GoldenRetriever AI Chat

No. If you can install a WordPress plugin from the admin, you can install GoldenRetriever AI Chat. The setup is a paste-a-key-and-save flow; the connection test on the Settings tab tells you exactly what to do if anything is off — usually it's a one-line message to send your hosting provider.

The free version uses Claude (Anthropic). Pro adds OpenAI (GPT) and Gemini (Google) as alternative providers. You bring the API key for whichever provider you choose; there is no markup, you pay your provider directly.

For a small business site receiving roughly 30 advisor conversations a day, expect £5–£15 per month with Claude Sonnet 4.6 (the recommended default). Claude Haiku is cheaper for higher volumes; Opus is more expensive but more capable. Anthropic bills you directly per token used.

Yes. The widget inherits your theme's primary colour (via --wp--preset--color--primary) and body font where the theme exposes them, falling back to a neutral indigo + system font stack. It's been verified against Twenty Twenty-Five and the major block themes.

Inside your WordPress database. The plugin doesn't have a server. It doesn't see your content, your visitors, or your conversations. The only outbound traffic for serving chat is the API call to Anthropic that you authorise by configuring an API key. (If you opt in to the optional Freemius connection for the in-admin upgrade path, basic licence/environment data also goes to Freemius — see "External services"; skipping that opt-in sends nothing.) Conversation history isn't stored server-side beyond a per-session browser cookie that lets visitors continue a conversation across page loads.

Yes. The "Try it" panel on the Settings tab lets you run real conversations against your configured model end-to-end, without any visitor seeing the widget. The Knowledge tab also includes a Test Retrieval panel that shows you which posts the advisor would quote for a given query.

The key is encrypted with AES-256-GCM tied to your WordPress salts before it's written to the database, and the option is stored with autoload=no. The admin UI only ever displays a masked form (••••XXXX). The plaintext key is read only at the moment a chat is being served.

The plugin enforces a per-IP rate limit (30 messages/hour by default) and a global rate limit (200 messages/hour by default) before any call to Anthropic. Both limits are filterable. The widget also caps individual messages at 4,096 characters and conversations at 20 messages.

Conversations go directly from your visitor's browser through your WordPress site to Anthropic and back. They are not logged on disk, not analytics-tracked, not sent to any third-party server we operate (we don't have one). Anthropic processes the messages per their published privacy policy.

Yes. GoldenRetriever AI Chat runs each post through the_content before indexing, which executes shortcodes and page-builder block output exactly the same way WordPress renders them for visitors. This means rendered text — including prices, product names, or any text generated by shortcodes — becomes searchable and can be quoted in answers. If you want to exclude a specific post from the index (for example, a page you keep published for internal reasons), use the grchat_index_render_content filter: return false for that post's ID and it will be skipped during indexing.

Some sites store their visible copy in shortcodes or page-builder plugins (Elementor, Divi, Bricks, a "Custom Code" plugin, etc.) that only register their handlers on front-end requests. In that context WordPress leaves the shortcode tags unexpanded, so the index sees [ccj id="tabmenu"] instead of real content. The same issue applies to injected widgets — pricing tables, review carousels, WooCommerce product tabs — that are added by JavaScript or page-builder render engines after WordPress processes the raw post content. GoldenRetriever AI Chat handles this by fetching every public page directly via its own front-end URL (a "loopback self-fetch") during indexing. The fetched page is parsed and chrome-stripped (navigation, header, footer, sidebars, scripts are all removed) so only the main content area reaches the index. The plugin then compares the chrome-stripped fetched text against the admin-context render and indexes whichever version is richer. No configuration is needed — run a full Reindex (Knowledge tab → Reindex all content) and the advisor picks up the full visitor-facing copy. The same applies to your homepage when it is a custom theme template or front-page widget rather than a WordPress page: a URL knowledge source for your homepage is created and refreshed automatically on every full reindex. The source appears in the Knowledge tab (labelled "Homepage (auto)") after the first reindex completes. Performance note: each post requires one HTTP fetch during the reindex. On most sites (batch size 10, 5-second timeout per fetch) a batch of 10 posts takes at most ~50 seconds. The reindex is resumable — if it is slow you can let it run in the background and come back; the progress bar will be right where you left it. The same fetch also runs when you save or update a post in the WordPress editor — expect a brief extra moment before the save completes while the plugin fetches the page's public version to keep the index current. To opt out of the auto homepage source, add this to your theme's functions.php or a custom plugin: add_filter( 'grchat_auto_homepage_source', '__return_false' ); To skip the loopback self-fetch for a specific post: add_filter( 'grchat_index_fetch_rendered', function( $do, $post ) { return 123 === $post->ID ? false : $do; // replace 123 with the post ID to skip }, 10, 2 ); To override the richer-wins threshold for a specific post (e.g. always use the fetched version): add_filter( 'grchat_index_prefer_fetched', function( $prefer, $post, $fetched_len, $rendered_len ) { return 456 === $post->ID ? true : $prefer; // replace 456 with the post ID }, 10, 4 );

Yes — Pro adds OpenAI + Gemini providers, multi-persona advisors, file-upload knowledge (PDF/DOCX/TXT/MD/CSV) with embeddings retrieval, lead routing to your inbox/CRM with full conversation context, and white-labelling. It installs as a separate plugin and works alongside the free version. The free plugin keeps working forever — Pro never holds it hostage.