Methodology: historical destination review and linked readiness analysis
Version H1 final, October 5, 2026. Scope: protocols 9, 10 and 11 only.
1. Frame and original collection
Candidates came from the pinned free Overture Maps Places 2026-08-19.0 extract, the retained mapping and seed readiness-us-2026-09, and original selection ranks. The forward batches excluded previously attempted initial domains, including the unrelated original 5,000 and all inherited protocol 7 and 8 attempts. Protocol 10 additionally excluded protocol 9 attempts; protocol 11 additionally excluded protocol 10 attempts. Failed and pending-retry initial domains were included in these exclusions.
Historical collection windows for the finally selected representatives were September 30 to October 1 for protocol 9, October 1 to October 3 for protocol 10, and October 4 for protocol 11. Qualification rules changed across these batches; scorer version 2 remained unchanged. Full prior source manifests and forward plans are retained privately.
All 11,423 original r9/r10/r11 result rows remain immutable. Their shared production identity identified ten repeated final destinations, leaving 11,413 historical destination representatives. Historical representative selection used earliest observation, then cohort and sample rank. The review queue contains 5,000 r9, 4,990 r10 and 1,423 r11 representatives. It was processed shortest historical text first, not randomly. The separate r12 pilot and unrelated original 5,000 are outside the analysis.
2. Complete current qualification pass
The immutable queue was joined to the frozen source to recover the recorded initial targets internally. The existing guarded fetch path followed their current redirects and assessed the resulting homepage. This does not prove that a current destination or response is identical to the historical response.
H1 retains protocol 12's minimum 80 qualifying parsed body-text characters and bounded frame, holding, cancellation, maintenance and setup-template rules, plus the exact whole-body service-provider error rule documented before the full review. Unrecognized surrounding business prose prevents that exact exclusion. The local sidecar did not change production qualification, migrations or scores.
The full pass ran October 5 from 03:33:58 to 06:34:00 UTC, concurrency two, with 11,413 unique starts and results and no automatic retries. Short text below 500 characters, replacement characters, generic/template concerns and cross-domain redirects triggered review rather than automatic rejection. The minimum content requirement stayed 80.
Robots, pinned DNS, SSRF restrictions, timeouts, maximum 32 probes and 3 MiB per execution, and existing encoding restrictions remained in force. No JavaScript or CSS execution or new decompression was added. A page unavailable to this bounded fetch can still work for a human browser.
3. Finite follow-up and adjudication
A predeclared second attempt covered 554 transient network, non-success HTTP or ENOTFOUND cases from the 574 initially inaccessible inputs. Sixteen robots-denied and four unsafe/unresolved cases were not retried. Follow-up ran once per eligible input at concurrency one from 08:05:20 to 08:39:23 UTC. This is not a retry-until-success process.
Individual content adjudication ran at concurrency one, combined network concurrency at most two while access follow-up ran. It covered 875 inputs from 08:02:00 to 09:02:22 UTC. Evidence was ephemeral, redacted and bounded to the first 1,000 text characters, with numeric diagnostics. Definite business, service, product or contextual navigation prose can support inclusion. Generic holding, setup, maintenance or error-only responses are excluded when the evidence supports that interpretation. Ambiguous truncated or inaccessible evidence remains inconclusive.
Short pages were not discarded merely for being short. Closure or relocation, non-English text and navigation-dominant business content are not inherently disqualifying. This is not a verification of legitimacy, active trading, US presence, independent ownership or source-sector accuracy. No manual claim covers all 10,777 pages: 9,588 inputs passed the automatic screen without flags; 450 redirect-only cases were reconciled from numeric evidence; 740 passed individual content adjudication.
Each input has a separate numeric result and decision history where relevant. No target was silently substituted. All flags are processed, but 530 inputs remain inconclusive: 460 access-related and 70 adjudication-related. They are not counted as qualified, failed businesses or zero readiness scores.
4. Current destination identity
Current final registrable domains use the existing tldts normalization with private domains enabled and a separate private local HMAC key. No production salt or fingerprint was retrieved. Hashes and key material are not included in public reports.
Every eligible current destination is counted once across all three cohorts. Mixed eligible/ineligible groups would be withheld; none remained. One additional eligible input collided at the current destination. Representative selection was lowest original protocol number, then lexical sample rank. Final accounting is 10,777 unique qualified destinations, 105 excluded inputs, 530 inconclusive inputs and one additional eligible duplicate, totaling 11,413.
Domain uniqueness is not business uniqueness. One business can own several domains; several businesses can share a normalized domain. Redirects and content can change after the review.
5. Readiness metrics and interpretation
A read-only database query on October 5 at 09:04:11 UTC selected the historical rows corresponding to the final current-qualified representatives. Its count matched 10,777 exactly. The saved aggregate preserves protocol-specific denominators of 4,703, 4,724 and 1,350. No pooled readiness mean or cross-protocol trend is calculated.
These are historical scanner observations, not an October 5 rescore. Current endpoint qualification cannot establish that historical prices, signals or business identity still apply. Representative linkage preserves provenance, not proof of unchanged content.
Definitions matter:
- Visible price and availability signals are scanner matches in fetched HTML-derived text, not rendered visibility or verified prices/inventory.
- An action path is a qualifying link or form target, not a successfully completed booking or purchase.
- Parseable JSON-LD means at least one JSON-LD root parsed as JSON, not that its schema is accurate or commercially useful.
- Structured price requires an Offer-type node somewhere and a nonempty recognized price field somewhere among collected nodes; those conditions need not occur on the same Offer.
- Structured action recognizes specified Action targets or Offer URL/action fields; it does not execute them.
- Structured availability can be an availability field, opening hours or delivery lead time. It is not necessarily live inventory.
- llms.txt passes only a root non-HTML, non-JSON Markdown-H1 format check. It is not evidence of useful guidance or AI consumption.
- MCP and OpenAPI checks are shallow JSON-shape detections at fixed paths. They do not prove working servers, authenticated access or executable transactions.
- Zero agent-file detections apply only to the exact paths and checks used, not all possible integrations.
6. Bias, limits and reporting rules
This is not nationally representative, a simple random prefix, a verified US small-business census or a direct trend comparison with the earlier 652-site study. Source errors, source-category mismatches, repeated forward exclusions, ordering, accessibility, robots rules, retries, destination collisions, qualification changes and incomplete frame coverage affect inclusion. Current qualification also selects on later availability and content. More observations do not remove these biases.
The entire historical coverage and all original scores remain preserved, including excluded and inconclusive inputs. Known predicate weaknesses in old cohorts are not population error estimates. The present full review supersedes sample-only conclusions for this review set, without retroactively certifying old gates or silently rewriting observations. No remaining records were validated merely by subtracting a few known defects.
Confidence intervals would not repair these design limits and are not presented. Sector rankings are omitted because source labels are unverified. Qualification is specific to the review time and method; it is not a permanent certification.
7. Reproducibility and privacy
The final reconciliation verified frozen source and helper checksums, unique request and result records, per-request bounds, completed decisions and current destination deduplication. Fourteen offline rule, classification and deduplication tests passed. The aggregate export matched the selected 10,777 historical representatives.
Private selection records and decision evidence are retained separately. Fetched bodies, title text, headers and raw destination URLs are not included in the retained review outputs or these public reports. No original result or score was rewritten. Only aggregate figures are published.