Methods and caveats
Data and methods for the agent readiness study
Five evidence layers answer different questions. Their sample sizes, methods, and claim boundaries remain separate.
Method in brief
The primary layer is a crawl of 2,642 unique Japanese lodging domains sourced from association directories. Each domain was crawled in an aggressive run and a compliant run. The compliant run is the primary public measurement.
Separate layers cover a curated 80 target mountain hut hunt, two qualitative fidelity cases, seven supervised booking cells, and one documented booking task across three agent sessions. Results from those layers are never blended into the primary crawl percentages.
Reachability crawl
N = 2,642 domains. Honest user agent, certificate verification, robots rules, one second delay, ten path limit, and 700 KB page limit.
Mountain hut hunt
N = 80 curated targets. Two discovery passes, including regional portal assistance. This is not a random sample.
Fidelity cases
N = 2 worked cases. The evidence supports a phenomenon, not a market rate. Detailed transcripts remain internal.
Booking interactions
n = 7 supervised cells in the initial layer, with later replication specimens. No booking was submitted and no percentage is published.
Narrative anchor
One booking task across three sessions. It is a documented experience, not a statistic.
The compliant crawl
The universe includes association listed properties with a published official website. Known travel marketplaces, social profiles, aggregators, and chain corporate domains were excluded during collection. This creates an optimistic selection bias because properties without a published website never enter the denominator.
The crawler recorded fetch outcomes, the first page where an email appeared, contact forms, phone links, LINE links, structured data, and recognizable booking engine fingerprints. Engine detection measures presence, not agent usability.
What this study does not claim
- No consumer AI recall percentage for named properties.
- No statistics from the retired 100 property depth frame.
- No claim that a named property cannot be found by an agent.
- No named vendor conclusion from private negative probes.
- No booking completion, uplift, leakage, or revenue rate.
- No claim that proxy measurements represent consumer AI behavior.
Reproducibility and public data
The public downloads contain aggregates only. They exclude property rows, domains, email addresses, contacts, internal identifiers, and raw session logs. The methods and correction record travel with the data.
The aggressive and compliant result snapshots were generated on July 23, 2026. The compliant snapshot supersedes earlier headline conclusions where the two conflict.
Data use and citation
The published aggregate values may be quoted with attribution to Spann and a link to this methods page. Publication does not grant access to property rows, contact data, raw logs, or internal research artifacts.
A citation should state the exact variable, denominator, collection period, and that the universe contains association sourced properties with a published website.
Questions and boundaries
Methods questions
Why were the sites crawled twice?
The first run used an aggressive crawler as a ceiling. The second used an honest user agent, certificate verification, robots rules, a page limit, and a delay. The compliant run is the primary published measurement.
Why are the sample sizes kept separate?
Each layer answers a different question. The 2,642 domain crawl supports market level reachability findings, while the mountain hut, fidelity, and booking layers support bounded qualitative observations.
What was corrected?
An early proxy inferred that only 16.1 percent of found emails appeared on the first page. Exact compliant crawl measurement found 52.7 percent. The earlier one in 26 composite is withdrawn and the corrected figure is one in 9.