Agentic Substrate Readiness

AI tools directly observed on-site — chatbots (Ada, Tidio AI, Chatbase, Botpress, Voiceflow, Dialogflow), AI content tools (Jasper, Copy.ai, Writer), AI personalisation (Dynamic Yield), AI search (Algolia AI), AI summarisation widgets.

API-first / headless / composable architecture, webhook infrastructure, OAuth or API-key surfaces for delegated agent identity, modern CMS with documented API access. (MCP manifests are scored separately below, not double-counted here.)

Workflow orchestration depth — detected automation platforms (Zapier, Make, n8n, Workato, native automations), AI-assisted automation patterns, evidence of multi-step orchestration.

Integration density between detected tools — the more tools speak to each other, the more agent-operable the stack overall.

The same llms.txt file is scored twice on different questions. Here it is read for agent-operability — does it help an agent understand the site structure and navigate it: sectioning quality, key pages indexed, a navigation path an agent could plan from the file alone. This reading can score high even when the citation-lens reading under Discoverability scores low. Same file, two scores.

A manifest at /.well-known/mcp.json or /.well-known/ai-plugin.json declaring tools and endpoints an agent can call. Presence is a strong positive signal; absence is the modal default for businesses today and not by itself a finding, but it forecloses downstream agentic integration paths.

Browser-native WebMCP tool registration — declarative HTML tool attributes on form/button elements, or imperative navigator.modelContext.registerTool calls (audited by Chrome Lighthouse 13.3+). The declarative API is detectable via static scrape; imperative runtime registration is not yet captured in our pipeline. Modal default in 2026 is absence — the spec is experimental — so absence is not a finding by itself.

schema.org / JSON-LD presence, validity, and coverage on the things that matter — Organization, Article/BlogPosting, FAQPage, HowTo, Product, Service, LocalBusiness, Event. Answer-relevant types are materially more likely to be extracted into AI-generated answers.

Explicit allow/deny posture in robots.txt for AI bots — training bots (GPTBot, Google-Extended, ClaudeBot) versus retrieval bots (ChatGPT-User, PerplexityBot, OAI-SearchBot). A differentiated posture (training blocked, retrieval allowed) is a valid strategic choice that scores positively. Retrieval bots blocked caps AEO regardless of other signals — no fetch, no answer-extraction.

Whether content is shaped to be lifted directly into an answer — clear question-and-answer formatting, specific verifiable claims, stable canonical URLs, content addressable by stable identifiers. Hedged or buried facts extract poorly.

The structured summary file read here for its citation/answer-extraction value only — does it help AI systems extract and cite you. On current evidence (Zyppy 2026; Google’s own guidance) llms.txt provides minimal-to-no measurable citation lift today, so most sites score low here regardless of file quality. The SAME file is scored separately under Tech Stack for agent-operability, where it can score high. Same file, two questions, two scores.

The GEO sub-score, holistic and off-site: third-party mentions, earned media and trade-press coverage, analyst and industry citations, presence in authoritative reference sources (Wikipedia, Crunchbase, directories), and dedicated AI-visibility tooling (Obsero, Otterly, Ayzeo) if detected. This is the "get named and recommended" axis, distinct from AEO’s "get into the answer."

Forms accessible without captcha-by-default; no hostile JS gates on essential conversion paths. Targeted captcha on high-risk forms is appropriate; blanket captcha on every form is hostile to legitimate agents. Sparse-homepage-form sites where the conversion model is sales-assisted or hybrid (with a dedicated /contact, /account-opening, or /quote path) are healthy patterns and not penalised — many regulated industries deliberately keep homepage forms minimal.

Cloudflare Bot Management, DataDome, PerimeterX, Akamai Bot Manager, Imperva, Kasada visible in headers or cookies. Presence is neutral-to-positive — it means the business has capability for differentiated bot policy. Absence on a high-stakes funnel (e-commerce checkout with no bot management) is a finding.

Form submissions persist into a system of record (POST destination is a system endpoint, not fire-and-forget mailto:); booking/order pathways have stable URLs; checkout returns addressable confirmation pages.

Calendly / Cal.com endpoints, programmatic checkout (Shopify Storefront API, Stripe Payment Links), exposed booking systems. Optional / positive-only — many businesses don’t have programmatic conversion paths and that’s not a finding for non-API-native businesses.

Can an autonomous agent reliably identify and complete each field of the conversion form? Label associations (label[for]/wrapping label/aria-label), autocomplete hints (autocomplete="email"/"name"), and the absence of the placeholder-only anti-pattern — placeholder text is a hint, not an accessible name, and disappears once the user types. High label coverage plus autocomplete on identity/contact fields means an agent can fill the form deterministically.

Does the privacy policy disclose AI usage and named AI vendors? When the audit detects a chat or AI-content vendor (Intercom, Tidio, Jasper, Copy.ai, etc.) but the privacy policy doesn’t name it, that’s a real disclosure gap. Strong AI/ML governance language — automated decision-making references, sub-processor disclosures, named third-party AI providers — scores positively.

Does the AI widget respect consent gating? When a chat widget loads BEFORE consent fires for jurisdictions that require it (PIPEDA, GDPR, UK GDPR), that’s real exposure. Consent-aware widget loading scores positively. When no chat or AI widget exists at all, this dimension is treated as un-evaluable rather than penalised.

Has the business addressed automated agent access in robots.txt and Terms of Service? Robots.txt directives for AI training and retrieval bots PLUS explicit ToS clauses on scraping, crawling, or text-and-data-mining is the strong-posture pattern. Either surface alone is partial.

Same evidence as the Conversion pillar’s anti-bot posture, framed differently. For Conversion: differentiated funnel-policy capability. For Trust & Security: can the business OBSERVE agent traffic at all? On a regulated entity, absence of any anti-bot infrastructure is a sharper finding here than at the funnel level.

AI-generated content disclosure markers detected in page HTML — attribute markers like data-ai-generated, AI-content meta tags, visible-text disclosure patterns ("Generated with AI", "AI-assisted"). C2PA binary metadata is not currently parsed. Absence is the modal case today; absence is only a finding when the business has detected AI content tooling without disclosure.