What the score is
HRI 2.0 is an index of independently recognised hotels. A hotel receives a score when at least one independent body has recognised it: an anonymous inspection guide (Forbes Travel Guide, Michelin Keys), an academy vote (The World's 50 Best Hotels) or an industry award (World Travel Awards). Hotels without such recognition are published with their facts, their sources and their dates, and without a score. 2,656 of 118,493 published hotels carry a score today.
Nothing anybody buys changes a score, a medal, a percentile or a ranking. Claim, corrections, photographs and the reply to a score are free. Membership of a paid network is shown on the hotel page and enters no part of the score.
How the score is made
Two components, then confidence, then calibration.
| Component | Weight | What it measures |
|---|---|---|
| Expert recognition (EXP) | 0.70 | The prestige-weighted value of the hotel’s recognitions (0.6) plus the value propagated through the expert graph from hotels recognised alongside it (0.4), with the saturation and the three-year half-life of version 1.3. The prestige table is published below. |
| Facility coherence (FAC) | 0.30 | The hotel’s confirmed facilities against what hotels of its location type usually have. A facility counts as present when an admitted source says so, and as absent only when the hotel’s own correction or declaration says so. A facility nobody has stated is unknown and costs nothing. |
Facility coherence, in one sentence. Confirming a facility you have can raise the score; confirming that you lack a facility most peers have can lower it; leaving a question unanswered changes nothing.
Confidence (CONF). The composite is shrunk toward the baseline of the hotel's declared class by a confidence factor built from four inputs: how many independent families recognise the hotel (0.4286, saturating at four), whether the hotel has a Wikipedia or Wikidata record (0.2143), how many recognitions it holds (0.2143, saturating at three), and whether it belongs to a recorded brand (0.1429). A hotel with little evidence sits closer to its class baseline.
Calibration and medals. The shrunk composite is mapped to the published 20 to 99.5 scale through six anchors cut over the rated set each night (table below); the top 3 percent of the rated set carry gold, the next 7 percent silver, the next 15 percent bronze.
Shown, not scored
Independent families. The page prints how many independent families of recognition agree on a hotel (one, two or three). This count is computed and shown; it carries no weight in version 2.0, because on today's data it separates fewer than one rated hotel in twenty. It will be re-measured for weighting when more families are held.
Memberships. Relais & Chateaux and Virtuoso membership is shown as a fact. It gates nothing and weighs nothing: membership carries a fee.
The guest's voice. Version 2.0 holds no guest review data: the sources version 1.3 used did not permit storing it. Guest scores return to the formula only when a source whose terms allow storage exists, through a new deposited version.
Anchors and parameters
The nightly anchor table: p = 0 / 0.05 / 0.5 / 0.95 / 0.995 / 1 of the rated set map to HRI 20 / 35 / 62 / 90 / 97 / 99.5, with the composite value at each anchor as of the last publish.
| Anchor | Percentile of the rated set | Composite at the anchor | HRI |
|---|---|---|---|
| Top of the rated set | 100th | 0.8197 | 99.5 |
| Top 0.5% begins | 99.5th | 0.7155 | 97.0 |
| Top 5% begins | 95th | 0.5558 | 90.0 |
| Median rated hotel | 50th | 0.3671 | 62.0 |
| Bottom 5% | 5th | 0.2782 | 35.0 |
| Scale floor | 0th | 0.2402 | 20.0 |
Last publish 2026-10-26, methodology 2.0
Class baselines (tier priors) as of the last publish - the value a hotel with no evidence is shrunk toward.
| Declared class | Baseline composite | Hotels |
|---|---|---|
| none declared | 0.1500 | 63,570 |
| 1-star | 0.1620 | 2,983 |
| 2-star | 0.1740 | 20,043 |
| 3-star | 0.1740 | 44,731 |
| 4-star | 0.1860 | 22,002 |
| 5-star | 0.2774 | 5,006 |
Prestige table. Version 1.3's rows for Forbes Travel Guide (5-Star 5.0, 4-Star 3.0, Recommended 2.5), Michelin Keys (Three 5.0, Two 3.5, One 2.0), The World's 50 Best Hotels (6.0) and World Travel Awards (World 2.0, Continent 1.5, Country 1.0, City 0.7); the Travelers' Choice rows of version 1.3 are removed. FAC: over-bonus +0.04, under-penalty -0.08, cap 5 each; peer cohorts by location type. CONF shares as above. EXP: 0.6 direct / 0.4 propagated, saturation k = 6.0, half-life 3 years.
What changed from version 1.3, and why
Version 1.3 read guest-review consensus from platform APIs whose terms do not permit storing or combining their content. Those inputs were withdrawn on 12 October 2026 and scores were paused. Version 2.0 is the index rebuilt on inputs HOTELSRATING has a stated right to hold: recognitions recorded as facts of public record with date and source, open-licensed data, the hotel's own submissions, and our own derived work. The full change record, the measurements behind it and the correction of record are in the deposit (DOI 10.5281/zenodo.23670609).
What this did to the numbers. Measured for the deposit on the corpus of 2026-10-21: 17,720 hotels were rated under 1.3 on the last night of 12 October; 2,674 are rated under 2.0. The hotels that left the rated set were rated on guest reviews alone. Among the hotels rated under both versions the median score moved by -22 points and 88.3 percent changed rank by more than 100 places. Medals are percentiles of the rated set, so a smaller rated set carries fewer medals. Tonight 2,656 hotels are rated under 2.0.
A score under 2.0 is not comparable to a score under 1.3. A hotel whose number fell did not get worse: it is compared with a smaller and stronger population under a formula without review consensus. Hotel pages mark the break in their history.
Validation
Nine tests run before every publish; a red test withholds the publish. Under version 2.0: T1 (hold-out of The World's 50 Best) passes with a restated bound - a winner whose only recognition is the held-out list becomes unrated, not neutral; its "top 2 percent" leg is measured over the full corpus, and the deposit states the denominator. T2 (gross class order) is skipped while fewer than 10 one-star hotels are rated and does not withhold publication; adjacent-class medians are telemetry. T3 carries the family-count expectation. T5 compares hotels with three or more families against hotels with one and prints both populations. T6 adds that no score references a review table. T4, T7, T8 and T9 are unchanged. The nine results of the last publish are in the publish log.
Sources and rights
| Source | Role | Rights |
|---|---|---|
| Forbes Travel Guide, Michelin Keys, The World’s 50 Best Hotels, World Travel Awards | Recognition: the gate and EXP | Facts of public record, per hotel, with date and source URL; removed on the body’s request with a recompute |
| Relais & Chateaux, Virtuoso | Shown only | Membership is a fact on the page; it enters no scoring channel |
| Wikipedia / Wikidata presence | A confidence input | CC0 |
| Brand affiliation | A confidence input | Operator pages |
| Declared star class | Selects the class baseline | The hotel’s own declaration or the seed import; no official registry rows in 2.0 |
| Facilities | FAC | The admitted-source allowlist; the owner’s corrections and declarations for absence |
| Location type and coordinates | The FAC peer cohort | OpenStreetMap under ODbL, credited; seed; owner |
| Platform ratings, reviews and photographs | None | None held; the dated deletion record is in the deposit |
Corrections, reply, claim
Report an error on any page without an account; claim your hotel for free to correct facts, add photographs and reply to the score. Corrections with evidence change the facts the score reads, up or down; the score follows the facts.
Personalisation
The Personalized Fit Index is defined only for rated hotels (PFI = HRI x (0.75 + 0.25 x coverage)). A personal search that includes unrated hotels orders them by matched facts and says so; no hotel is ranked by an imputed score.
Integrity Policy
Hotels cannot pay to influence a score, a medal, or a ranking position. Commercial products sold by the operator affect management tooling and data confidence inputs only through evidence, never through the formula's treatment of evidence. The full policy is the Independence Charter.
Versioning and Changelog
| Version | Date | Changes |
|---|---|---|
| 2.0 | pending the first 2.0 publish | Version 2.0. Review consensus (CRS) withdrawn; trajectory (TRJ) removed; membership bodies out of every scoring channel; the family-agreement term defined, shown and unweighted; confidence renormalised to four inputs; facility coherence no longer treats an unknown facility as absent; the gate becomes independent recognition; weights EXP 0.70 / FAC 0.30. Deposited at 10.5281/zenodo.23670609. |
| 1.3 | August 2026 | Version 1.3, August 2026. The peer group for the Delivery component (FAC) is keyed by location type alone; the declared star class leaves the key. Until now the expectation set was a function of a hotel's own declaration, so declaring a class and an inventory bought a longer list of things to be found short of rather than credit for what was there: a property that disclosed its facilities scored no better than one that disclosed nothing, both at the neutral point. The location type is derived from geography, with rivers excluded by decision, across five settings (beach, city, countryside, mountain, urban); a sixth value, other, means we hold no usable coordinates and is a geocoding gap rather than a setting. No component weight, source trust weight, prestige weight, calibration anchor, medal band or validation test changed. Because the peer key changed, published scores move: 2,943 of 5,127 rated properties change their FAC value and 3,066 move more than one hundred rank places, and 2,941 of those changes are in properties whose own declarations did not move at all. Deposited at 10.5281/zenodo.21935353 under the same concept DOI 10.5281/zenodo.20670591. |
| 1.2 | July 2026 | Version 1.2, July 2026. Validation test T2 revised: adjacent-tier coherence is retired and a gross tier-order check takes its place. The suite remains nine tests. No component weight, source trust weight, prestige weight, calibration anchor or medal band changed, and no scoring input was touched. The concept DOI printed in the version 1.1 record is corrected. Deposited at 10.5281/zenodo.21636656. |
| 1.1 | 2026-07-04 | Added Michelin Keys as an expert-recognition source in the EXP component. Michelin Keys are an anonymous-inspection hotel distinction, the same class of signal as Forbes Travel Guide, and enter the published prestige table at three tiers (Three Keys 5.0, Two Keys 3.5, One Key 2.0) alongside the existing award programs. They stack with a hotel's other awards under the same saturating formula, so recognitions add sublinearly rather than compounding. No component weight changed (CRS 0.40, EXP 0.35, FAC 0.15, TRJ 0.10 are unchanged), no calibration anchor changed, no medal band changed, and no review-consensus input was touched. Because medals are relative percentile bands, adding a new expert source re-ranks the rated corpus: some hotels gain a medal and some lose one even though their own inputs did not change. Key distinctions are read from the official Michelin Guide's public listing data. No commercial product can affect this or any signal; the prestige table is set by published editorial judgment, not by payment. This version is deposited as a new release under the same concept DOI 10.5281/zenodo.20670591; version 1.0 remains independently citable at its own version DOI. This version's own DOI: 10.5281/zenodo.21192233. |
| 1.0 | 2026-06-12 | Methodology v1.0, synchronized with the deposited specification, DOI: 10.5281/zenodo.20670592. |
| 1.0 | 2026-06-12 | Nine-test validation gate enforced on nightly recomputes: scores publish only after the full suite passes against the staged recompute; a failed test blocks publication and alerts the operator. |
| 0.9 | June 2026 | Initial public methodology: four-component HRI, Bayesian shrinkage, cross-source normalization and disagreement discount, ExpertRank trust graph, confidence shrinkage, anchored calibration, percentile medals. Trajectory component held neutral pending dated review ingestion. |
Material changes to weights, prestige values, or calibration anchors appear here with effective dates and a brief rationale. Scores recompute nightly; methodology changes do not. The 2.0 row carries its publish date from the first 2.0 publish.
Questions, Corrections, Challenges
If you believe a score is wrong, the breakdown on the hotel's page shows which part to challenge, and the claim process above is how to challenge it. If you believe the methodology is wrong, we want to hear that too: [email protected].