{
"$comment": "Source registry for Agent Lighthouse evidence dossiers. Every claim in docs/evidence cites an id from this file. accessed = the research pass that built the registry; verified = the date that URL was last resolved, taken from the dossier that cites it and falling back to the accessed date.",
"accessed": "2026-08-20",
"sources": [
{
"id": "S1",
"title": "MCP Specification 2026-07-28 — Streamable HTTP Transport",
"url": "https://modelcontextprotocol.io/specification/2026-07-28/basic/transports/streamable-http",
"type": "spec",
"publisher": "Model Context Protocol (Anthropic / MCP Working Groups)",
"verified": "2026-08-20",
"keyFindings": "Revision 2026-07-28 REMOVED the GET stream endpoint and protocol-level sessions (Mcp-Session-Id, Last-Event-ID). Server MUST expose one POST endpoint. Server MUST validate Origin; if Origin is present and invalid it MUST return 403 Forbidden. Every POST MUST carry MCP-Protocol-Version, Mcp-Method, and (for tools/call, resources/read, prompts/get) Mcp-Name headers; these are 'REQUIRED for compliance'. Header value MUST match the _meta body value or server MUST return 400 + JSON-RPC code -32020 HeaderMismatch. Unknown protocol version -> 400 + UnsupportedProtocolVersionError. Unknown method -> 404 + -32601. x-mcp-header constraints defined; clients MUST reject (exclude from tools/list) tools that violate them. Servers SHOULD send X-Accel-Buffering: no on SSE. GET/DELETE to endpoint SHOULD now return 405."
},
{
"id": "S10",
"title": "Lighthouse audit source: agent-accessibility-tree.js",
"url": "https://raw.githubusercontent.com/GoogleChrome/lighthouse/main/core/audits/agentic/agent-accessibility-tree.js",
"type": "repo",
"publisher": "Google Chrome / Lighthouse",
"verified": "2026-08-20",
"keyFindings": "Implementation is a filter over artifacts.Accessibility.violations against ~37 TARGET_RULES from axe (button-name, link-name, input-button-name, label, autocomplete-valid, aria-allowed-attr, aria-required-attr, aria-valid-attr-value, tabindex, table/definition-list rules). Binary score: any violation scores 0. Crucially it inherits axe's blind spots — axe cannot fail an element that has no interactive semantics at all, and autocomplete-valid only validates tokens that are already present, never their absence."
},
{
"id": "S11",
"title": "WebSuite: Systematically Evaluating Why Web Agents Fail",
"url": "https://arxiv.org/html/2406.01623v1",
"type": "study",
"publisher": "arXiv",
"verified": "2026-08-20",
"keyFindings": "Per-UI-primitive success rates for natbot and SeeAct. Worst patterns: slider interaction 0% for both agents; tooltip-based information retrieval 0% for both; complex form filling 12.5% (natbot) / 0% (SeeAct). Aggregate: operational actions 85.2%/76.2%, menu navigation 93.8%/81.3%, informational actions 43.8%/40.6%. Taxonomy covers click (button, link, icon button, slider, switch, accordion, dropdown menu, dialog button, snackbar), type (text/date/phone), select (checkbox, multicheck, select, datagrid row)."
},
{
"id": "S12",
"title": "Text fragments",
"url": "https://web.dev/articles/text-fragments",
"type": "vendor-doc",
"publisher": "Google / web.dev",
"verified": "2026-08-20",
"keyFindings": "Confirms a shipped answer-surface consumer: \"Clicking a featured snippet takes the user directly to the featured snippet text on the source web page. This works thanks to automatically created Text Fragments URLs.\" Support: Chrome 89+, Edge 89+, Firefox 131+, Safari 18.2+. Restates the boundary rule: \"Each of prefix-, start, end, and -suffix can only match text within a single block-level element, but full start,end ranges can span multiple blocks.\" Opt-out header: Document-Policy: force-load-at-top."
},
{
"id": "S13",
"title": "MCP Specification (latest) — index",
"url": "https://modelcontextprotocol.io/specification/latest",
"type": "spec",
"publisher": "Model Context Protocol",
"verified": "2026-08-20",
"keyFindings": "Confirms the current authoritative revision is 2026-07-28 (schema/2026-07-28/schema.ts). Lists optional extensions negotiated in capabilities: Tasks (io.modelcontextprotocol/tasks), MCP Apps (io.modelcontextprotocol/ui), Skills over MCP. Restates that annotations describing tool behavior 'should be considered untrusted, unless obtained from a trusted server'."
},
{
"id": "S14",
"title": "MCP Specification 2025-06-18 — Transports (superseded baseline)",
"url": "https://modelcontextprotocol.io/specification/2025-06-18/basic/transports",
"type": "spec",
"publisher": "Model Context Protocol",
"verified": "2026-08-20",
"keyFindings": "The legacy-era shape used for backward-compat detection: Mcp-Session-Id assigned in the InitializeResult, GET opens a standalone SSE stream or returns 405, DELETE terminates a session, Last-Event-ID resumability, and servers SHOULD assume 2025-03-26 when MCP-Protocol-Version is absent. Also documents the deprecated 2024-11-05 HTTP+SSE detection path (GET returns an `endpoint` event as the first SSE event)."
},
{
"id": "S15",
"title": "browser-use DOM extraction: enhanced_snapshot.py",
"url": "https://raw.githubusercontent.com/browser-use/browser-use/main/browser_use/dom/enhanced_snapshot.py",
"type": "repo",
"publisher": "Browser Use",
"verified": "2026-08-20",
"keyFindings": "Parses CDP DOMSnapshot for exactly these computed styles: display, visibility, opacity, overflow, overflow-x, overflow-y, cursor, pointer-events, position, background-color — plus bounding boxes, client rects, scroll rects, paint order and stacking contexts, and a CDP isClickable flag. Confirms production agents infer interactivity from cursor style and occlusion/paint order, so cursor:pointer-without-role and overlay occlusion are first-class, measurable inputs to a real agent's world model."
},
{
"id": "S16",
"title": "Playwright MCP server",
"url": "https://github.com/microsoft/playwright-mcp",
"type": "repo",
"publisher": "Microsoft",
"verified": "2026-08-21",
"keyFindings": "Default mode is 'Playwright's accessibility tree, not pixel-based input'; browser_snapshot returns interactive elements with roles and accessible names, and every action tool takes a 'target' = 'exact target element reference from the page snapshot'. Coordinate clicking exists only behind the optional --caps=vision flag. Therefore an element absent from the a11y snapshot is literally unaddressable by the default toolchain."
},
{
"id": "S17",
"title": "Lighthouse overview — audit categories",
"url": "https://developer.chrome.com/docs/lighthouse/overview",
"type": "vendor-doc",
"publisher": "Google",
"verified": "2026-08-20",
"keyFindings": "Confirms the shipped Agentic Browsing category composition: WebMCP integration (registered tools, forms missing declarative WebMCP support, schema validity), Discoverability (llms.txt), Accessibility for agents, Layout Stability. No DOM-operability, occlusion, control-substitution, or state-introspection auditing."
},
{
"id": "S18",
"title": "Why Do LLM-based Web Agents Fail? A Hierarchical Planning Perspective",
"url": "https://arxiv.org/abs/2603.14248",
"type": "study",
"publisher": "arXiv",
"verified": "2026-08-20",
"keyFindings": "Decomposes failures across planning, execution and replanning layers and concludes 'low-level execution remains the dominant bottleneck', arguing that 'improving perceptual grounding and adaptive control, not only high-level reasoning, is critical'. Supports prioritising DOM-level operability checks over content/semantics checks when predicting agent task failure."
},
{
"id": "S19",
"title": "Operator System Card",
"url": "https://cdn.openai.com/operator_system_card.pdf",
"type": "vendor-doc",
"publisher": "OpenAI",
"verified": "2026-08-20",
"keyFindings": "Documents confirmation prompts before 'actions that affect the state of the world (e.g., before completing a purchase or sending an email)' — 92% recall on 607 risky-action tasks, reducing mistake risk ~90% — plus watch mode forcing supervision on high-impact sites, and proactive refusal of high-risk categories. Implication for site authors: an agent must be able to observe and verify state transitions before and after acting, which requires persistent, machine-readable confirmation of what changed."
},
{
"id": "S2",
"title": "Playwright: Auto-waiting / Actionability checks",
"url": "https://playwright.dev/docs/actionability",
"type": "vendor-doc",
"publisher": "Microsoft",
"verified": "2026-08-20",
"keyFindings": "Before click/check/fill/selectOption, Playwright enforces five checks: Visible (non-empty bounding box, not visibility:hidden), Stable (same bounding box over 2 animation frames), Receives Events (element is the hit target at the action point — overlays cause failure), Enabled (not [disabled]/aria-disabled), Editable (not readonly/aria-readonly). Fill requires visible+enabled+editable. This is the exact gate every Playwright-based agent (Playwright-MCP, browser-use, most CUA harnesses) passes through, so each check is a directly testable site-side failure cause."
},
{
"id": "S3",
"title": "MCP Specification 2026-07-28 — Versioning and Compatibility",
"url": "https://modelcontextprotocol.io/specification/2026-07-28/basic/versioning",
"type": "spec",
"publisher": "Model Context Protocol",
"verified": "2026-08-20",
"keyFindings": "'There is no negotiation handshake.' Terminology: Modern = 2026-07-28+ (per-request _meta); Legacy = 2025-11-25 and earlier (initialize handshake). Unsupported version MUST return error code -32022 with data.supported[] and data.requested. Verbatim compatibility matrix: Modern client + Legacy server = FAILS. Legacy client + Modern server = FAILS. Only dual-era implementations bridge. Extensions negotiated via capabilities.extensions map with mandatory reverse-DNS prefix."
},
{
"id": "S4",
"title": "MCP Specification 2026-07-28 — Tools",
"url": "https://modelcontextprotocol.io/specification/2026-07-28/server/tools",
"type": "spec",
"publisher": "Model Context Protocol",
"verified": "2026-08-20",
"keyFindings": "tools/list result set MUST NOT vary per-connection or as a side effect of other requests (MAY vary by authorization). Servers SHOULD return tools in deterministic order — rationale given verbatim: enables client caching and 'improves LLM prompt cache hit rates'. inputSchema MUST be a valid JSON Schema object (not null); defaults to JSON Schema 2020-12. Tool names SHOULD be 1-128 chars, case-sensitive, only [A-Za-z0-9_.-], unique within a server. Full x-mcp-header constraint list including static-reachability rule (chain of only `properties` keys; never through items/oneOf/anyOf/allOf/not/if/then/else/$ref). Clients MUST exclude violating tools from tools/list. If outputSchema present, servers MUST conform. Clients MUST treat annotations as untrusted."
},
{
"id": "S5",
"title": "MCP Specification 2026-07-28 — Caching",
"url": "https://modelcontextprotocol.io/specification/2026-07-28/server/utilities/caching",
"type": "spec",
"publisher": "Model Context Protocol",
"verified": "2026-08-20",
"keyFindings": "'Servers MUST include caching hints on results with resultType: \"complete\"' for server/discover, tools/list, prompts/list, resources/list, resources/templates/list, resources/read. ttlMs is an integer ms; servers MUST provide ttlMs >= 0. If ttlMs is absent clients SHOULD assume 0 = immediately stale. cacheScope is exactly \"public\" or \"private\". Servers MUST apply the same cacheScope to all pages of a paginated list. Public scope on an authenticated endpoint may be shared across access tokens — servers MUST NOT rely on cacheScope for access control."
},
{
"id": "S6",
"title": "MCP Specification 2026-07-28 — Authorization",
"url": "https://modelcontextprotocol.io/specification/2026-07-28/basic/authorization",
"type": "spec",
"publisher": "Model Context Protocol",
"verified": "2026-08-20",
"keyFindings": "'MCP servers MUST implement OAuth 2.0 Protected Resource Metadata (RFC9728).' Authorization servers MUST provide RFC8414 or OIDC Discovery. Servers SHOULD include a scope parameter in the WWW-Authenticate challenge. Example verbatim: `WWW-Authenticate: Bearer resource_metadata=\"https://mcp.example.com/.well-known/oauth-protected-resource\", scope=\"files:read\"`. Insufficient scope -> 403 with error=\"insufficient_scope\". Servers SHOULD NOT include offline_access in WWW-Authenticate scope or in PRM scopes_supported. Canonical server URI rules: no fragment, scheme required, prefer no trailing slash. Servers MUST validate token audience; MUST NOT accept or transit other tokens."
},
{
"id": "S7",
"title": "MCP Specification 2026-07-28 — Authorization Server Discovery",
"url": "https://modelcontextprotocol.io/specification/2026-07-28/basic/authorization/authorization-server-discovery",
"type": "spec",
"publisher": "Model Context Protocol",
"verified": "2026-08-20",
"keyFindings": "PRM document returned by the MCP server MUST include authorization_servers with at least one entry (stronger than RFC 9728, where it is OPTIONAL). Two discovery mechanisms, both of which clients MUST support: WWW-Authenticate resource_metadata, then well-known probing in order — path-inserted (https://example.com/public/mcp -> https://example.com/.well-known/oauth-protected-resource/public/mcp) then root. AS metadata probing order for issuers with a path: /.well-known/oauth-authorization-server/{path}, /.well-known/openid-configuration/{path}, {path}/.well-known/openid-configuration; without a path: /.well-known/oauth-authorization-server then /.well-known/openid-configuration. Clients MUST reject a metadata doc whose issuer differs from the issuer used to build the URL."
},
{
"id": "S8",
"title": "MCP Security Best Practices (2026-07-28)",
"url": "https://modelcontextprotocol.io/docs/2026-07-28/tutorials/security/security_best_practices.md",
"type": "spec",
"publisher": "Model Context Protocol",
"verified": "2026-08-20",
"keyFindings": "Token passthrough: 'MCP servers MUST NOT accept any tokens that were not explicitly issued for the MCP server.' Scope minimization: 'Common Mistakes' list names publishing all possible scopes in scopes_supported and using wildcard/omnibus scopes (*, all, full-access). State handle hijacking replaces session hijacking now that MCP is stateless: servers MUST NOT treat possession of a state handle as authentication; SHOULD use non-deterministic handles bound server-side to the authenticated user. SSRF section: clients SHOULD require HTTPS for all OAuth-related URLs and block private/link-local ranges (169.254.0.0/16 etc.)."
},
{
"id": "S9",
"title": "RFC 9728 — OAuth 2.0 Protected Resource Metadata",
"url": "https://www.rfc-editor.org/rfc/rfc9728.html",
"type": "spec",
"publisher": "IETF",
"verified": "2026-08-20",
"keyFindings": "`resource` is the only REQUIRED metadata parameter; scopes_supported and resource_name are RECOMMENDED; authorization_servers is OPTIONAL at the RFC level. Section 3 well-known construction: insert /.well-known/oauth-protected-resource between host and path, removing any terminating slash after the host (https://resource.example.com/resource1 -> https://resource.example.com/.well-known/oauth-protected-resource/resource1). Section 3.3 validation: the retrieved `resource` value MUST be identical to the resource identifier used to build the request URL; on mismatch the response data MUST NOT be used. Section 7.7 recommends blocking private/reserved IP ranges."
},
{
"id": "a2a-agent-discovery",
"title": "A2A Protocol — Agent Discovery",
"url": "https://a2a-protocol.org/latest/topics/agent-discovery/",
"type": "spec",
"publisher": "Linux Foundation / A2A Project",
"verified": "2026-08-20",
"keyFindings": "Normative path: `https://{agent-server-domain}/.well-known/agent-card.json`, following RFC 8615. Three discovery mechanisms defined: Well-Known URI, Curated Registries, Direct Configuration."
},
{
"id": "a2a-repo",
"title": "a2aproject/A2A",
"url": "https://github.com/a2aproject/A2A",
"type": "draft-spec",
"publisher": "Linux Foundation / A2A Project",
"verified": "2026-08-20",
"keyFindings": "25,433 stars, 2,575 forks, created 2025-03-25, last push 2026-08-18 — the largest and most active agent-protocol repo in this domain by an order of magnitude over agents.json."
},
{
"id": "a2a-spec",
"title": "A2A Protocol Specification v1.0.0",
"url": "https://a2a-protocol.org/latest/specification/",
"type": "spec",
"publisher": "Linux Foundation / A2A Project",
"verified": "2026-08-20",
"keyFindings": "Version 1.0.0. Agent Card is 'a JSON metadata document published by an A2A Server that describes its identity, capabilities, skills, service endpoint, and authentication requirements.' This URL is the IANA change-controller reference for the registered `agent-card.json` well-known suffix."
},
{
"id": "acceptmarkdown-agent-matrix",
"title": "Which AI agents support Markdown content negotiation? (status matrix)",
"url": "https://acceptmarkdown.com/status",
"type": "dataset",
"publisher": "acceptmarkdown.com",
"verified": "2026-08-20",
"keyFindings": "Per-agent Accept-header matrix, tests Apr–Jun 2026, last updated 2026-06-22. Methodology: capture Accept headers in server logs, trigger agents at specific URLs, inspect access logs. SEND text/markdown (7): Claude Code ('Accept: text/markdown, text/html, */*'), Copilot Chat, Copilot CLI, Cursor, Microsoft Copilot, OpenClaw, OpenCode (q=1.0). PARTIAL — Codex CLI (OpenAI): no markdown Accept preference, but 'fetches the canonical URL as HTML first, then parses the response for ' and requests the markdown version separately. This is the ONLY named consumer of the rel=alternate link tag I found, and it is single-sourced. NO SUPPORT (HTML only): Aider, ChatGPT, Claude.ai, Cline, Devin, Gemini variants, Grok, Perplexity, v0, Windsurf, Zed."
},
{
"id": "acp-architecture",
"title": "ACP Concepts: Architecture",
"url": "https://agenticcommerce.dev/docs/concepts/architecture",
"type": "spec",
"publisher": "Agentic Commerce Protocol",
"verified": "2026-08-20",
"keyFindings": "Four roles: Buyer, Agent, Seller (merchant of record, implements Checkout API), Payment Provider (tokenizes credentials with allowance constraints). IMPORTANT NEGATIVE RESULT: the architecture documents no seller discovery mechanism — no registry, no .well-known URL, no automatic endpoint discovery. Seller onboarding is out-of-band/manual. This means any 'ACP endpoint discovery' audit check would be speculative today, and endpoint conformance testing must accept an operator-supplied base URL."
},
{
"id": "acp-capabilities",
"title": "ACP Concepts: Capability Negotiation",
"url": "https://agenticcommerce.dev/docs/concepts/capability-negotiation",
"type": "spec",
"publisher": "Agentic Commerce Protocol",
"verified": "2026-08-20",
"keyFindings": "Sellers advertise support via a `capabilities` object in API responses, not via HTTP headers. capabilities.payment.handlers identifies payment methods, PSP and instrument schemas. Interventions negotiated as the intersection of agent and seller support, with types including 3ds, biometric, address_verification. Extensions declared per-session. Same capabilities structure used by both parties; request-vs-response context indicates who is declaring."
},
{
"id": "acp-checkout-openai",
"title": "Agentic Checkout Specification",
"url": "https://developers.openai.com/commerce/specs/checkout/",
"type": "spec",
"publisher": "OpenAI / Stripe (Agentic Commerce Protocol)",
"verified": "2026-08-20",
"keyFindings": "Five merchant-hosted HTTPS+JSON endpoints: POST /checkout_sessions (201), POST /checkout_sessions/{checkout_session_id}, POST /checkout_sessions/{id}/complete, POST /checkout_sessions/{id}/cancel (405 if not cancelable), GET /checkout_sessions/{id} (404 if absent). Request headers: Authorization, Accept-Language, User-Agent, Idempotency-Key, Request-Id, Content-Type, Signature, Timestamp (RFC 3339), API-Version. Response MUST echo Idempotency-Key and Request-Id. Session object carries id, status, currency (lowercase ISO 4217), line_items, fulfillment_options, totals, messages, links, payment_provider. Error envelope: {type, code, message, param(JSONPath)}. Message error codes include missing, invalid, out_of_stock, payment_declined, requires_sign_in, requires_3ds. Link types include terms_of_use, privacy_policy, seller_shop_policies."
},
{
"id": "acp-checkout-spec",
"title": "Agentic Commerce Protocol — Agentic Checkout Spec",
"url": "https://developers.openai.com/commerce/specs/checkout",
"publisher": "OpenAI",
"type": "spec",
"keyFindings": "Five REST endpoints: POST /checkout_sessions, POST /checkout_sessions/{id}, /complete, /cancel, GET /checkout_sessions/{id}. Scoped to checkout only, not search or catalog browsing. REST-only delivery; \"In the future, the Agentic Checkout Spec will support MCP servers\". No in-page tools, no HTML attributes, no WebMCP reference.",
"verified": "2026-08-20",
"accessed": "2026-08-21"
},
{
"id": "acp-feed",
"title": "Agentic Commerce Protocol — Product Feed Spec",
"url": "https://agentic-commerce-protocol.com/docs/commerce/specs/feed",
"type": "spec",
"publisher": "OpenAI + Stripe",
"verified": "2026-08-20",
"keyFindings": "Three specs: Agentic Checkout, Product Feed, Delegated Payment (Apache 2.0). Feed required fields include seller_name, seller_url, seller_privacy_policy, seller_tos (last two required if checkout enabled), shipping as country:region:service_class:price. Recommended: offer_id, color, size, size_system, gender. Formats TSV/CSV/XML/JSON. Feeds are PUSHED to OpenAI at an agreed endpoint (not published at a public well-known URL), refresh every 15 min — so no external audit can fetch a merchant's ACP feed."
},
{
"id": "acp-feed-openai",
"title": "OpenAI Product Feed Specification",
"url": "https://developers.openai.com/commerce/specs/feed/",
"type": "vendor-doc",
"publisher": "OpenAI",
"verified": "2026-08-21",
"keyFindings": "Required feed fields: item_id (<=100 chars), title (<=150), description (<=5000), brand (<=70), url, image_url, price (number + ISO 4217), availability enum {in_stock,out_of_stock,pre_order,backorder,unknown}, is_eligible_search, is_eligible_checkout, target_countries. CONDITIONAL hard requirements: availability_date required if availability is pre_order/backorder; seller_privacy_policy AND seller_tos required if is_eligible_checkout=true; gtin OR mpn required unless identifier_exists=no. Recommended: group_id, listing_has_variations, variant_dict, size, size_system, color, item_group_title, review_count, star_rating, q_and_a, reviews, age_restriction. Optional fulfillment field `shipping` uses format country:region:service_class:price:handling_days:transit_days. Returns fields: accepts_returns, return_deadline_in_days (positive int), accepts_exchanges, return_policy (URL, HTTPS preferred). sale_price must be less than price. Formats: .txt/.tsv/.csv (+.gz), UTF-8, lowercase underscore header row. Parser auto-selects between OpenAI schema and a Google-compatible profile."
},
{
"id": "acp-feed-spec",
"title": "Product Feed Specification — Agentic Commerce Protocol",
"url": "https://agentic-commerce-protocol.com/docs/commerce/specs/feed",
"type": "draft-spec",
"publisher": "Agentic Commerce Protocol",
"verified": "2026-08-20",
"keyFindings": "Defines its own field taxonomy (item_group_title, color, size, offer_id, …) as \"a structured source of truth\"; does not reference schema.org vocabulary, GTIN, or Google Merchant Center compatibility. The emerging agentic-commerce standard is deliberately feed-native, not markup-native."
},
{
"id": "acp-github",
"title": "Agentic Commerce Protocol (ACP) specification repository",
"url": "https://github.com/agentic-commerce-protocol/agentic-commerce-protocol",
"type": "draft-spec",
"publisher": "OpenAI + Stripe (Founding Maintainers)",
"verified": "2026-08-20",
"keyFindings": "Apache 2.0, date-versioned releases (latest stable 2026-04-17): openapi.agentic_checkout.yaml, openapi.delegate_payment.yaml, JSON schemas; cart, feed, orders, auth and MCP surfaces. CRITICAL FOR SIGNAL 9: the repo defines NO website-side discovery mechanism — no well-known URL, no HTML link tag, no sitemap-based or markup-based endpoint discovery. Merchant endpoints are registered out-of-band."
},
{
"id": "acp-home",
"title": "Agentic Commerce Protocol homepage",
"url": "https://agenticcommerce.dev/",
"type": "spec",
"publisher": "OpenAI + Stripe",
"verified": "2026-08-20",
"keyFindings": "Open standard (Apache 2.0) for programmatic commerce flows between buyers, AI agents and businesses, developed jointly by Stripe and OpenAI. Confirms the protocol is genuinely open and implementable by any merchant, not gated to a single platform."
},
{
"id": "acp-openapi-cart",
"title": "openapi.cart.yaml (spec version 2026-04-17)",
"url": "https://raw.githubusercontent.com/agentic-commerce-protocol/agentic-commerce-protocol/main/spec/2026-04-17/openapi/openapi.cart.yaml",
"type": "spec",
"publisher": "Agentic Commerce Protocol",
"verified": "2026-08-20",
"keyFindings": "Seller-hosted pre-checkout Cart API at https://seller.example.com: POST /carts, GET /carts/{id}, PUT /carts/{id} (full replacement), POST /carts/{id}/cancel. Required headers Authorization (Bearer) and API-Version (e.g. 2026-04-17); Idempotency-Key required on POST; Content-Type application/json on POST/PUT; optional Accept-Language, Request-Id. The cart response carries a `continue_url` (e.g. https://seller.example.com/cart/cart_abc123) used to hand the buyer off to the seller's own checkout UI. No payment configuration in this phase; expired carts return 404. This is the lower-lift adoption path versus full delegated-payment Instant Checkout."
},
{
"id": "acp-openapi-checkout",
"title": "openapi.agentic_checkout.yaml (spec version 2026-04-17)",
"url": "https://raw.githubusercontent.com/agentic-commerce-protocol/agentic-commerce-protocol/main/spec/2026-04-17/openapi/openapi.agentic_checkout.yaml",
"type": "spec",
"publisher": "Agentic Commerce Protocol",
"verified": "2026-08-20",
"keyFindings": "CheckoutSession REQUIRED fields (9): id, status, currency, line_items, totals, fulfillment_options, messages, links, capabilities. links[].type enum expanded to 8 values: terms_of_use, privacy_policy, return_policy, shipping_policy, contact_us, about_us, faq, support. status enum (11): incomplete, not_ready_for_payment, requires_escalation, authentication_required, ready_for_payment, pending_approval, complete_in_progress, completed, canceled, in_progress, expired. totals[].type enum (12): items_base_amount, items_discount, subtotal, discount, fulfillment, tax, fee, gift_wrap, tip, store_credit, total, amount_refunded. Message error codes extended with low_stock, quantity_exceeded, coupon_invalid, coupon_expired, minimum_not_met, maximum_exceeded, region_restricted, age_verification_required, approval_required, unsupported, not_found, conflict, rate_limited, expired, intervention_required. API-Version is YYYY-MM-DD, required on all requests. Response headers Idempotency-Key and Request-Id are required echoes."
},
{
"id": "acp-openapi-feed",
"title": "openapi.feed.yaml (spec version 2026-04-17)",
"url": "https://raw.githubusercontent.com/agentic-commerce-protocol/agentic-commerce-protocol/main/spec/2026-04-17/openapi/openapi.feed.yaml",
"type": "spec",
"publisher": "Agentic Commerce Protocol",
"verified": "2026-08-20",
"keyFindings": "Merchant-hosted feed API (server https://merchant.example.com): POST /feeds, GET /feeds/{id}, GET /feeds/{id}/products, PATCH /feeds/{id}/products. Product object REQUIRES id and variants[] (array of Variant). Variant REQUIRES id and title; optional description, url, barcodes, price, list_price, unit_price, availability, categories, condition, variant_options, media, seller, marketplace. Critically, the model is variant-first: every sellable thing is a Variant with its own id, price and availability — a product without resolvable variants has no purchasable unit."
},
{
"id": "acp-payment",
"title": "Delegated Payment Specification",
"url": "https://developers.openai.com/commerce/specs/payment.md",
"type": "spec",
"publisher": "OpenAI / Stripe",
"verified": "2026-08-20",
"keyFindings": "PSP-or-merchant-hosted endpoint POST /agentic_commerce/delegate_payment. Same nine headers as checkout (Authorization, Accept-Language, User-Agent, Idempotency-Key, Request-Id, Content-Type, Signature, Timestamp, API-Version). Request requires payment_method (type card only; card number type fpan or network_token) and allowance {reason:'one_time', max_amount, currency (ISO 4217), checkout_session_id, merchant_id, expires_at (RFC 3339)}; optional billing_address, risk_signals, metadata. Success 201 returns id = vault token in the form vt_… — single-use, scoped outside PCI scope, bounded by max_amount and expiry."
},
{
"id": "acp-repo",
"title": "agentic-commerce-protocol repository",
"url": "https://github.com/agentic-commerce-protocol/agentic-commerce-protocol",
"type": "repo",
"publisher": "Agentic Commerce Protocol (OpenAI + Stripe, Apache 2.0)",
"verified": "2026-08-20",
"keyFindings": "Date-versioned spec releases: 2025-09-29 (initial), 2025-12-12 (fulfillment), 2026-01-16 (capability negotiation), 2026-01-30 (extensions, discounts, payment handlers), 2026-04-17 (current stable: cart, feed, orders, authentication, MCP), plus unreleased/. Artifacts under spec//: openapi/, json-schema/, openrpc/. openapi/ for 2026-04-17 contains openapi.agentic_checkout.yaml, openapi.agentic_checkout_webhook.yaml, openapi.cart.yaml, openapi.delegate_authentication.yaml, openapi.delegate_payment.yaml, openapi.feed.yaml. No .well-known or discovery mechanism anywhere in the repo."
},
{
"id": "acp-security",
"title": "ACP Concepts: Security",
"url": "https://agenticcommerce.dev/docs/concepts/security",
"type": "spec",
"publisher": "Agentic Commerce Protocol",
"verified": "2026-08-20",
"keyFindings": "All ACP endpoints use HTTPS and send/receive JSON. Bearer-token authentication between agent and seller; sellers retrieve tokens through the agent's application. Mandatory headers: Authorization: Bearer , Content-Type: application/json, Accept: application/json."
},
{
"id": "acp-sellers",
"title": "ACP Getting Started: Sellers",
"url": "https://agenticcommerce.dev/docs/getting-started/sellers",
"type": "spec",
"publisher": "Agentic Commerce Protocol",
"verified": "2026-08-20",
"keyFindings": "Seller obligations: implement the five HTTPS/JSON checkout endpoints; calculate all amounts (item prices, discounts, taxes, shipping); manage inventory and availability; process payments through their PSP; fulfil orders. Sellers must declare capabilities in EVERY checkout response, including payment handlers (handler id, name, version, PSP reference, configuration) and optional extensions. Sellers must validate payment handler IDs against declared capabilities and manage state transitions to ready_for_payment."
},
{
"id": "acp-spec-confirm",
"title": "Agentic Commerce Protocol",
"url": "https://www.agenticcommerce.dev/",
"publisher": "OpenAI + Stripe",
"type": "spec",
"keyFindings": "Completion/confirmation flows run through agent-facing endpoints ('Publish your checkout configuration with a traditional API or MCP') integrated with the merchant's commerce backend and payment processor. No schema.org, no confirmation-page markup.",
"verified": "2026-08-21",
"accessed": "2026-08-21"
},
{
"id": "aeo-auditor-long-tail",
"title": "aeo-auditor ('Think Lighthouse, but for AI coding agents')",
"url": "https://github.com/JhontyBarreto/aeo-auditor",
"type": "article",
"publisher": "JhontyBarreto — 0 stars",
"verified": "2026-08-20",
"keyFindings": "Representative of a very large 0-star long tail confirmed via the GitHub search API in this session: AJ-EN/aeo-audit ('PageSpeed for AI Agents'), saai07/GeoLens, devakrishnar/AEO-Analyzer, akimovpro/seo-for-ai, api-evangelist/gleefulai-visibility, truos-official/Generative-Engine-Optimization-Audit, Third-Sun-Pro/geo-auditor, msiroya21/GEO-Auditor, echorank-com/geo-checker-cli, indexcraft/ai-crawler-access-toolkit, plus a dozen llms.txt generator/validator kits. Near-identical check surfaces (llms.txt, robots.txt AI-bot rules, structured data, content structure), near-zero adoption, and not one publishes evidence grading. The category is crowded but undifferentiated on rigour."
},
{
"id": "agent-lighthouse-own-measurement",
"title": "Direct measurement of vendor llms.txt files and markdown content negotiation (2026-08-20)",
"url": "https://llmstxt.org/",
"type": "dataset",
"publisher": "Agent Lighthouse (this research, reproducible via curl)",
"verified": "2026-08-21",
"keyFindings": "ORIGINAL EMPIRICAL DATA, reproducible. (1) CONTENT NEGOTIATION LIVE: curl -H 'Accept: text/markdown' returned 'content-type: text/markdown; charset=utf-8' from platform.claude.com/docs, docs.stripe.com/api, and developers.cloudflare.com/fundamentals/. (2) LINK TAG LIVE: developers.cloudflare.com emits ; docs.stripe.com and nextjs.org/docs emit none. Stripe instead emits Link: ; rel=\"service-meta\". (3) llms-full.txt SIZES ARE ABSURD: Anthropic 33.5 MB, Cloudflare 57.3 MB, OpenAI 6.1 MB, Perplexity 4.1 MB; Stripe returns 404. These exceed every production context window by 1-2 orders of magnitude. (4) FORMAT CONFORMANCE across 7 major vendor llms.txt files (Anthropic, Cloudflare, Perplexity, Stripe, OpenAI, GitHub, Vercel): 7/7 have an H1 and markdown links (all would PASS Lighthouse), but only 5/7 have a blockquote summary — Anthropic and Stripe omit it. Link counts range 38 (OpenAI) to 1,872 (Vercel); sizes 5.8 KB to 215 KB."
},
{
"id": "agent-ready-websites",
"title": "Designing Agent-Ready Websites for AI Web Agents",
"url": "https://arxiv.org/abs/2607.12056",
"publisher": "arXiv",
"type": "study",
"verified": "2026-08-20",
"keyFindings": "Structures agent readiness as 'agent interpretability, agent executability, and agent decision reliability', supported by 'machine readability, semantic clarity, agent actionability, and contextual decision-reliability signals', with improvements in 'structural clarity, action cues, evidence signals, and temporal validity indicators'. Framework identity, framework detection and client-side rendering are absent from the framework; it also notes GEO metrics do not fully assess agent-readiness.",
"accessed": "2026-08-21"
},
{
"id": "agentic-commerce-protocol",
"title": "Agentic Commerce Protocol",
"url": "https://www.agenticcommerce.dev/",
"publisher": "OpenAI + Stripe",
"type": "spec",
"keyFindings": "'An open standard for programmatic commerce flows between buyers, AI agents, and businesses.' Merchants 'Publish your checkout configuration with a traditional API or MCP'. No schema.org markup, no JSON-LD, no potentialAction anywhere in the protocol.",
"verified": "2026-08-21",
"accessed": "2026-08-21"
},
{
"id": "agents-json-com-dead",
"title": "agents-json.com — upstream repo's listed homepage, unreachable",
"url": "https://agents-json.com",
"publisher": "Wild Card AI",
"type": "vendor-doc",
"keyFindings": "(Domain does not resolve — the dead domain IS the evidence.) DNS resolves to 216.92.3.51 but both HTTPS and HTTP time out (curl exit 28, http_code=000); WebFetch returned ECONNREFUSED. The project has no live website at either of its two domains.",
"verified": "2026-08-20",
"accessed": "2026-08-21"
},
{
"id": "agents-txt-spec",
"title": "agents.txt specification",
"url": "https://agents-txt.com/spec/",
"publisher": "agents-txt.com (community)",
"type": "spec",
"keyFindings": "Only found source placing security.txt in an agent context: \"security.txt is a human-readable vulnerability disclosure channel published at /.well-known/security.txt; agents.txt is a machine-readable capability declaration. Sites that take agent payments or authentication SHOULD publish both.\" Does not reference Referrer-Policy, Permissions-Policy or preconnect. Community-maintained, no vendor adoption.",
"verified": "2026-08-20",
"accessed": "2026-08-21"
},
{
"id": "agentsjson-org-dead",
"title": "agentsjson.org — audit's docsUrl, domain no longer resolves",
"url": "https://agentsjson.org/",
"publisher": "Wild Card AI (defunct)",
"type": "vendor-doc",
"keyFindings": "(Domain does not resolve — the dead domain IS the evidence.) WebFetch failed with getaddrinfo ENOTFOUND; dig returns no A record; curl returns 'Could not resolve host: agentsjson.org' (http_code=000). This is the exact URL the audit ships as its docsUrl, so users are sent to a dead domain.",
"verified": "2026-08-20",
"accessed": "2026-08-21"
},
{
"id": "agentswelcome",
"title": "AGENTS WELCOME",
"url": "https://agentswelcome.dev/",
"publisher": "agentswelcome.dev (demo site)",
"type": "article",
"keyFindings": "Agent-first demo site listing security.txt as part of a \"complete machine-discovery surface\" alongside sitemap.xml and Atom feeds, while assigning the authoritative machine-readable role to /.well-known/agents.json. Self-describes as a demonstration, not a standard. No claim that agents parse security.txt.",
"verified": "2026-08-20",
"accessed": "2026-08-21"
},
{
"id": "ahrefs",
"title": "Ahrefs Brand Radar",
"url": "https://ahrefs.com/brand-radar",
"type": "vendor-doc",
"publisher": "Ahrefs",
"verified": "2026-08-20",
"keyFindings": "AI visibility tracking, competitive benchmarking, AI citation discovery, YouTube/TikTok/Reddit monitoring, custom prompts, dashboard. Explicitly no llms.txt, no AI-bot robots.txt rules, no agent schema audits."
},
{
"id": "ahrefs-ai-traffic-study",
"title": "AI traffic study: how much traffic do AI chatbots actually send?",
"url": "https://ahrefs.com/blog/ai-traffic-study/",
"type": "study",
"publisher": "Ahrefs",
"verified": "2026-08-20",
"keyFindings": "3,000 websites analyzed across seven AI platforms (ChatGPT, Claude, Copilot, Gemini, Perplexity, Jasper, Mistral), segmented by monthly visitor volume. 63% of sampled sites receive at least one AI-sourced visit. But AI chatbots generate only 0.12% of average website views and 0.17% of visitors. ChatGPT, Perplexity and Gemini are 98% of AI referrals; ChatGPT alone 50%. Small sites capture proportionally more (0.56% of visitors) than large ones (0.13%). Caveat: Copilot stopped passing referrer data around October 2024, so some AI traffic is misattributed to direct — real figures are somewhat higher."
},
{
"id": "ahrefs-aio-top10",
"title": "Update: 38% of AI Overview Citations Pull From The Top 10",
"url": "https://ahrefs.com/blog/ai-overview-citations-top-10/",
"type": "study",
"publisher": "Ahrefs",
"verified": "2026-08-20",
"keyFindings": "863K keyword SERPs and 4M AI Overview URLs, data current as of 2 March 2026 (>2x the July 2025 study). Only 38% of AI-Overview-cited pages rank in the top 10 for the query; 31.2% rank 11–100 and 31.0% rank beyond 100. Organic blue links only: 37.1% / 26.2% / 36.7%. Down sharply from ~76% top-10 in July 2025. Attributed to query fan-out — sources recurring across simulated sub-queries get cited rather than those ranking for the direct query. YouTube is the single most-cited domain, supplying 18.2% of citations from outside the top 100."
},
{
"id": "ahrefs-brand-correlation",
"title": "An Analysis of AI Overview Brand Visibility Factors (75K Brands Studied)",
"url": "https://ahrefs.com/blog/ai-overview-brand-correlation/",
"type": "study",
"publisher": "Ahrefs",
"verified": "2026-08-20",
"keyFindings": "75,000 brands (Domain Rating >40, highest-volume keyword ≥800 monthly searches); millions of AI Overview responses via Brand Radar; ~26% of domains had zero mentions. Spearman correlations with AI Overview brand visibility: brand web mentions 0.664, brand anchors 0.527, brand search volume 0.392, Domain Rating 0.326, referring domains 0.295, branded organic traffic 0.274, backlinks 0.218, URL Rating 0.18, site pages 0.17. Off-site brand signals beat backlinks by ~3x. Explicit caveats: 'correlation ≠ causation' and 'all the factors we studied revealed moderate to very weak correlations'. No author-level or on-page brand-consistency variable was measured."
},
{
"id": "ahrefs-brand-radar",
"title": "Ahrefs Brand Radar",
"url": "https://ahrefs.com/brand-radar",
"type": "vendor-doc",
"publisher": "Ahrefs",
"verified": "2026-08-20",
"keyFindings": "Tracks brand mentions across AI Overviews/AI Mode, ChatGPT, Copilot, Gemini, Perplexity and Grok against a database of '476M+ total monthly prompts', described as 'search-backed prompts, not synthetic ones', and indexes YouTube/TikTok/Reddit as citation sources. Explicitly a visibility monitor, not a technical AI-readiness auditor — site-side checking is left to the separate Site Audit product, which has no AI-crawler audit surface described here."
},
{
"id": "ahrefs-brand-visibility-correlations",
"title": "Top Brand Visibility Factors in ChatGPT, AI Mode, and AI Overviews (75k Brands Studied)",
"url": "https://ahrefs.com/blog/ai-brand-visibility-correlations/",
"type": "study",
"publisher": "Ahrefs",
"verified": "2026-08-20",
"keyFindings": "75,000 brands, millions of AI responses, Spearman correlations across ChatGPT / AI Mode / AI Overviews. Structured data and schema markup are NOT among the measured factors. Top correlates: YouTube mentions (~0.737), branded web mentions (0.656–0.709), branded anchors (0.511–0.628), branded search volume (0.352–0.466). Relevant as an absence-of-evidence datapoint: the largest AI-visibility correlation study did not consider schema worth measuring."
},
{
"id": "ahrefs-freshness-17m",
"title": "Do AI Assistants Prefer to Cite Fresh Content?",
"url": "https://ahrefs.com/blog/do-ai-assistants-prefer-to-cite-fresh-content",
"type": "study",
"publisher": "Ahrefs",
"verified": "2026-08-21",
"keyFindings": "16.975M cited URLs across ChatGPT, Perplexity, Gemini, Copilot, Google AI Overviews and organic Google SERPs (Brand Radar + Content Explorer). AI-cited pages average 1,064 days old (2.9 years) vs 1,432 days for organic results = 25.7% fresher. Per-platform average age since publication: ChatGPT 958 days (strongest recency bias), Copilot 1,056, Gemini 1,118, Perplexity 1,166, Google AI Overviews 1,432 — AI Overviews are NOT fresher than organic. Author caveats: the average cited page is still 2.9 years old, so AI assistants still favour established content; warns against updating without meaningful change, noting 'low-quality, irrelevant content that's updated every day will not have a magic positive effect.'"
},
{
"id": "ahrefs-schema-ai-citations",
"title": "We Tracked 1,885 Pages Adding Schema. AI Citations Barely Moved.",
"url": "https://ahrefs.com/blog/schema-ai-citations/",
"type": "study",
"publisher": "Ahrefs",
"verified": "2026-08-21",
"keyFindings": "Quasi-experiment: 1,885 pages that added JSON-LD Aug 2025–Mar 2026, matched 1:3 against ~4,000 never-JSON-LD control URLs on different domains with similar pre-period citation levels; 30-day pre/post windows; four tests (two-sample t-test, matched difference-in-differences, event study, alternate-window robustness). Results: Google AI Mode +2.4% (indistinguishable from zero), ChatGPT +2.2% (indistinguishable from zero), Google AI Overviews −4.6% (significant, ~1-in-2,500 chance of randomness, ≈12 daily citations per page). Schema types pooled (Article, FAQ, Product, HowTo, Organization) — no per-type isolation. Authors' caveats: sample restricted to pages already heavily cited (100+ AIO citations in Feb 2025), so it cannot speak to first-time discovery/parsing/indexing effects; cannot separate schema from concurrent page changes."
},
{
"id": "ahrefs-short-vs-long-aio",
"title": "Short vs. Long Content in AI Overviews: The Data Says Both Work",
"url": "https://ahrefs.com/blog/short-vs-long-content-in-ai-overviews/",
"type": "dataset",
"publisher": "Ahrefs",
"verified": "2026-08-20",
"keyFindings": "560,346 AI Overviews, 1,677,876 cited URLs, 174,048 pages with valid data after boilerplate removal. 53.4% of citations go to pages under 1,000 words; 36.8% are 350–1,000 words; only 16% exceed 2,000 words. 'Near-zero correlation between word length and being cited' (Spearman 0.04); average word counts for citation positions 1–3 varied only 1,270–1,291. Median word count by type: product listings 315, blog articles 1,166, podcast audio 1,226. Conclusion favors structure over length: 'answer the query directly', use clear headings, 'write for humans first'. Establishes that raw length is not a citation lever."
},
{
"id": "ai-catalog-consuming",
"title": "Consuming catalogs guide",
"url": "https://raw.githubusercontent.com/Agent-Card/ai-catalog/main/docs/guides/consuming-catalogs.md",
"publisher": "Linux Foundation / Agent Card Working Group",
"type": "spec",
"keyFindings": "Prescribes consumer discovery order: HTTP Link header rel=\"ai-catalog\", then HTML , then fall back to GET /.well-known/ai-catalog.json. Names no deployed consumers.",
"verified": "2026-08-20",
"accessed": "2026-08-21"
},
{
"id": "ai-catalog-io",
"title": "AI Catalog — Common AI Catalog and Registry Standard",
"url": "https://ai-catalog.io/",
"publisher": "Linux Foundation / Agent Card Working Group",
"type": "spec",
"keyFindings": "Defines the AI Catalog JSON container published at /.well-known/ai-catalog.json; maintained by the Agent Card WG under the Linux Foundation. Names no concrete consumers or adopters.",
"verified": "2026-08-20",
"accessed": "2026-08-21"
},
{
"id": "ai-catalog-io-meta",
"title": "ai-catalog.io specification site",
"url": "https://ai-catalog.io/",
"publisher": "Linux Foundation / Agent Card Working Group",
"type": "spec",
"keyFindings": "Describes the catalog as a JSON container for MCP servers, A2A agents, Claude Code plugins, datasets and model cards with optional Trust Manifest extensions for attestation/provenance — the spec's answer to 'who owns this', in place of owner/contact fields.",
"verified": "2026-08-20",
"accessed": "2026-08-21"
},
{
"id": "ai-catalog-readme",
"title": "Agent-Card/ai-catalog README",
"url": "https://raw.githubusercontent.com/Agent-Card/ai-catalog/main/README.md",
"publisher": "Linux Foundation (Agent Card WG)",
"type": "repo",
"keyFindings": "Calls itself a temporary Linux Foundation working repo; states A2A and MCP steering committees will vote on adoption — i.e. pre-ratification. 210 stars, actively pushed 2026-08-20.",
"verified": "2026-08-20",
"accessed": "2026-08-21"
},
{
"id": "ai-catalog-repo",
"title": "Agent-Card/ai-catalog — working repository for common AI Card standard",
"url": "https://github.com/Agent-Card/ai-catalog",
"type": "draft-spec",
"publisher": "Agent-Card working group",
"verified": "2026-08-20",
"keyFindings": "210 stars, created 2025-10-29, last push 2026-08-20 (active). This is the protocol-agnostic catalog format that SEP-2127 delegates domain-level discovery to. Not an IETF/W3C document; no IANA registration for `ai-catalog.json`."
},
{
"id": "allenai-crawler",
"title": "AI2Bot — Allen Institute for AI crawler",
"url": "https://allenai.org/crawler",
"type": "vendor-doc",
"publisher": "Allen Institute for AI",
"verified": "2026-08-20",
"keyFindings": "UA: 'Mozilla/5.0 (compatible) AI2Bot (+https://www.allenai.org/crawler)'. Crawls 'to find web content' that is 'used to train open language models'. Page offers the UA string so operators 'can be used to filter or reject traffic from our crawler if desired'. PARTIAL NEGATIVE: the fetched page does not contain an explicit robots.txt-compliance sentence or a canonical disallow snippet."
},
{
"id": "almanac-markup-2024",
"title": "Web Almanac 2024 — Markup",
"url": "https://almanac.httparchive.org/en/2024/markup",
"type": "dataset",
"publisher": "HTTP Archive",
"verified": "2026-08-20",
"keyFindings": "Median 594 elements per mobile page (p90 1,716); median HTML transfer size 33 kB desktop / 32 kB mobile; 10.5% of mobile pages serve HTML uncompressed; 86% of mobile pages contain at least one HTML comment and 26% still ship IE conditional comments; SVG present on 51.6% of pages. Population baseline for calibrating per-page token budgets and for the claim that dead markup ships at scale."
},
{
"id": "almanac-page-weight-2024",
"title": "Web Almanac 2024 — Page Weight",
"url": "https://almanac.httparchive.org/en/2024/page-weight",
"type": "dataset",
"publisher": "HTTP Archive",
"verified": "2026-08-20",
"keyFindings": "Median total page 2,652 kB desktop / 2,311 kB mobile; p90 8,375 kB / 7,680 kB; median desktop homepage loads ~18 kB of HTML against ~1,054 kB images and ~613 kB JS. Useful contrast: for a rendering browser HTML is ~1% of weight, but for a non-rendering AI crawler that HTML document is ~100% of what gets tokenized — so HTML-internal waste is the entire agent-side cost."
},
{
"id": "amazon-science-rufus",
"title": "The technology behind Amazon's GenAI-powered shopping assistant, Rufus",
"url": "https://www.amazon.science/blog/the-technology-behind-amazons-genai-powered-shopping-assistant-rufus",
"type": "article",
"publisher": "Amazon Science",
"verified": "2026-08-20",
"keyFindings": "Rufus data sources: the entire Amazon catalogue, customer reviews, community Q&A posts, \"public information on the web\", plus queries to internal systems and Stores APIs. No mention of schema.org, structured data formats, or crawling merchant markup. Counter-evidence: optimizing site Product markup has no documented path into Rufus."
},
{
"id": "amazonbot-docs",
"title": "Amazonbot",
"url": "https://developer.amazon.com/amazonbot",
"type": "vendor-doc",
"publisher": "Amazon",
"verified": "2026-08-20",
"keyFindings": "UA: 'Mozilla/5.0 AppleWebKit/537.36 (KHTML, like Gecko; compatible; Amazonbot/0.1) Chrome/W.X.Y.Z Safari/537.36'. 'Amazonbot is used to improve our products and services' and 'may be used to train Amazon AI models'. 'Automated crawling from these listed user agents respects the Robots Exclusion Protocol, honoring the user-agent and the allow/disallow directives.' Explicit negative on crawl-delay: 'They do not support the crawl-delay directive.' Also documents a separate Amzn-SearchBot which 'does not crawl content for generative AI model training'."
},
{
"id": "anthropic-bots",
"title": "Anthropic — Does Anthropic crawl data from the web?",
"url": "https://support.claude.com/en/articles/8896518-does-anthropic-crawl-data-from-the-web-and-how-can-site-owners-block-the-crawler",
"type": "vendor-doc",
"publisher": "Anthropic",
"verified": "2026-08-21",
"keyFindings": "Three tokens with distinct purposes: ClaudeBot (training), Claude-User (live user-initiated fetch), Claude-SearchBot (search quality). IP list at https://claude.com/crawling/bots.json. Anthropic states IP-based blocking 'may not work correctly or persistently guarantee an opt-out' — robots.txt product tokens are the sanctioned control surface."
},
{
"id": "anthropic-bots-json",
"title": "Anthropic crawler IP prefixes (bots.json)",
"url": "https://claude.com/crawling/bots.json",
"type": "dataset",
"publisher": "Anthropic",
"verified": "2026-08-20",
"keyFindings": "Live JSON with creationTime (observed 2026-08-18T23:56:36Z) and a 'prefixes' array of IPv4 CIDRs (e.g. 216.73.216.0/22, plus many /32s). Confirms Anthropic now publishes verifiable crawler IP ranges, enabling WAF allow-listing without UA spoofing risk."
},
{
"id": "anthropic-browser-use-address",
"title": "Browser use tool",
"url": "https://platform.claude.com/docs/en/docs/agents-and-tools/tool-use/browser-use-tool",
"publisher": "Anthropic",
"type": "vendor-doc",
"keyFindings": "Documents that the agent reads the page as an accessibility tree via `read_page`. Combined with the snapshot experiment, this establishes that the representation Claude receives contains no -specific signal to act on.",
"verified": "2026-08-20",
"accessed": "2026-08-21"
},
{
"id": "anthropic-browser-use-aside",
"title": "Browser use tool",
"url": "https://platform.claude.com/docs/en/docs/agents-and-tools/tool-use/browser-use-tool",
"publisher": "Anthropic",
"type": "vendor-doc",
"keyFindings": "`read_page` returns 'the page's accessibility tree as text with each element tagged with a reference'; the doc instructs developers to prefer accessibility-tree references over pixel coordinates. Establishes the named vendor agent that receives the `complementary` landmark my experiment observed.",
"verified": "2026-08-20",
"accessed": "2026-08-21"
},
{
"id": "anthropic-browser-use-images",
"title": "Browser use tool",
"url": "https://platform.claude.com/docs/en/docs/agents-and-tools/tool-use/browser-use-tool",
"publisher": "Anthropic",
"type": "vendor-doc",
"keyFindings": "Confirms `read_page` returns the page's accessibility tree as text — the representation in which empty-alt images are already absent, so the audit's 'agents flag false content gaps' scenario has no node to flag.",
"verified": "2026-08-20",
"accessed": "2026-08-21"
},
{
"id": "anthropic-browser-use-tool",
"title": "Browser use tool (browser_toolset_20260801)",
"url": "https://platform.claude.com/docs/en/agents-and-tools/tool-use/browser-use-tool",
"type": "vendor-doc",
"publisher": "Anthropic",
"verified": "2026-08-21",
"keyFindings": "First-party proof that Claude's browser agent perceives pages via the accessibility tree, not raw DOM or pixels. read_page: 'Return the page's accessibility tree as text with each element tagged with a reference such as [ref_2]'; filter=interactive returns only visible interactive elements; depth default 15; output capped at 50,000 characters and truncated. find: 'Search for elements matching a natural-language description such as \"search field\" or \"add to cart button\"'. form_input sets a form element's value by ref. get_page_text: 'Return the page's visible text as plain text, prioritizing the main article content'. Security guidance: 'build page reads from what the page renders (the accessibility tree or visible text), not raw DOM source, so hidden text doesn't reach Claude.' Element references come from read_page and find. Same tool surface is exposed by Anthropic's Claude-in-Chrome MCP server (read_page = 'Get an accessibility tree representation of elements on the page')."
},
{
"id": "anthropic-chrome",
"title": "Piloting Claude for Chrome",
"url": "https://claude.com/blog/claude-for-chrome",
"type": "vendor-doc",
"publisher": "Anthropic",
"verified": "2026-08-20",
"keyFindings": "Red-team attack success rate 23.6% in autonomous browsing mode, 11.2% after mitigations; a browser-specific challenge set went 35.7% -> 0%. Names the exact vectors: 'hidden malicious form fields in a webpage's Document Object Model (DOM) invisible to humans, and other hard-to-catch injections such as through the URL text and tab title that only an agent might see.' This is the vendor-documented basis for auditing hidden inputs and a11y/metadata attributes."
},
{
"id": "anthropic-citations-docs",
"title": "Citations — Claude Platform Documentation",
"url": "https://platform.claude.com/docs/en/docs/build-with-claude/citations",
"type": "vendor-doc",
"publisher": "Anthropic",
"verified": "2026-08-20",
"keyFindings": "Documents the citation unit for a major model provider: 'Document contents are \"chunked\" to define the minimum granularity of possible citations. For example, sentence chunking lets Claude cite a single sentence or chain together multiple consecutive sentences to cite a paragraph or longer passage.' 'For plain text documents: Content is chunked into sentences that can be cited from.' 'For PDFs: ... content is chunked into sentences.' 'For custom content documents: Your provided content blocks are used as-is and no further chunking is done.' Citation formats: character indices (plain text), page numbers (PDF), block indices (custom). Confirms the extraction unit is the sentence, not the section or the page."
},
{
"id": "anthropic-claude-code-tools-reference",
"title": "Tools reference — Claude Code docs (WebFetch behavior)",
"url": "https://code.claude.com/docs/en/tools-reference",
"type": "vendor-doc",
"publisher": "Anthropic",
"verified": "2026-08-20",
"keyFindings": "STRONGEST A-GRADE VENDOR STATEMENT IN THIS DOMAIN, and it is about markdown, not llms.txt: \"WebFetch sets a User-Agent header beginning with Claude-User, and an Accept header that prefers Markdown over HTML so servers that support content negotiation can return Markdown directly.\" This is an explicit vendor commitment that a named agent consumes markdown alternates via content negotiation. No equivalent statement exists anywhere for llms.txt."
},
{
"id": "anthropic-claude-connectors",
"title": "About custom connectors (remote MCP servers) — Claude Help Center",
"url": "https://support.claude.com/en/articles/11175166-about-custom-connectors-remote-mcp-servers",
"publisher": "Anthropic",
"type": "vendor-doc",
"keyFindings": "Directly contradicts the audit's named consumer. Claude requires manual URL entry, not auto-discovery: users 'Navigate to Customize > Connectors', 'Click + then Add custom connector', and 'Add your connector's remote MCP server URL.' No auto-detection from website HTML or /.well-known/ files is mentioned.",
"verified": "2026-08-20",
"accessed": "2026-08-21"
},
{
"id": "anthropic-claude-custom-connectors",
"title": "Third party connectors with remote MCP — Claude Docs",
"url": "https://claude.com/docs/connectors/custom/remote-mcp",
"type": "vendor-doc",
"publisher": "Anthropic",
"verified": "2026-08-20",
"keyFindings": "'You can manually add any third-party connector to Claude as long as you have the URL of that remote MCP server.' Steps for every plan tier are: navigate to Connectors → Add custom connector → 'Enter the remote MCP server URL'. Discovery is either the curated Connectors Directory or manual URL entry. NO .well-known lookup, NO domain-derived discovery is documented anywhere on the page."
},
{
"id": "anthropic-claude-for-chrome",
"title": "Piloting Claude for Chrome",
"url": "https://claude.com/blog/claude-for-chrome",
"publisher": "Anthropic",
"type": "announcement",
"keyFindings": "Defines instructions hidden in web pages/DOM as prompt injection; describes system-prompt hardening and 'advanced classifiers to detect suspicious instruction patterns' to make Claude NOT follow page-embedded instructions.",
"verified": "2026-08-20",
"accessed": "2026-08-21"
},
{
"id": "anthropic-claudebot",
"title": "Does Anthropic crawl data from the web, and how can site owners block the crawler?",
"url": "https://support.claude.com/en/articles/8896518-does-anthropic-crawl-data-from-the-web-and-how-can-site-owners-block-the-crawler",
"publisher": "Anthropic",
"type": "vendor-doc",
"keyFindings": "\"Anthropic's Bots respect 'do not crawl' signals by honoring industry standard directives in robots.txt\" plus non-standard Crawl-delay support and IP JSON lists. No HTTP response headers, no security headers, no security.txt, no page-speed or resource-hint guidance.",
"verified": "2026-08-21",
"accessed": "2026-08-21"
},
{
"id": "anthropic-computer-use-images",
"title": "Computer use tool",
"url": "https://platform.claude.com/docs/en/docs/agents-and-tools/tool-use/computer-use-tool",
"publisher": "Anthropic",
"type": "vendor-doc",
"keyFindings": "Computer use perceives only screenshots and zoom images, with no accessibility tree or DOM access. A second named agent for which ARIA roles on images are entirely inert.",
"verified": "2026-08-20",
"accessed": "2026-08-21"
},
{
"id": "anthropic-computer-use-tool",
"title": "Computer use tool",
"url": "https://platform.claude.com/docs/en/agents-and-tools/tool-use/computer-use-tool",
"type": "vendor-doc",
"publisher": "Anthropic",
"verified": "2026-08-20",
"keyFindings": "COUNTER-EVIDENCE for a blanket 'agents read the a11y tree' claim. The desktop computer-use tool is screenshot-only: 17 member actions (screenshot, zoom, clicks, scroll, type, key, wait) with no DOM or accessibility-tree access. Doc redirects web work to the browser use tool: 'For tasks that stay inside webpages, the browser use tool is the closer fit: its member tools read and act on the page itself.'"
},
{
"id": "anthropic-contextual-retrieval",
"title": "Introducing Contextual Retrieval",
"url": "https://www.anthropic.com/news/contextual-retrieval",
"type": "vendor-doc",
"publisher": "Anthropic",
"verified": "2026-08-20",
"keyFindings": "Vendor-documented mechanism for why self-contained passages retrieve better. On a chunk lacking its own context: 'However, this chunk on its own doesn't specify which company it's referring to or the relevant time period, making it difficult to retrieve the right information or use the information effectively.' Measured effect of restoring self-containment by prepending context: contextual embeddings alone reduce top-20 retrieval failure rate 35% (5.7% → 3.7%); with contextual BM25, 49% (→2.9%); with reranking, 67% (→1.9%). Note the fix is applied inside the RAG pipeline, not by the publisher — so this proves the mechanism, not the publisher-side lever."
},
{
"id": "anthropic-crawler",
"title": "Does Anthropic crawl data from the web, and how can site owners block the crawler?",
"url": "https://support.claude.com/en/articles/8896518-does-anthropic-crawl-data-from-the-web-and-how-can-site-owners-block-the-crawler",
"publisher": "Anthropic",
"type": "vendor-doc",
"keyFindings": "'Anthropic's Bots respect \"do not crawl\" signals by honoring industry standard directives in robots.txt' for ClaudeBot, Claude-User and Claude-SearchBot; examples shown are root-level Disallow: / only, with path granularity not explicitly confirmed.",
"verified": "2026-08-21",
"accessed": "2026-08-21"
},
{
"id": "anthropic-crawler-docs",
"title": "Does Anthropic crawl data from the web, and how can site owners block the crawler?",
"url": "https://support.claude.com/en/articles/8896518",
"type": "vendor-doc",
"publisher": "Anthropic",
"verified": "2026-08-20",
"keyFindings": "Documents ClaudeBot (training), Claude-User (user-directed fetch), Claude-SearchBot (search indexing). States bots 'respect do not crawl signals by honoring industry standard directives in robots.txt'. COUNTER-EVIDENCE: no mention of sitemaps, RSS, canonical, nofollow, or how URLs are discovered."
},
{
"id": "anthropic-crawler-framework",
"title": "Does Anthropic crawl data from the web, and how can site owners block the crawler?",
"url": "https://support.claude.com/en/articles/8896518-does-anthropic-crawl-data-from-the-web-and-how-can-site-owners-block-the-crawler",
"publisher": "Anthropic",
"type": "vendor-doc",
"verified": "2026-08-21",
"keyFindings": "Provides no technical specification about rendering capabilities or content discovery for ClaudeBot, Claude-User or Claude-SearchBot; scope is robots.txt honoring and Crawl-delay support.",
"accessed": "2026-08-21"
},
{
"id": "anthropic-crawler-support",
"title": "Does Anthropic crawl data from the web, and how can site owners block the crawler?",
"url": "https://support.claude.com/en/articles/8896518-does-anthropic-crawl-data-from-the-web-and-how-can-site-owners-block-the-crawler",
"type": "vendor-doc",
"publisher": "Anthropic",
"verified": "2026-08-21",
"keyFindings": "Canonical URL moved from support.anthropic.com to support.claude.com (301). Three documented tokens: ClaudeBot ('collecting web content that could potentially contribute to their training'), Claude-User ('When individuals ask questions to Claude, it may access websites using a Claude-User agent'), Claude-SearchBot ('navigates the web to improve search result quality'). Blanket compliance statement: 'Anthropic's Bots respect do not crawl signals by honoring industry standard directives in robots.txt'. Explicitly names support for the 'Crawl-delay extension to robots.txt'. IP verification list at https://claude.com/crawling/bots.json. NEGATIVE: 'anthropic-ai' and 'Claude-Web' are NOT mentioned anywhere on the current page."
},
{
"id": "anthropic-crawlers",
"title": "Does Anthropic crawl data from the web, and how can site owners block the crawler?",
"url": "https://support.claude.com/en/articles/8896518-does-anthropic-crawl-data-from-the-web-and-how-can-site-owners-block-the-crawler",
"type": "vendor-doc",
"publisher": "Anthropic",
"verified": "2026-08-21",
"keyFindings": "Anthropic operates three crawlers: ClaudeBot (collects web content for model training and development), Claude-User (retrieves sites when a user directs Claude to specific content), Claude-SearchBot (indexes web content to improve search result quality). Notable NEGATIVE finding: the document contains no guidance whatsoever on how content is selected, ranked or cited by Claude — no mention of authorship, credentials, publication dates, structured data, brand or source authority. It covers only crawler transparency and robots.txt control. This is direct absence-of-documentation counter-evidence for every authority signal in this domain."
},
{
"id": "anthropic-cu-tool",
"title": "Computer use tool — security and prompt injection guidance",
"url": "https://platform.claude.com/docs/en/agents-and-tools/tool-use/computer-use-tool",
"type": "vendor-doc",
"publisher": "Anthropic",
"verified": "2026-08-20",
"keyFindings": "'In some circumstances, Claude will follow commands found in content even when they conflict with your instructions. For example, instructions on webpages or contained in images might override your instructions.' Classifiers run on screenshots to flag injections and force user confirmation. Also recommends asking a human to confirm consequential actions — the vendor-side counterpart to the site-side 'don't hide consequence behind a GET link' check."
},
{
"id": "anthropic-define-tools",
"title": "Anthropic — define tools",
"url": "https://platform.claude.com/docs/en/agents-and-tools/tool-use/define-tools",
"type": "vendor-doc",
"publisher": "Anthropic",
"verified": "2026-08-21",
"keyFindings": "A tool `name` \"Must match the regex `^[a-zA-Z0-9_-]{1,64}$`\", so an operationId carrying spaces, punctuation or more than 64 characters cannot be registered verbatim."
},
{
"id": "anthropic-docs-head",
"title": "Anthropic developer docs page HTML (live head inspection)",
"url": "https://platform.claude.com/en/docs/overview",
"publisher": "Anthropic",
"type": "vendor-doc",
"keyFindings": "27 tags in head, none referencing llms.txt or llms-full.txt — despite Anthropic serving a 33.5 MB llms-full.txt. Discovery is done via a text pointer inside llms.txt instead.",
"verified": "2026-08-20",
"accessed": "2026-08-21"
},
{
"id": "anthropic-docs-md",
"title": "Intro to Claude (markdown alternate of the HTML docs page)",
"url": "https://platform.claude.com/docs/en/intro.md",
"type": "vendor-doc",
"publisher": "Anthropic",
"verified": "2026-08-20",
"keyFindings": "Fetching the HTML page URL with .md appended returns clean markdown with YAML front matter (title/url/description) instead of an HTML document — production evidence of the llms.txt .md-alternate convention on a major vendor's docs. Note the alternate embeds MDX components (, , ), which is a fidelity wrinkle a fidelity check must tolerate."
},
{
"id": "anthropic-llms-txt",
"title": "Anthropic Developer Documentation llms.txt",
"url": "https://platform.claude.com/llms.txt",
"publisher": "Anthropic",
"type": "vendor-doc",
"keyFindings": "Spec-compliant llms.txt that ends with 'For more comprehensive documentation, see llms-full.txt' — an AI vendor both publishing and cross-linking the file. The referenced llms-full.txt returns HTTP 200 at 33.5 MB.",
"verified": "2026-08-20",
"accessed": "2026-08-21"
},
{
"id": "anthropic-tool-use-overview",
"title": "Anthropic — tool use overview",
"url": "https://platform.claude.com/docs/en/agents-and-tools/tool-use/overview",
"type": "vendor-doc",
"publisher": "Anthropic",
"verified": "2026-08-21",
"keyFindings": "\"Claude determines when to call a tool based on the user's request and the tool's description.\""
},
{
"id": "ap2",
"title": "Agent Payments Protocol (AP2)",
"url": "https://ap2-protocol.org/",
"type": "draft-spec",
"publisher": "Google",
"verified": "2026-08-20",
"keyFindings": "Open protocol for the agent economy, positioned as an extension of A2A and the Universal Commerce Protocol. Current terminology is Checkout Mandate (open = user constraints pre-cart; closed = authorization for a finalized checkout) and Payment Mandate (open = payment constraints for autonomous execution; closed = authorization for a specific amount) — note the older 'Intent Mandate'/'Cart Mandate' names are no longer used. Built on Verifiable Digital Credentials (VDCs): tamper-evident cryptographically signed objects forming a non-repudiable audit trail. x402 appears in payment samples. NEGATIVE RESULT: no merchant-side endpoints, agent cards, or well-known URLs are specified on the overview, so AP2 offers no site-auditable surface today."
},
{
"id": "apievangelist-api-catalog-adoption",
"title": "Only Four API Providers Publish a Real .well-known/api-catalog Right Now",
"url": "https://apievangelist.com/blog/2026/05/22/four-providers-publishing-well-known-api-catalog/",
"type": "study",
"publisher": "API Evangelist",
"verified": "2026-08-20",
"keyFindings": "Published 2026-05-22. Methodology: 518 parallel HTTPS requests across 74 providers × 6 host candidates (apex, docs., developer., developers., api., platform.). Result: only FOUR served a valid RFC 9727 Linkset — Cloudflare (1 entry), Memesio (2), Merge.dev (10), Zuplo (1). Two clean 404s; SIXTY-EIGHT returned 200 OK with an HTML body — 'an HTML 200 at a well-known path lies, which is worse than a 404.' Stripe and Twilio have not implemented it."
},
{
"id": "apple-app-search-web-markup",
"title": "App Search Programming Guide: Mark Up Web Content",
"url": "https://developer.apple.com/library/archive/documentation/General/Conceptual/AppSearch/WebContent.html",
"type": "vendor-doc",
"publisher": "Apple (Documentation Archive)",
"verified": "2026-08-21",
"keyFindings": "ARCHIVED documentation. \"To annotate your web content so that users can see rich search results, use standards-based markup for structured data, such as that defined at Schema.org.\" Applebot-supported schemas listed: AggregateRating, Offers, PriceRange, InteractionCount, Organization, Recipe, SearchAction, ImageObject; Open Graph also supported. Ranking claim: \"the popularity of a URL and the amount of structured data available\" contribute to relevancy/ranking in Spotlight/Safari/Siri. The only vendor doc naming SearchAction as consumed — but it is archived and predates Apple Intelligence."
},
{
"id": "apple-applebot-about",
"title": "About Applebot",
"url": "https://support.apple.com/en-us/119829",
"type": "vendor-doc",
"publisher": "Apple",
"verified": "2026-08-21",
"keyFindings": "Page resolves (title: 'About Applebot - Apple Support') and is Apple's canonical Applebot reference documenting Applebot and Applebot-Extended robots.txt tokens. CAVEAT: page body is JS-rendered and could not be extracted verbatim by automated fetch; only the document title was recoverable. Treat quoted specifics as unverified by us."
},
{
"id": "apple-applebot-archived-2025",
"title": "About Applebot — archived snapshot, 2 March 2025 (Wayback Machine)",
"url": "https://web.archive.org/web/20250302012726/https://support.apple.com/en-us/119829",
"type": "vendor-doc",
"publisher": "Apple / Internet Archive",
"verified": "2026-08-20",
"keyFindings": "Verbatim: 'Applebot accesses many kinds of resources from web servers, including but not limited to robots.txt, sitemaps, RSS feeds, HTML, sub resources needed to render pages such as javascript, Ajax requests, images, and more.' This is the strongest single vendor statement that an AI-serving crawler (Applebot powers Siri, Spotlight, Safari and Apple Intelligence grounding) reads sitemaps AND RSS feeds. Removed in the June 2026 revision."
},
{
"id": "apple-applebot-docs",
"title": "About Applebot",
"url": "https://support.apple.com/en-us/119829",
"type": "vendor-doc",
"publisher": "Apple",
"verified": "2026-08-21",
"keyFindings": "Strongest single AI-specific head-tag source. \"Applebot supports robots meta tags in HTML documents.\" Directives: noindex (\"won't appear in Spotlight or Siri Suggestions\"); nosnippet (\"Applebot won't generate a description or web answer for the page. Any suggestions to visit this URL will only include the page's title. Apple will not use data tagged nosnippet as additional context and up-to-date content when AI models are used to generate output for display in Apple products and services\"); nofollow; none; all. Supports the applebot-specific meta name and X-Robots-Tag. Ranking factors listed are engagement, relevancy, links, location, webpage design characteristics — no head-tag factors named."
},
{
"id": "apple-applebot-training-privacy",
"title": "Applebot model training and individual privacy rights",
"url": "https://support.apple.com/en-us/120320",
"type": "vendor-doc",
"publisher": "Apple",
"verified": "2026-08-20",
"keyFindings": "Apple's stated position: web publishers can use standard robots.txt directives to direct Applebot not to crawl, or to direct Apple not to use their content to train Apple's foundation models (Applebot-Extended). Applebot-Extended is a training/AI-use opt-out token only and does not remove a site from Apple search surfaces (Siri/Spotlight/Safari Suggestions). CAVEAT: page body JS-rendered; title-only extraction succeeded."
},
{
"id": "applebot",
"title": "About Applebot",
"url": "https://support.apple.com/en-us/119829",
"publisher": "Apple",
"type": "vendor-doc",
"keyFindings": "Explicit path-level examples: 'User-agent: Applebot / Allow: / / Disallow: /private/' and, for the generative-AI training token, 'User-agent: Applebot-Extended / Disallow: /private/'. Confirms AI-relevant crawlers honor per-directory rules.",
"verified": "2026-08-21",
"accessed": "2026-08-21"
},
{
"id": "applebot-doc",
"title": "About Applebot",
"url": "https://support.apple.com/en-us/119829",
"publisher": "Apple",
"type": "vendor-doc",
"keyFindings": "Applebot powers Spotlight, Siri, Safari search and Apple Intelligence. The ONLY schema.org property Apple documents supporting is isAccessibleForFree ('Applebot supports the schema.org isAccessibleForFree property'). No speakable, no Actions, no HowTo. Directly refutes the audit's 'Siri uses speakable' claim.",
"verified": "2026-08-21",
"accessed": "2026-08-21"
},
{
"id": "ard-conformance-example",
"title": "ARD conformance example basic/ai-catalog.json",
"url": "https://raw.githubusercontent.com/ards-project/ard-spec/main/conformance/examples/basic/ai-catalog.json",
"publisher": "ARD Project",
"type": "spec",
"keyFindings": "Top-level keys are specVersion, host, entries. Entries carry identifier, displayName, type (media type), url, description, representativeQueries, capabilities. No `services`, `owner`, `contact`, or `lastUpdated`.",
"verified": "2026-08-20",
"accessed": "2026-08-21"
},
{
"id": "ard-spec",
"title": "Agentic Resource Discovery (ARD) Specification",
"url": "https://agenticresourcediscovery.org/spec/",
"type": "draft-spec",
"publisher": "ARDS Project",
"verified": "2026-08-20",
"keyFindings": "v0.9 Draft, status 'Proposal', dated 2026-05-28. Manifest ai-catalog.json requires specVersion, host, entries. Each entry requires identifier (urn:air:::), displayName, type, and exactly one of url|data. Four discovery mechanisms: /.well-known/ai-catalog.json, robots.txt 'Agentmap:', , DNS SVCB records."
},
{
"id": "ard-spec-discovery",
"title": "ARD Specification §6.1 — discovery mechanisms",
"url": "https://raw.githubusercontent.com/ards-project/ard-spec/main/spec/ard.md",
"publisher": "ARD Project (Google/Microsoft/Hugging Face authors)",
"type": "spec",
"keyFindings": "Four advertisement mechanisms: well-known URI; robots.txt 'Agentmap:' entry; ' in the of a document'; DNS SVCB with TXT fallback (DNS-AID draft).",
"verified": "2026-08-20",
"accessed": "2026-08-21"
},
{
"id": "ard-spec-entry-url",
"title": "ARD Specification — entry schema (url vs data)",
"url": "https://raw.githubusercontent.com/ards-project/ard-spec/main/spec/ard.md",
"publisher": "ARD Project",
"type": "spec",
"keyFindings": "Each entry carries either a `url` remote reference or embedded `data`; `type` is an IANA media type identifying the artifact. Confirms url is the field to validate, and that entries without url are legal.",
"verified": "2026-08-20",
"accessed": "2026-08-21"
},
{
"id": "ard-spec-fields",
"title": "ARD Specification §4.1 manifest fields",
"url": "https://raw.githubusercontent.com/ards-project/ard-spec/main/spec/ard.md",
"publisher": "ARD Project (Google/Microsoft/Hugging Face authors)",
"type": "spec",
"keyFindings": "Required top-level: specVersion, host, entries. Entry fields: identifier (URN), displayName, type, url|data, description, capabilities, representativeQueries; optional version, updatedAt, tags, metadata, trustManifest. No owner/contact/lastUpdated; no services.",
"verified": "2026-08-20",
"accessed": "2026-08-21"
},
{
"id": "ard-spec-repo",
"title": "ards-project/ard-spec (repository)",
"url": "https://github.com/ards-project/ard-spec",
"type": "repo",
"publisher": "ards-project / Linux Foundation",
"keyFindings": "§4.2 requires exactly one of `url` or `data` per entry, so an entry that embeds its artifact has no endpoint to dereference and is still fully conformant.",
"verified": "2026-08-24"
},
{
"id": "ardvark-crawler",
"title": "ardvark — ARD crawler/probe (GitHub code search hits)",
"url": "https://github.com/HelgeSverre/ardvark",
"publisher": "HelgeSverre",
"type": "repo",
"keyFindings": "Code search shows internal/crawler, internal/probe, internal/fetch and internal/mcpserver all referencing ai-catalog.json — an independent crawler that fetches and probes catalogs and their URLs. (Identified via GitHub code search; repo page not individually fetched.)",
"verified": "2026-08-20",
"accessed": "2026-08-21"
},
{
"id": "aria-apg-dialog-modal",
"title": "ARIA Authoring Practices — modal dialog pattern",
"url": "https://www.w3.org/WAI/ARIA/apg/patterns/dialog-modal",
"type": "spec",
"publisher": "W3C",
"verified": "2026-08-21",
"keyFindings": "Requires a dialog to have \"a value set for the aria-labelledby property that refers to a visible dialog title\" or \"a label specified by aria-label\"."
},
{
"id": "aria-in-html-address",
"title": "ARIA in HTML",
"url": "https://www.w3.org/TR/html-aria/",
"publisher": "W3C",
"type": "spec",
"keyFindings": "Assigns `address` the implicit `role=group`. This is the only formal machine semantics the element carries, and an unnamed group conveys no contact-information meaning; Chromium in practice exposes it as generic.",
"verified": "2026-08-20",
"accessed": "2026-08-21"
},
{
"id": "aria-in-html-aside",
"title": "ARIA in HTML",
"url": "https://www.w3.org/TR/html-aria/",
"publisher": "W3C",
"type": "spec",
"keyFindings": "Ratified spec assigning `aside` the implicit `role=complementary` (and `address` the implicit `role=group`). Gives