What specific structured information does an AI travel agent need to confidently recommend, compare and book a specific hotel and a specific room for a specific traveler — beyond name, address, price and star rating — including room-level attributes, noise and light, accessibility, real measured Wi-Fi speeds, policies, family and business needs, and neighborhood walking times? Which of these fields are missing, wrong, or unstructured in OTA listings and hotel websites today, with named sources and concrete evidence?
This report draws on 130 source-linked evidence rows from a web scan conducted 2026-10-24: 70 rows across 17 independent hosts covering room mapping, accessibility, structured schemas and content quality, plus 60 field-definition rows across 4 hosts covering Wi-Fi, accessibility, policies and walking routes. They are evidence samples, not a census of every OTA or hotel site. The answer is a versioned, provenance-bearing graph joining traveler requirements → property → canonical room type → physical room → rate plan and stay dates → policies → measured observations → route-specific neighborhood facts, with every value carrying granularity, source, timestamp, confidence and guarantee status.
Cell number = evidence rows; hover or tap a cell for the actual field names. "Structured" counts include schema booleans (e.g. Wi-Fi yes/no) that exist but miss the needed granularity. Family fit and business fit needs appear in the taxonomy but drew no classified rows in this scan — itself a gap. 11 study-context rows carry no state and sit in the table only.
U.S. DOJ reservation rule: accessible hotels and guest-room features must be identified and described in enough detail for independent assessment, and the specific reserved accessible room must be held and guaranteed. A legal requirement — the floor, not the ceiling, for room-level data.
A documented case: the "closest" room to accessible parking in a Santa Fe hotel actually required a route about 20 times the ADA guideline because elevator routing was not represented. Straight-line proximity and property centroids are inadequate; agents need route-network distances from geocoded entrances.
A Taiwan study of economy-hotel reviews on Agoda and Booking documents brochure/actual inconsistency, reserved doubles made of two joined singles, loud or uncontrollable HVAC, weak water pressure and missing non-slip mats — facts that surface only after stays, in review prose.
FTC rulemaking and state AG actions document hotels and OTAs excluding mandatory resort, destination and cleaning fees from advertised rates. Total mandatory price, fee triggers and cancellation deadlines must be structured and attached to the specific rate plan and stay dates, not the property.
An audit of twelve LLMs (3,024 choice sets each) shows guest rating carries 35.7% of revealed importance and review volume 9.4%, while eco-certification is over-weighted (+11.6 pp) and stated reasons track revealed weights imperfectly. Reputation signals dominate; they do not substitute for traveler-fit facts — hence provenance and debiasing.
The remedy across all ten groups: field-level source (hotel/PMS/OTA/inspection/device/review inference), observed and verified timestamps, validity window, unit, locale, scope from property down to physical room and route, confidence, conflict set, and whether the supplier contractually guarantees the value.
| Group | Field | State | Concrete evidence | Source |
|---|
Method: 130 source-linked evidence rows from a web scan conducted 2026-10-24 — 70 rows across 17 independent hosts (no URL above 18.6% of rows) on room mapping, accessibility, schemas and content quality, plus 60 field-definition rows across 4 hosts (no URL above 26.7%) on Wi-Fi, accessibility, policies and walking. Matrix cells count evidence rows per field group × observed state; "structured" means a typed field exists somewhere, not that it meets the needed granularity. Evidence weight: government/legal and primary schemas highest; peer-reviewed studies next; industry audits and vendor claims labeled as such; individual cases are illustrations. 11 study-context rows without a state appear in the table only. Evidence text truncated to 260 characters for space. Samples, not a census of all OTAs or hotel sites.