, every alternative explanation is run. If there is a simpler story in the data that explains the same pattern, it is found and the candidate is killed. What survives is what no simpler explanation could account for.

03Ship

Only findings that close with a mechanism

A finding does not ship as a rate. It ships with the mechanism: why the pattern exists, what the lever is, and what decision it changes. Read from what guests actually write, not reservation records: a guest's own words about coming back, or never again, are not the same thing as a confirmed booking.

The one we caught

A national loyalty "collapse" that turned out to be how the data was collected, not guests leaving.

A pooled national trend looked damning on first read: the share of guests who'd recommend the hotel fell from 37.8% in 2019 to 26.1% in 2023, an apparent 11.7-point collapse, large enough to flag 64% of the 344 hotels we could test as declining. Checked platform by platform, the story fell apart.

What we almost shipped
Portuguese hotel guest loyalty is collapsing nationally.
37.8% → 26.1%

Full-year 2019 versus full-year 2023, the pooled national share of guests who'd recommend the hotel, unadjusted. An 11.7-point fall, on its face a market-wide decline.

n=80,800 (2019) · n=223,202 (2023) · naive pooled recommend-rate trend.
What actually held up
Flat the whole time. A scrape-onset platform-mix artifact, not a decline.
~30–34%

Booking barely had any reviews in this record before 2022, and Booking's own scoring counts fewer reviews as a recommendation than any other platform does. As Booking's share of the sample grew, the pooled number fell on its own, even though nothing about how guests actually felt had moved. Correct for that platform mix, and the real trend across these 500 hotels is flat the whole time, roughly 30 to 34%, for the entire 2019–2026 window.

Method + denominator

Platform+seasonal two-way fixed-effects adjustment, re-estimated on the full 500-hotel corpus, not just the eligible sub-panel. Confirmed via 100% non-Booking platform-subset replication and a shuffled-month population placebo (null result).

No competitor selling review analytics publishes the number they almost shipped and didn't. This is the one we can point to. What's the last statistic your own team almost shipped that a second look would have killed?

See the same generate, attack, ship process on the finished study, Study 01 →

Honest about the edges

What this release does not cover.

The same habit that caught the false loyalty-collapse story above applies here too. Four limits, stated plainly.

No per-property score or ranking is public. The locator above returns existence facts only, never where a hotel sits in its set.
Regional and named-portfolio detail, which cities, which groups, stays warm-private, not published here.
This is a dated snapshot, not a live feed. Re-reads are planned on a periodic cadence, not continuously.
Expansion beyond Portugal is not scheduled.
Method, perimeter, and what this study will and will not claim
Will claim
Rates and proportions with 95% CI and stated denominators on every published number
Rates = share of mentioners (guests who raised a topic), never a share of all reviewers
Only findings that cleared a pre-set statistical threshold; killed candidates are disclosed, not hidden, PT500-M71 above is one of them
Any hotel in Portugal can be read from the outside: pre-bid target, rival, or asset under review
Will not claim
No hotel is ranked against any other. A loyalty signal from public reviews is not a quality verdict or consumer award
That findings from this cohort apply to any hotel not in the dataset, or predict individual property outcomes
That loyalty is confirmed: it is inferred from what guests chose to write, never from booking data
That this cohort substitutes for a bespoke engagement. A paid read builds a peer group specific to the named hotel
Inference
Loyalty is read from what guests chose to write. It signals return intent, never confirms a return happened. Reviews skew toward the delighted and the furious; the analysis accounts for that, but does not correct it away.
Selection
Recognition programs (Michelin Keys, Forbes Travel Guide, Leading Hotels of the World, Relais & Chateaux, Small Luxury Hotels, and equivalents) plus highest-rated properties on platform, hotels only. 500 of Portugal's recognised and highest-rated hotels: 141 authority-recognized, 359 highest-scored.
Source
Public reviews only, collected under GDPR Art. 6(1)(f) (legitimate interest). No hotel-systems data, no operator cooperation, no first-party guest information. No property was asked to participate.
Cadence
First release published 2026-07-10. Findings refresh on a periodic re-read cadence, not continuously; each re-read is adversarially tested again before anything changes.
Not a benchmark
This cohort is not the set a paying client is compared against. A paid read builds a peer group specific to the named hotel, its tier, and its competitive market.
From finding to decision

The national index tells you where you sit. The paid read tells you why, and what to do about it.

Same process, one hotel, 14 days, no operator cooperation required: why its guests signal return, why some flag the stay as overpriced even at a high rating, which friction pattern is driving that gap. Every angle run. Every surviving finding closed on a decision.

This is not a subscription the hotel buys to watch itself. It is a finished, outsider read built once for the person deciding on the asset.

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.

Which named hotel would you want read first?

Book a 20-min intro → See Study 01, published →

Fast triage: a red/yellow/green verdict on a named asset, 48 hours →

A quicker first look at one named hotel, before you commit time to it: fine, worth a closer look, or a genuine problem.

One-off per engagement, no subscription, no onboarding, no recurring commitment. Scope and price agreed on a short scoping call.

Want the next wave of findings the day they land?

You're on the list. New findings come to you first.

Email collected under GDPR Art. 6(1)(a) (consent). Used only to notify you when new findings publish. No further contact without a new request. See the privacy policy.

Book a 20-min intro →