The TH Analytica Framework assesses observable website signals that can support technical access, retrieval and reliable interpretation by search and AI-assisted systems.
It complements traditional SEO analysis with semantic clarity, machine-readable context, open-web presence, governance, agent-readiness and AI-visibility resilience checks. It cannot observe proprietary model internals or guarantee a downstream result.
Before scored technical dimensions, every analysis establishes the organisation's intended positioning and compares it with the positioning observable from public evidence.
Layer 0 defines the core positioning, target audiences, markets and languages, Desired Query Space, Excluded Query Space, evidence basis, expected next actions and confidence gaps. It is a mandatory, non-scored strategic gate. A strong technical score cannot compensate for an unclear or unconfirmed positioning foundation.
Quick Checks can only assess observable public positioning clarity. Full Analyses must record Layer 0 as confirmed, inferred, open or conflict and keep any uncertainty visible in downstream recommendations and prompt tests.
Prompt monitoring, content prioritisation and AI-system test sets should be derived from the Desired Query Space rather than from generic keyword or prompt volume.
Full methodology and reporting rules: positioning-foundation.md.
Technical stability of the website, accessibility and crawlability.
This dimension includes LLM Discovery / Locale Resolution. For multilingual websites, the analysis compares the canonical /llms.txt with the language variants the website actually advertises through hreflang, canonicals, sitemaps and public routes. Reachability, status, soft-fallback behaviour, content quality and available freshness headers are documented. The mere presence of llms.txt, ai.txt or another proprietary discovery file does not add readiness points.
Full methodology and reporting rules: llm-discovery-locale-resolution.md.
Presence in search engines, public web corpora and consistent external sources. A Common Crawl record is evidence of corpus presence only; it does not prove model training or live retrieval.
For local and location-dependent organisations this dimension includes the mandatory Local AI Data Sources & Entity Consistency module. Where relevant and actually verifiable, the analysis compares first-party website and structured data with Google Business Profile, Bing Places, Apple Business / Apple Maps presence, relevant map or POI surfaces and authoritative industry directories. Presence alone creates no score bonus; consistency, completeness, freshness and authority are assessed. Sources that were not checked remain not_measured and are not treated as missing.
Full methodology and reporting rules: local-ai-data-sources.md.
Visual AI Visibility is an evidence-based review within Dimensions 2 and 6, not an extra score axis. With authorised Search Console access or an export, record the availability of Google's multimodal search filter, preserve the baseline, and report visual search separately from generative-AI appearance, recommendations and business outcomes. Without data, mark it not_measured. Inspect image accessibility, precise alt text, meaningful filenames, image quality and page/entity context without claiming they are special AI ranking factors. See visual-ai-visibility.md.
Clear definition of services, topics and entities.
Legal pages, transparency, authorship and credibility indicators.
This dimension includes the mandatory Source Integrity & Evidence Readiness module for full analyses. Evidence Readiness assesses whether material claims can be checked against clear, current and authoritative evidence. External AI Source Verification is reported separately and records whether a named or linked source actually exists and supports the concrete statement. A structural readiness score is not treated as proof of external source use.
Standard full-analysis depth prioritises 10 to 15 business-critical answers or claims for source-level verification. Untested systems or claims remain not measured.
Full methodology and reporting rules: source-integrity-evidence-readiness.md.
Consistency of messaging across the website and external profiles.
How clearly the website explains what the company does.
How clearly automated systems can interpret available actions, constraints and approval steps, and whether organisation-controlled generative interfaces can communicate within defined, testable and auditable boundaries. Real-world actions remain subject to security controls and explicit human approval.
This dimension includes a mandatory Agent Readiness Signal Maturity classification whenever agent-discovery, agent-interface or agent-protocol signals are assessed. The Full Analysis must separate (1) the current specification/ecosystem status of a mechanism, (2) client-specific implementation evidence and (3) downstream outcome evidence. An established web standard does not by itself prove agent adoption, and an emerging or experimental mechanism does not earn a score merely by existing. DNS-AID is treated as an applicability-gated, diagnostic-only experimental discovery mechanism unless its authoritative status changes. ARD and UCP are assessed through the mandatory applicability rules in agentic-protocol-readiness.md: ARD for discoverable agent resources, UCP for real transactional use cases. See agent-readiness-signal-maturity.md and agentic-protocol-readiness.md.
This dimension includes AI Output & Liability Readiness for customer-facing chatbots, voice assistants and agents. Applicable assessments review approved sources of truth, high-risk claims, grounding, uncertainty behaviour, output guardrails, regression testing, traceability, incident response and human escalation. The framework distinguishes independent statements by third-party AI products from outputs of systems deployed, configured or presented by the organisation.\n\nIt also includes a mandatory AI Interaction & Data Governance applicability check. Where an organisation-controlled AI interface accepts open or semi-open user input, the Full Analysis reviews sensitive-input exposure, processing purposes, consent or legal-basis status, secondary use, advertising/profiling/model-improvement reuse, data minimisation, memory, retention, deletion, processors, transfers, security, transparency and human escalation. Every applicable assessment also performs a necessity test: if the same user outcome can be achieved with clearer website content, structured data, bounded forms or defined Agent Actions, no chatbot required is a valid result.
It also includes Agent Communication Readiness as a mandatory full-analysis module. This assesses whether an external AI agent, when it has an authorised communication tool, can identify a legitimate telephone, email, contact-form or booking/request channel, formulate the required inquiry and reach a clear human handover or verified service confirmation. Public contact data does not itself prove that a third-party agent can call or send email, and it does not grant permission for external communication.
Within that module, Voice / Call Agent Readiness is mandatory when telephone or voice contact is material to the customer journey. It assesses canonical phone data, service hours, IVR/menu navigability, AI-agent disclosure handling, consistency of prices/availability/policies with authoritative digital sources, required personal data, confirmation boundaries, cancellation/modification paths and human handover. Vendor-specific calling features are dated applicability evidence, not scoring criteria, and live call tests require explicit authorisation.
Every full analysis must generate a client-specific /.well-known/agent-contact-readiness.json draft. Contact details, permissions, response times and capabilities must be verified from authoritative evidence or remain explicitly unverified. Real phone, email, message, booking or form tests require explicit customer authorisation.
Agent Governance is a separate mandatory result inside this dimension. Agent Readiness evaluates whether an agent can interpret a possible action and its prerequisites. Agent Governance evaluates what the agent is allowed to do, under which identity, permission and approval rules, with which financial or data-access limits, and how decisions, execution, revocation and incidents can be audited. Capability is not treated as permission.
Agent Readiness without Agent Governance is incomplete. The mandatory Agentic Control & Containment review within Agent Governance covers authority, communication, goals, privilege/credential escalation, collective behaviour, human oversight, auditability, containment/revocation and technical policy enforcement. It adds no score dimension. Quick Checks inspect public declarations only; Full Analyses record per-control applicability and evidence, with operational approval blocked for affected actions when required boundaries remain unverified or fail. See Agent Governance and the July 2026 evidence case.
Every full analysis must generate or update a client-specific /.well-known/ai-governance.json draft. Published policies alone are not evidence that a control is technically enforced; unknown controls remain unverified until supported by authoritative evidence or an authorised test.
Agent Portability is a separate mandatory result inside this dimension. It evaluates whether organisation-controlled knowledge, instructions, skills, interfaces, governance rules, credentials and fallback paths are sufficiently separated from a single model provider to be reused, exported or reconnected in another compatible agent runtime. Portability reduces avoidable provider lock-in but does not imply identical behaviour across providers.
Every full analysis must identify provider-bound dependencies, model-neutral sources of truth, reusable skills/workflows, interface dependencies, governance and credential boundaries, data/export paths and fallback options. It must generate a client-specific /.well-known/agent-portability.json draft. This TH Analytica artefact is not a universal standard and does not prove support by third-party systems.
The core maturity path is: found → understood → correctly answered → contactable with clear boundaries → permitted within defined controls → portable across compatible runtimes → safely acted → auditable and revocable.
For event-related analyses, this dimension also includes the mandatory applicability module Attendance Conversion & Event Agent Readiness. It checks whether ticketing or registration systems can support reminders, attendance confirmation, explicit cancellation or place release, waiting-list reallocation and measurable no-show reduction workflows within defined identity, consent and approval boundaries. Silence or inferred intent is never treated as permission for a consequential ticket action.
Full methodology and reporting rules: agent-readiness-signal-maturity.md, agentic-protocol-readiness.md, ai-output-liability-readiness.md, ai-interaction-data-governance.md, agent-communication-readiness.md, agent-governance.md, agent-portability.md and attendance-conversion-agent-readiness.md.
How strongly observed AI visibility depends on individual source families or individual AI/search systems. Full analyses should distinguish first-party resilience from third-party dependence and flag concentration risk where one source ecosystem or one AI system dominates the result.
This dimension also includes AI Zero-Click Presence & Direct Action Measurement. It separates observed AI-answer presence, zero-click capability, structural Direct Action Readiness and independently attributable zero-click business outcomes. Traffic decline, low CTR or Search Console zero-click patterns are never treated as proof that an AI system answered the user's question. Quick Checks may assess public Zero-Click Readiness signals, but actual AI Zero-Click Presence and outcome attribution remain not_measured unless separately tested.
The assessment uses observable citations, source-bearing answers and controlled cross-system tests. It does not infer hidden ranking factors, training data or proprietary model internals.
Full methodology and reporting rules: source-concentration-risk.md and ai-zero-click-presence.md.
How clearly a physical venue can be identified and how its rules for photography, video, audio, livestreaming, visual AI analysis and biometric identification are communicated to people and compatible AI clients. Every analysis and generated file is additionally governed by the mandatory Privacy-by-Design Standard.
This experimental dimension assesses canonical venue identity, visible human notices, an HTTPS policy, Entry and Exit QR markers, bounded validity, manual override and evidence of client support. It does not claim that a QR code can universally disable third-party cameras, microphones or AI glasses.
Every full analysis must pass through an applicability gate. A relevant venue operated or controlled by the organisation activates the detailed review. If the analysed offer has no such venue, it is recorded as not applicable, and this dimension is excluded from the overall score. An unknown status remains open for human review and must not be scored.
Full specification and reporting rules: physical-ai-governance.md, full-analysis-physical-ai-governance.md and privacy-by-design-standard.md.
Future full analyses may activate an optional Natural Query & AI Intent Layer. This layer connects real first-party Google Search Console demand with semantic website analysis and separate AI-answer testing.
It is not an AI-prompt detector. Search Console data must never be used to claim that a query came from ChatGPT, Gemini, AI Overviews or another AI product.
Every analysis records one of three data states:
- GSC LIVE DATA – authorised Search Console access is available.
- GSC EXPORT DATA – the customer or its agency supplies CSV/XLSX data; this is the preferred low-friction mode where direct access is unnecessary.
- PUBLIC INTENT MODEL – no first-party GSC data is available; likely natural-language intents are modeled from public signals and must be labelled as modeled rather than observed demand.
When GSC data is available, the analysis can classify natural-language query clusters, map them to JTBD and decision intent, measure Answer Coverage and Semantic Alignment, assign a Prompt Opportunity Score, and convert priority clusters into answer blocks, FAQ improvements, service-page clarifications and controlled AI test prompts.
The analysis should also record whether the customer itself has appropriate access to Search Console, GA4, Google Business Profile, Bing Webmaster Tools and Tag Manager where relevant. Agency management is compatible with the framework, but business-critical properties should ideally remain accessible to the customer as owner/admin or equivalent.
Full methodology and reporting rules: natural-query-ai-intent-layer.md.
The framework visualizes core dimensions using an AI Visibility Radar.
The radar helps companies prioritise observable structural weaknesses. A higher assessment score does not promise that an AI product will mention or recommend a company.
Source Concentration & AI Visibility Resilience should normally be reported separately as a concentration-risk view because it is based on cross-system output observations rather than only on website structure.
The TH Analytica Framework can be used for:
• Positioning Foundation and Desired / Excluded Query Space definition before technical optimisation
• AI Readiness analysis of websites
• evaluation of AI visibility
• evaluation of source concentration and cross-system visibility risk
• AI Zero-Click Presence & Direct Action Measurement across observed AI answers, surfaced actions and separately attributable business outcomes
• Source Integrity & Evidence Readiness, including claim-level verification of business-critical AI answers
• improving structured information for AI systems
• assessing controlled AI outputs, claim guardrails and human escalation for organisation-operated conversational systems\n• assessing AI Interaction & Data Governance for open chat, voice and agent inputs, including sensitive-data exposure, secondary use, retention, transparency and safer bounded alternatives\n• measuring Recommendation Effectiveness & Robustness with baseline → intervention → re-test comparisons across relevant decision situations
• classifying Agentic-Web Signal Maturity across specification/ecosystem status, client implementation evidence and downstream outcome evidence, including applicability-gated DNS-AID diagnostics
• assessing Agent Communication Readiness, including Voice / Call Agent Readiness across telephone, IVR/menu flows, AI-agent disclosure, information consistency, confirmation boundaries, email, forms and booking/request paths with explicit authorisation boundaries
• assessing Agent Governance across identity, permissions, allowed/prohibited actions, approval thresholds, limits, auditability, revocation and incident escalation
• assessing Agent Portability across model-neutral sources of truth, reusable skills/workflows, interfaces, governance and credential separation, exportability and fallback continuity
• assessing Attendance Conversion & Event Agent Readiness across reminders, attendance confirmation, explicit cancellation / release, waiting-list reallocation, no-show measurement and governed agent actions
• optional analysis of natural-language search demand when first-party GSC data is available
• Visual AI Visibility with a documented multimodal Search Console export and asset audit, where access is authorised
• physical-place identity and visual governance for venues where the module is applicable
- Crawling is the automated fetching of public pages.
- Indexing or corpus presence records content in a search index or dataset.
- Model training adjusts model parameters using selected data and cannot be inferred from crawler access or a public corpus record.
- Retrieval selects sources at query time; a source may be retrieved without having been used for model training.
- Generation produces an answer and may still omit a technically accessible source.
Files such as robots.txt, ai.txt and llms.txt can communicate preferences and context, but support differs by provider. They do not override law, contracts or provider-specific controls. llms.txt locale resolution is a diagnostic consistency check, not evidence of downstream retrieval or recommendation and not a presence-based score bonus.
Crawler identities should be evaluated by function. Training crawlers, search/retrieval crawlers and user-triggered fetchers must not automatically be treated as equivalent.
Every indexable knowledge, service and case-study page should resolve one primary user question or decision, provide a direct answer first, define important entities and support material claims with a primary source, reproducible first-party measurement or explicit evidence label.
The module rejects rigid word-count formulas, invented ranking factors and causal claims based only on sequence. It keeps search performance, observed AI output, business actions and governance or agent-readiness signals separate.
Editorial pages should expose author and dates. Structured data and machine-readable representations must match visible content. The operational review explicitly records Answer Block Readiness (with 30–80 words as a useful diagnostic range, not a quota), Semantic Heading Clarity and Visible Content ↔ Structured Data Parity. These are clarity and consistency diagnostics, not claimed ranking factors. Files such as llms.txt or ai.txt may provide context where supported but are not universal standards or proof of retrieval.
Full controls, page template and audit checklist: evidence-first-content-standard.md.
Every full analysis reports structural Evidence Readiness separately from observed external AI Source Verification. A cited or named source is not accepted merely because it looks plausible: the source must exist and support the concrete claim. Standard depth prioritises 10 to 15 business-critical answers or claims, while untested systems remain not measured.
Full controls and reporting rules: source-integrity-evidence-readiness.md.