Pitwall AI Manifest for llms.txt

Description

Pitwall AI Manifest does two jobs with the same content. For your visitors, it is a chat assistant that answers their questions from your own pages, through the WordPress AI Client or with no AI at all. For AI assistants and agents, it publishes your site in a form they can read accurately: a curated llms.txt, clean Markdown copies of your pages, and a search API.

llms.txt is a file format: a short, curated description of a site that sits at /llms.txt. Most plugins that write it dump every post you have ever published into one file. That is the opposite of what the format is for: it exists to fit a site into an AI’s limited context window, so twenty well-described links beat eight hundred bare ones.

This plugin is built around three ideas.

It asks you to write the part that matters. The Site context tab captures what no crawler can work out: who you are, what you sell, who you serve, where you operate, and who you are not. That last field stops a model from confidently giving you a similarly named company’s prices, locations or reviews.

It curates instead of listing everything. You define sections with rules (post type, category or tag, order, a limit), pin the handful of pages that define the business, and feature or exclude single posts from the editor sidebar. There is a size budget, and the file stops cleanly when it reaches it.

It serves clean Markdown. Every public post and page is also available at its address plus .md: the title, a few facts and the text, without menus, share buttons, comment forms or related-post widgets. An AI agent fetching your page gets a few kilobytes of substance instead of a few hundred kilobytes of HTML, whether or not it ever reads llms.txt.

An honest note on adoption

No major AI company has committed to reading llms.txt, and server logs show AI crawlers asking for it rarely. Treat it as a cheap, low-risk bet on machine-readable publishing, and treat the .md copies as the part that helps today. This plugin does not promise you rankings in AI search, because nobody can.

Features

• A curated llms.txt at /llms.txt, following the llmstxt.org structure
• An optional llms-full.txt with the full text of every post in llms.txt
• Clean Markdown copies at /your-page.md, linked from each page’s head, with a canonical link back to the page
• Hand-written organisation context: summary, description, products and services, audience, geography, languages, who you are not, how to cite you
• “Scan my site” fills empty fields from your site title, SEO plugin settings, WooCommerce store address, translation plugin and contact page
• Section rules with post type, category or tag, order and a limit per section
• Pinned links chosen from a searchable list of your pages, worked out again on every rebuild, so a changed address never leaves a dead link
• Pinned-link suggestions from your main menu
• Feature, exclude or describe single posts from the editor sidebar
• Reads meta descriptions and noindex settings from Yoast SEO, Rank Math and SEOPress, for each post and for each post type
• Rebuilds wait a few minutes after a change, so saving posts stays fast, and visitors get the last copy while a rebuild runs
• Size budgets on both files
• Conditional requests (ETag and Last-Modified), so repeat fetches are cheap
• A live preview in the admin, and WP-CLI commands

Structured data

The same context also produces a schema.org Organization block in your page head: name, description, logo, social profiles, address, languages, area served and founding date. One round of typing, two results.

The difference is worth being clear about: llms.txt is a bet on where things are going, while structured data is read by search engines today.

If Yoast SEO, Rank Math, SEOPress or All in One SEO is active, the plugin says so, because two Organization blocks that disagree are worse than one. Leave this off unless you have switched the other one off first.

Setting up many sites

Set one site up properly, export it as a template, and apply it to the next site in a minute. A template carries the decisions about how the plugin works: section rules, output settings, the search API, the schema type, and how the chat widget answers and looks.

It leaves out everything about one business: the site context, the address, the pinned links (they point at post IDs, which are other pages on another site), the curated answers, and the widget’s own words and contact details. “Export everything” is there as a backup of one site.

An import replaces only what the file holds and goes through the same checks as the settings form, so an edited file cannot bring in a value you could not have typed.

Search API

The plugin also splits your content into passages and indexes them, so any assistant can search the site: a custom GPT, an MCP server, the chat widget, or an automation in n8n or Make.

GET /wp-json/ltxt/v1/search?q=...     Matching passages as JSON, with their addresses
GET /wp-json/ltxt/v1/context?q=...    The same, as one Markdown block ready for a prompt
GET /wp-json/ltxt/v1/manifest         Your llms.txt with the site name and summary
GET /wp-json/ltxt/v1/post/{id}        One post in full, as Markdown

Passages follow your headings, about 250 words each, and are searched with your database’s full-text index. No account, no key and no cost per search: it all runs on your own database. Each result links to the heading it came from.

Requests are limited per visitor address and can require a key. The endpoints only return content that is already public, so the key is for keeping casual traffic away and cutting off one integration without touching the others.

Chat widget

A chat window on your site that answers visitors’ questions. Two of its four ways of answering cost nothing to run, because your site already holds the answers.

• Curated answers only. Free. You write the questions you get asked and the exact answers. Nothing can be invented, it never changes, and it is instant.
• Curated answers, then site search. Free, and the default. When no curated answer fits, it quotes the closest passage from your site with a link. It quotes rather than rewrites, so nothing is invented, but the wording has to match: “shipping” does not find a page that only says “delivery”.
• Curated answers, then AI. Paid only when no curated answer fits. Your curated answers handle the questions they match for free; anything else goes to the AI, which answers from the passages your site search finds.
• AI for every question. Paid on every question. Every question goes to the AI, with the passages and, when one fits, your curated answer as the wording to use. Replies read naturally and follow the conversation.

The AI answers are grounded in your own content and told never to state a price, time, phone number or policy that is not in it. Page text is marked as content rather than instruction, so a page cannot talk the assistant out of its rules, and the conversation history sent back by the browser only counts when this site signed it.

Cost controls, because a public chat box is something anyone can call: a daily limit on AI answers (past it, the assistant answers the free way instead of going quiet), a limit per visitor per minute, a cap on how much of the conversation is remembered, and a ceiling of 1,024 tokens per answer by default.

AI answers go through the WordPress AI Client only. On WordPress 7.0 and later, the AI provider you set up for the whole site under Settings, Connectors answers, on your account with it, and WordPress picks the model. This plugin never contacts an AI service itself and has no AI key of its own. The Chat widget tab says whether a provider is set up, and links to Settings, Connectors.

Without it, the AI features switch off and nothing else changes. On WordPress before 7.0, or with no AI provider set up, the two ways of answering that use AI answer the free way instead (a curated answer, or a passage with a link), a notice in the admin says how to switch the AI on, and everything else in the plugin works as usual.

Talking to a person

Every assistant meets a question it cannot answer, and “I don’t know” on its own is a dead end. Add a WhatsApp number (with a first message), an email address or a contact page, and the widget offers them: after a question it could not answer, by default, so it appears exactly when it helps and only once per conversation.

Smarter answers, still free

The free way of answering knows which words mean the same (“cost”, “price”, “preço” and “Preis” find the same answer) in English, Portuguese, Spanish, French and German, and takes your own groups of words too. It forgives a typo, follows a short follow-up (“and an extra driver?”), quotes the one or two sentences that answer rather than a block of text, and offers the curated question that came close (“Did you mean…?”) when none quite fits.

Starting the conversation

Up to three suggested questions under the greeting, for visitors to tap: your own, or your first three curated questions. A nudge, a small bubble by the launcher after a few seconds on chosen pages (“Questions about the car?”), once per visit. Answers are formatted: bold, lists and links.

A conversation that follows the visitor

The conversation stays with the visitor from page to page in the same browser tab. Follow a link in an answer and the chat is open on the next page, mid-conversation.

Multilingual sites

With Polylang or WPML, the widget’s words and the curated answers are translated where you translate everything else, and each visitor gets them, and answers from the pages, in the language of the page.

Look

The brand colour takes a hex code, var(--your-theme-colour), rgb(), hsl() or a colour name. A CSS variable is the one worth using: point it at your theme’s brand colour and the widget follows a rebrand by itself.

Also: three sizes, light, dark or follow the device, rounded or square corners, either bottom corner, words beside the launcher icon, an avatar image or the business’s initials, and a subtitle. Movement is subtle and switches off for visitors who ask for reduced motion.

Trying it before you go live

Who can see it has three settings: everyone, logged-in users, or administrators only. Start with administrators and the widget appears on the real site for you and nobody else, with a preview badge so you cannot forget. The rule is enforced by the endpoint too, so it is private rather than hidden.

Visitor questions

Every question is logged under AI Manifest, Questions, with a filter for the ones the site could not answer: content gaps, questions real visitors asked that your website does not answer yet. The log can be exported as CSV, and a weekly email can bring you the unanswered ones every Monday.

Entries hold the question and the answer, and no IP address, cookie or visitor identifier. Visitors can still type personal details into a chat box, so treat the log as personal data and keep entries no longer than useful. It deletes old entries every day.

WP-CLI

wp ai-manifest generate       # rebuild llms.txt now
wp ai-manifest show           # print llms.txt
wp ai-manifest show --full    # print llms-full.txt
wp ai-manifest markdown 42    # print one post as Markdown
wp ai-manifest index          # rebuild the search index
wp ai-manifest search "..."   # search the index
wp ai-manifest flush          # clear the cache and the rewrite rules

Hooks

Filters:

• Settings and output: ltxt_settings, ltxt_document, ltxt_full_document, ltxt_credit_line, ltxt_agency
• What gets in: ltxt_section_posts, ltxt_include_post, ltxt_can_serve, ltxt_is_noindex, ltxt_seo_description_keys, ltxt_serve_markdown
• Markdown: ltxt_post_markdown, ltxt_markdown_drop_classes, ltxt_apply_content_filters, ltxt_content_type, ltxt_robots_reference
• Search API: ltxt_is_indexable, ltxt_search_results, ltxt_client_ip
• Structured data: ltxt_schema, ltxt_schema_conflicts
• Scan my site: ltxt_suggestions, ltxt_contact_slugs
• Chat widget: ltxt_render_widget, ltxt_faq_match, ltxt_faq_stopwords, ltxt_synonyms, ltxt_faq_import_candidates
• AI answers: ltxt_chat_system_prompt, ltxt_ai_client_model_preference

Actions: ltxt_rebuilt, ltxt_index_complete

The settings screen describes each one.

External services

This plugin does not connect to any external service itself. It contacts no AI provider, and it never contacts pitwallwp.com or any other server. It loads nothing from outside your site in your visitors’ browsers.

Only the chat widget’s AI answers can send anything outside your site, and only if you choose them: while the widget is set to “Curated answers, then AI” (and then only for questions no curated answer fits) or “AI for every question”, on WordPress 7.0 or later, with an AI provider set up under Settings, Connectors. With the free ways of answering, nothing is ever sent.

The WordPress AI Client hands each AI answer to the AI provider you set up under Settings, Connectors, on your account with it. Which company receives it, and whose terms and privacy policy apply, is your choice in WordPress.

What is sent, for each AI answer: the visitor’s question; the earlier messages of the same conversation; the passages from your site that the plugin’s own search found for the question; the curated answer that fits it, if any; your site’s llms.txt and name; and the instructions for the assistant, including any extra guidance you wrote. Nothing that identifies the visitor (IP address, cookies, login) is sent.

You pay the provider for these answers on your own account.

Privacy

The plugin suggests wording for your privacy policy under Settings, Privacy, covering the questions log and, when you use AI answers, what is sent to the AI provider.

The chat keeps the conversation in the visitor’s browser tab (sessionStorage), so it follows them from page to page. Closing the tab deletes it; nothing about it is stored on your server beyond the questions log.

To stop the chat and search endpoints being flooded, the plugin counts requests per visitor address for a minute at a time. It stores a one-way, salted hash of the address, never the address itself, and deletes it within minutes.