How consistent are hotel listings across OTAs?

Key findings

Pooled across every published city study. Pilot runs contribute to none of these totals. The comparison was corrected on 6 October 2026 and London was measured before it, so this pooled rate spans two methods and is provisional until London is re-scanned — see the limits.

A "mismatch" is a field where at least two successfully read listings stated comparable values that do not agree after normalisation. A field no listing stated, or that only one listing stated, is reported as not comparable and is excluded from every rate.

Studies by city

Each study has its own page carrying its sample, its selection rule, its per-field and per-channel rates and its anonymised examples. A run below 30 scored properties is listed here as a pilot and publishes no rate.

These two rates are not directly comparable with each other. Most of the distance between them is the comparison change of 6 October 2026, not a difference between the two cities: London was measured before wording that states no amount stopped counting as a disagreement, and it has not been recomputed because that run's raw sources are not retained. Read each rate against its own study, not against the other.

Methodology

This is the method every city study follows, stated once here rather than repeated on each one, so there is a single copy to check a published rate against.

How the study is run

1. 1Select the properties by a rule that can be restated — a city, a segment, a published list — and record the rule rather than the names.
2. 2Discover each property's live public listings on the supported channels, keeping a record of which channels returned nothing.
3. 3Read each discovered listing and extract the factual fields with a schema, keeping the original wording as evidence.
4. 4Compare each field only where at least two listings were read successfully and both stated something comparable, after normalising times, currency, booleans and whitespace.
5. 5Score each property from the comparable fields alone, and exclude any property where fewer than two sources produced comparable evidence.
6. 6Aggregate only across properties measured the same way in the same window, and publish the sample size, the window and the exclusions alongside every rate.

How a mismatch is defined

A field is comparable when at least two listings were read successfully and both stated a value for it. A comparable field is a conflict when those values still differ after normalisation — 3 PM and 15:00 are the same value, "Free cancellation" and "Non-refundable" are not. A field only one listing stated is not comparable, and a listing that could not be read is unavailable. Neither is ever counted as agreement.

The consistency score is the weighted share of comparable fields that agree, on a 0–100 scale, and is only produced when at least two sources contributed comparable evidence. Properties without a score are reported as excluded rather than averaged in.

The field definitions, the confidence threshold and the normalisation rules are the scanner's own — the study uses no separate pipeline. They are documented in the scanner methodology.

Why a study waits for 30 properties

A measured number is not automatically a publishable one. At one property, "100% of properties had at least one mismatch" is true, useless, and trivially quoted out of context, and the denominator printed beside it does not undo the headline. 30 is the point at which a share stops swinging by tens of percentage points per property. Below it a run keeps its page and its record, and publishes no rate.

Limits

• This measures content parity — whether the public listings state the same facts — not rate parity, and not whether any single listing is correct. Two channels can agree and both be out of date.
• Only channels the scanner supports are read. A property with a strong presence elsewhere is measured on fewer sources, not penalised for the rest.
• A listing that could not be read is reported as unavailable. It never counts as agreement and never enters a denominator. Some channels serve an automated bot-verification page rather than the listing, so their coverage in any run can be far lower than the number of listings found — read the coverage table on the city study before comparing channels to each other.
• Extraction is automated. A field is only compared when the extraction clears a confidence threshold; below it the value is shown as evidence but is not scored.
• Each sample is whatever its published selection rule produced. It is not a random sample of all hotels, and it does not support a claim about any market as a whole.
• Listings change continuously. A rate describes the scan window stated on that study and nothing outside it.
• Channels do not all report a field the same way, and some of what is counted as a conflict is a difference in reporting rather than in what the hotel offers. In the London sample Booking.com recorded parking as available for 59% of properties against 90-100% on the other channels, and Agoda recorded pets as allowed for 86% against roughly 44% elsewhere. Two fields were withdrawn from scoring entirely for this reason — see below — and the ones that remain should be read as what the channels publish, not as a settled account of the property.
• Two fields are shown as evidence but never scored. The swimming pool flag was withdrawn because Booking.com reported a pool for 11% of the London properties against Expedia's 78%, with no comparable gap on gym (90-97%) or air conditioning (97-100%) — the channels are answering different questions. Pet restrictions were withdrawn because they are free sentences with no shared vocabulary, where "service animals allowed" and "service animals welcome" scored as a disagreement.
• The comparison was corrected on 2026-10-06, which makes the two cities' rates not directly comparable. Wording that states no amount — "fee", "charges apply", a price quoted in a second currency, a deposit given without its basis, a cot free for one age band — is no longer read as disagreeing with a stated figure. Fifteen of the Amsterdam census's seventy-five conflicts were of that kind, and removing them moved its rate from 78% to 65%. London was measured before the correction and has not been recomputed, because that run's raw sources are not retained here; its rate overstates on the same axis by an amount we have not quantified. The pooled figure above spans both methods and should be read as provisional until London is re-scanned.

How to cite this data

Related references