| 开发者 | ajmaurya |
|---|---|
| 更新时间 | 2026年11月12日 15:54 |
| PHP版本: | 7.4 及以上 |
| WordPress版本: | 7.1 |
| 版权: | GPL-2.0-or-later |
| 版权网址: | 版权信息 |
llms.txt that AI agents like ChatGPT, Claude, and Perplexity read.
It makes your site legible to AI agents — and proves it. The built-in readiness scanner runs 33 checks against your live site, scores you 0–100 with a letter grade (A+ to F) and on the Level 0–5 agent-readiness ladder (Not Ready → Basic Web Presence → Bot-Aware → Agent-Readable → Agent-Integrated → Agent-Native), and shows the full evidence trail — every request and response — behind every verdict. Failing checks get a Fix now button that enables the right feature and re-scans that single check to prove it went green; two more (Organization details, a real MCP server) get a guided button to the right screen; anything the plugin can't fix in WordPress (DNS records, server config) gets a copy-paste prompt for your coding agent.
This is the technical, agent-readiness layer of answer engine optimization (AEO) and generative engine optimization (GEO): it improves your AI visibility by making your content legible to AI agents through llms.txt, JSON-LD, Markdown negotiation, and AI crawler rules. It is not a content SEO plugin — it doesn't write titles or meta descriptions, and it doesn't track brand mentions or AI citations. It prepares your WordPress site so AI search engines and agents can read, trust, and route on it, then verifies that over real HTTP.
The scan → fix → verify loop
/.well-known/*, static-file rules 404 /llms.txt). When that happens the Dashboard says so and hands you copy-paste nginx and Apache fixes.[agent_ready_badge] shortcode (and an opt-in public /agent-ready-badge.svg) shows the score, grade and Level as a static SVG; on request, the Dashboard looks up this site's score on ora.ai and is-agentic.com (read-only, keyless) next to ours.POST /wp-json/ajaco/v1/scan, /scan/check, /fix, GET /external, POST /llms/preview, a public GET /health; wp agent-ready scan|status|fix with --format=summary|json|agent and --min-score / --min-level CI gates that exit non-zero.llms.txt from fixed rules and hand you the result. This one hands you the controls.
/llms.txt will serve, with entry, byte and token counts, before you save./llms-full.txt serves the full content of the same curated entries. Password-protected content is always excluded from both./.well-known/api-catalog linkset advertising your REST API, plus a Link: rel="api-catalog" header on every response./.well-known/mcp/server-card.json describing the site to MCP-aware agents. The plugin does not run an MCP server itself. When the official WordPress MCP Adapter (or any plugin registering REST routes under the mcp namespace) is active, the card advertises its real Streamable HTTP endpoint and the Application Password login; otherwise it states plainly that the REST API is the only machine interface.mcpServer check connects to it (JSON-RPC initialize) to prove it answers. Without one, the check fails with an "Install the MCP Adapter" button (the adapter needs WordPress 6.9+)./.well-known/ard.json (and the predecessor /.well-known/ai-catalog.json) indexing every enabled agent-facing endpoint, linked from every page with rel="ard" / rel="ai-catalog"./.well-known/agent-skills/index.json listing six skills plus per-skill SKILL.md artifacts with verifiable sha256 digests.<link rel> tags and Link headers for the sitemap, OpenAPI document, llms.txt and catalogs, so agents find every file from the pages they already visit.<lastmod> dates to the WordPress core sitemap on versions that lack them; a no-op where core or an SEO plugin already provides them.Accept: text/markdown, the page is served as clean Markdown that opens with a YAML frontmatter block (title, url, site, modified), with X-Markdown-Tokens for context budgeting. Every page advertises its Markdown twin with <link rel="alternate" type="text/markdown"> and a matching Link header, and missing pages return a real 404 as Markdown. Browsers (which send text/html) are unaffected.sameAs, contactPoint and address to the Organization node./openapi.json returns a complete spec generated dynamically from rest_get_server(), including plugin-registered REST routes.document.modelContext.registerTool() per the W3C WebMCP draft (Chrome origin trial), each with a typed input schema and a readOnlyHint annotation, with a fallback for the earlier preview API.Content-Signal: ai-train=no, search=yes, ai-input=no directive to robots.txt per contentsignals.org. Composes with Yoast/Rank Math/AIOSEO — their additions are preserved, our directive lands last./auth.md honestly documenting how agents authenticate to your REST API via Application Passwords (creation, Basic-auth usage, scope, revocation), and the MCP endpoint when one is detected. No fictional OAuth endpoints.llms.txt plugins only generate a file from fixed rules. AJ Agent Crawl Optimizer also scans your live site over real HTTP, scores it 0–100 with a grade and the Level 0–5 ladder, shows the request/response evidence behind every verdict, and fixes failures in one click — then re-scans to prove the check turned green. It writes per-bot robots.txt rules for 20 AI crawlers and lets you hand-curate what AI agents read instead of auto-dumping every post. Scan → score → fix → verify, not generate-and-hope.
Why use it
/{subsite}/... paths automatically.aria-label, copy buttons announce success via aria-live, and decorative icons are hidden from screen readers.
== Installation ==aj-agent-crawl-optimizer folder to /wp-content/plugins/./{key}.txt for ownership verification automatically.Run the built-in readiness scanner. AJ Agent Crawl Optimizer checks your live site against 33 AI-agent-readiness checks — Cloudflare's 21 isitagentready.com checks plus 12 modelled on what ora.ai / is-agentic.com grade for — scores you 0–100 with a grade and a Level 0–5 ladder from Not Ready to Agent-Native. Every verdict shows the real request and response behind it, and failing checks get a one-click Fix now button that re-scans to prove it passed.
llms.txt is a plain-Markdown file at /llms.txt that gives AI agents a curated index of your most important content, proposed at llmstxt.org. AJ Agent Crawl Optimizer lets you hand-curate it: write the intro, pick which post types (and WooCommerce products) appear, and set a per-post "Summary for AI agents" that overrides the excerpt. It is a surface AI agents can route on — not a guaranteed path to AI citations, but a machine-readable map of what your site is for.
The plugin writes explicit robots.txt rules for 20 AI crawlers — GPTBot, ChatGPT-User and OAI-SearchBot (OpenAI), ClaudeBot, Claude-User and Claude-Web (Anthropic), PerplexityBot and Perplexity-User, Google-Extended, CCBot and Bytespider among them — each allowed or blocked per a per-bot policy. Presets like "Allow search & user requests, block training" make it a single choice. Remember: to be cited by an AI engine you must allow its crawler.
There is no guaranteed path, but three things help and this plugin does all three: allow the AI search crawlers (OAI-SearchBot, PerplexityBot, Google-Extended), publish the machine-readable surfaces agents look for (llms.txt, JSON-LD, Markdown negotiation), and write a clear "Summary for AI agents" on key pages. The readiness scanner then verifies each of these is actually live over real HTTP — so you fix what's broken instead of guessing.
/llms.txt is a curated index — a Markdown list of your key pages with links and short summaries, so an agent can decide what to fetch. /llms-full.txt serves the full Markdown content of those same curated entries in one file, for agents that want everything inline. The plugin serves both, and password-protected content is always excluded from each.
Most plugins only generate an llms.txt file from fixed rules. AJ Agent Crawl Optimizer is an agent-readiness scanner and fixer: it audits your live site over real HTTP against 33 checks, scores it 0–100 with a grade and a Level 0–5 ladder, shows the request/response evidence behind every verdict, fixes failures in one click, and re-verifies. It covers the answer engine optimization (AEO) and generative engine optimization (GEO) groundwork — structured data, AI crawler rules, and MCP discovery — that plain generators and content-writing SEO plugins skip.
Yes. WooCommerce products are detected automatically and can appear as their own section in your curated llms.txt, so AI agents can see your catalogue. You control whether products appear, how many, and in what order, and you can exclude individual products or give any product a per-product "Summary for AI agents."
Dashboard is the verifier: it scans your live site over real HTTP, shows your score, grade and Level 0–5 with evidence for every check, and fixes failures with one click. Settings is the switchboard: the per-feature toggles, wizard, and testing tools. A toggle can be ON in Settings while the Dashboard still shows a fail — that gap usually means a page cache or server rule is intercepting the endpoint, and the evidence view shows exactly what happened.
The scanner runs Cloudflare's 21 isitagentready.com checks plus 12 checks modelled on the ora.ai / is-agentic.com catalog and the AgentReady.org standard. The 0–100 score weights every check; checks that don't apply or can't be reached drop out of the denominator, bonus checks (Web Bot Auth, DNS-AID, A2A card, commerce protocols) add up to 5 points and never subtract, and grades use the ora.ai bands (A+ 95, A 86, B 70, C 48, D 28). The Level 0–5 ladder stays identical to Cloudflare's: Level 1 needs 2 of robots.txt/sitemap/Link headers; Level 2 adds AI bot rules and Content Signals; Level 3 adds Markdown negotiation; Level 4 adds one of MCP card/Agent Skills/API catalog/A2A card; Level 5 adds two of Web Bot Auth, all four integrations, and auth metadata. Nothing counts unless the scan verifies it — enabling a toggle earns nothing until the endpoint actually responds. Commerce checks are informational and never scored.
The plugin never creates files: every endpoint (/.well-known/, /llms.txt, /openapi.json, /auth.md, /agent-ready-badge.svg) is served virtually by WordPress, so hosts that forbid file creation are fully supported. What can break things is the web server intercepting requests before WordPress runs — typically an nginx dot-path deny rule returning 403 on /.well-known/, or a static-file rule returning 404 for /llms.txt-style paths. The readiness scan detects exactly this (feature enabled + server blocking) and the Dashboard shows a hosting notice with copy-paste nginx and Apache fixes you can apply yourself or forward to your hosting support.
Yes — it is curated, not auto-generated from fixed rules. Under Agent Ready → llms.txt you write the intro, pick which post types appear (custom post types and WooCommerce products show up automatically), and set each section's heading, item count and order, plus your own Markdown block. A live preview shows the file as you edit, before you save.
Individual entries are steered from the editor: the "Agent Ready (llms.txt)" panel has an Include in llms.txt toggle (excluded pages stay out of both /llms.txt and /llms-full.txt) and a Summary for AI agents field that overrides the excerpt for that entry — write it for a model deciding whether to fetch the page. The defaults keep the same sections and ordering as before, so upgrading needs no attention until you decide to curate.
Yes: wp agent-ready scan (add --format=json for machines or --format=agent for a markdown fix report), wp agent-ready fix <check> or --all-safe, and wp agent-ready status. Add --min-score=<n> or --min-level=<n> to scan or status for a CI gate that exits non-zero. The same operations are available over REST at /wp-json/ajaco/v1/scan, /wp-json/ajaco/v1/fix and /wp-json/ajaco/v1/external (admin capability required; /wp-json/ajaco/v1/health is public).
No. JSON-LD and the identity meta tags auto-suppress when Yoast SEO, Rank Math, All in One SEO, SEOPress, The SEO Framework, Slim SEO, Squirrly SEO, Schema Pro, or SASWP is active. The settings page shows a clear notice when this happens. Our robots_txt filter runs at PHP_INT_MAX priority, so any rules your SEO plugin adds (sitemap URLs, custom Disallow, etc.) are preserved — our Content-Signal line is appended last.
No. Browsers send Accept: text/html and get the normal HTML response. The .well-known/..., /llms.txt, and /openapi.json endpoints are paths a normal user never visits. What the enabled features add to a page is invisible to visitors: a JSON-LD <script> block and Open Graph meta tags (when no SEO plugin is active), a few <link rel> tags and Link headers pointing at the agent-facing files, and a small WebMCP script tag that no-ops in browsers without the API.
Yes. Every endpoint resolves at both the root and per-subsite paths automatically — /llms.txt, /blog/llms.txt, /.well-known/api-catalog, /blog/.well-known/api-catalog, etc. Each subsite has its own settings.
Use the ajaco_required_capability filter. The plugin defaults to manage_options (admin-only); change it to any capability of your choice, e.g. edit_pages for editors:
add_filter( 'ajaco_required_capability', function () { return 'edit_pages'; } );
Check three things: (1) the IndexNow API Key field is filled in, (2) you're publishing a post of a public post type — revisions and autosaves are skipped, (3) you're on production. Pings are non-blocking, so failures are silent. Tail your server log for outbound requests to api.indexnow.org.
Yes. The plugin exposes nineteen filter hooks. Examples:
add_filter( 'ajaco_required_capability', function () { return 'edit_posts'; } );
Delegate plugin access to a non-admin role.
add_filter( 'ajaco_skill_definitions', function ( $skills ) { return $skills + [ 'products' => [ 'type' => 'information-retrieval', 'description' => 'WooCommerce products', 'endpoint' => rest_url( 'wc/v3/products' ) ] ]; } );
Register custom skills that ship in the Agent Skills Index as SKILL.md artifacts with verifiable sha256 digests.
add_filter( 'ajaco_content_signal', function () { return 'ai-train=yes, search=yes, ai-input=yes'; } );
Customize the Content-Signal directive (e.g. permit AI training).
Other hooks: ajaco_json_ld_graph (add Product, Recipe, Event… schema), ajaco_llms_txt_content and ajaco_llms_full_txt_content (append to or replace the files), ajaco_llms_post_types, ajaco_llms_exclude_post, ajaco_openapi_spec, ajaco_api_catalog_linkset, ajaco_ai_catalog, ajaco_mcp_server_card, ajaco_mcp_endpoint (declare an MCP server the plugin cannot detect), ajaco_auth_md_content, ajaco_ai_bot_list, ajaco_ai_bot_policy, ajaco_commerce_signals, ajaco_scan_sslverify, ajaco_active_seo_plugin. The Help tab's "For Developers" section on the settings page lists every hook with a description.
Per-post curation is stored in two post meta fields: _ajaco_llms_exclude (boolean) and _ajaco_llms_summary (string, 300 chars). Both are registered with show_in_rest, so they round-trip through the REST API and WP-CLI (wp post meta update 42 _ajaco_llms_summary "...").
Use the Reset to Defaults button under the Save Changes button on the settings page. It clears every plugin option, wipes cached endpoint outputs, and shows a confirmation dialog before wiping anything.
Minimal. The OpenAPI document is cached for a day in a transient and only regenerates on plugin activation/deactivation or theme switch. The llms.txt body is cached for an hour and invalidates on post edits, site name/description changes, and setting toggles. Endpoint handlers run on init only when the request path matches; on every other request they early-return after one path check. Markdown conversion happens in an output-buffer callback for requests that ask for Markdown, so normal page loads are untouched. The hosted-scanner lookup runs only when you click it and is cached for an hour.
The plugin doesn't include a built-in access log. To verify activity, check your server access log for User-Agents like GPTBot, ClaudeBot, OAI-SearchBot, Google-Extended, PerplexityBot, CCBot, Bytespider, etc. hitting any of the plugin endpoints.
Yes. Every feature is an independent toggle. Uncheck what you don't want and Save Changes. The corresponding behavior reverts immediately on the next request — no rewrite flush, no cache flush needed (the plugin handles cache invalidation itself).
Only in three situations you trigger yourself; otherwise the plugin makes no outbound requests at all. Publishing features (manifests, JSON-LD, robots.txt) only answer incoming requests.
initialize request to your own MCP endpoint, and the DNS-AID check queries public DNS-over-HTTPS resolvers (https://cloudflare-dns.com/dns-query, with https://dns.google/resolve as fallback) for _agents records of your own domain. Only query names derived from your domain are sent, never visitor data. Policies: https://developers.cloudflare.com/1.1.1.1/privacy/public-dns-resolver/ and https://developers.google.com/speed/public-dns/privacyhttps://api.indexnow.org/indexnow containing exactly your site host, your IndexNow key (generated by you at Bing Webmaster Tools) and the published URL. No visitor data or content. Terms and policies: https://www.indexnow.org/documentation, https://www.indexnow.org/faq, https://www.bing.com/webmasters/help/webmaster-guidelines-30fba23a, https://www.microsoft.com/en-us/privacy/privacystatementhttps://ora.ai/api/score/{host} and https://is-agentic.com/api/v1/report?url=…), sending your host name / home URL and nothing else; the answer is cached for an hour. These APIs never start a scan of your site. Policies: https://ora.ai/privacy and https://is-agentic.com/Everything lives on your own server — no telemetry, analytics or third-party logging. Option rows store toggle states, the per-bot crawler policy, the Organization details you enter, the llms.txt curation config, the most recent scan result (with evidence snapshots of your own site's responses), a compact score history (at most 30 entries) and the IndexNow key (plain text — keep your database secure). Transients cache the OpenAPI document, the llms.txt bodies and the last hosted-scanner lookup. No request data, IP addresses, User-Agents or visitor information is recorded, and the plugin sets no cookies. Deleting the plugin (not just deactivating) removes every option, the per-post curation meta and all transients; multisite networks have every site cleaned in turn.
mcp namespace, or the new ajaco_mcp_endpoint filter). The server card, agent resource catalog, API catalog linkset, llms.txt and auth.md then advertise the real Streamable HTTP endpoint; the mcpServer check sends a JSON-RPC initialize and passes on a valid result or an authentication challenge. Without a server the card now says plainly that the REST API is the only machine interface, and the check offers an "Install the MCP Adapter" button (WordPress 6.9+).document.modelContext.registerTool() with typed input schemas, titles and readOnlyHint annotations (search_site_content, list_recent_posts, list_site_pages, get_site_info), with a fallback to the earlier navigator.modelContext.provideContext() preview. The WebMCP check recognises both surfaces.<link rel> tags and Link headers for the sitemap, OpenAPI, llms.txt and catalogs, a rel="alternate" type="text/markdown" link on every page, YAML frontmatter on served Markdown, and Markdown 404 pages.[agent_ready_badge] shortcode, plus an opt-in public /agent-ready-badge.svg for embedding elsewhere.GET /wp-json/ajaco/v1/external.--min-score=<n> and --min-level=<n> on scan and status exit non-zero below the threshold.ajaco/v1 — POST/GET /scan, POST /scan/check, POST /fix, and a public GET /health (also the API catalog's RFC 9727 status target).wp agent-ready scan|status|fix [--all-safe] with summary, json, and agent (markdown fix report) formats._ajaco_llms_exclude / _ajaco_llms_summary post meta (REST-exposed, so WP-CLI and the REST API can set them).type: skill-md and a digest field with sha256: prefix (previously type: information-retrieval and a sha256 key, which external scanners rejected); $schema corrected to the published schema URI.endpoint (SEP-1649 structure) with type streamable-http; protocolVersion bumped to 2025-06-18.Vary: Accept on all frontend responses so page caches and CDNs key correctly on the Accept header; buffer unwind at shutdown is now bounded (no hang on non-removable output buffers); str_contains() replaced for WP < 5.9 compatibility./openapi.json path (query-var form kept as an alias).service-desc only advertised when the OpenAPI feature is enabled (no more dead link), and its media type now matches what the endpoint serves.User-agent: * group so its scope is deterministic under RFC 9309 group semantics.Tom's Blog & Café instead of Tom's Blog & Café.