Dragon Internal Links

Your AI key now stays with its provider and password-protected posts are never sent to AI, and orphan counts update as soon as you apply a link.

What's new in 1.1.13

• Fixed: your AI API key now belongs to the provider you entered it for. If you change provider without entering a new key, the old key is removed and the Settings screen asks for the new provider's key, so it is never sent to another provider.
• Fixed: password-protected posts are left out of suggestions, as the post being linked from and as the page being linked to, so their text is never sent to the AI provider or shown as a suggestion's context.
• Fixed: a page that gains its first link through Apply or an edit leaves the Orphan Posts list straight away, and one that loses its last link joins it, without waiting for the next full scan. A link from a post to itself no longer hides it from the orphan list.
• Fixed: links are no longer added inside code samples (code blocks, inline code, keyboard and sample text), or inside a bare URL or email address such as an embed's URL, which used to stop the embed from showing.
• Fixed: suggestions in languages with multibyte letters were sometimes lost because their context was cut in the middle of a character. Capitalised non-English words, such as "Über" and "über", now count as the same word.
• Fixed: a suggestion that was already applied or dismissed, for example in another browser tab, cannot be applied a second time.
• Fixed: an AI API key is now saved exactly as you type it. A key containing a percent sign followed by two hex digits, an angle bracket or a double space was altered when saved.
• The Orphan Posts headings show the real totals, with a note when only the first posts are listed.
• The plugin's database tables are checked after they are created, so a failed setup is retried on the next admin page instead of breaking every scan.
• When you opt in to deleting data on uninstall, only this plugin's own old settings are removed, not other plugins' settings that happen to share the old "dil_" prefix.
• The settings form checks its security token and your permission in the same step that saves, and reads every field through a WordPress sanitizer. Only valid post type names and category IDs are stored.
• Clearer wording on the Settings screen and in the readme about what AI ranking sends and how often: one request per post every time suggestions are generated, including each scheduled full scan, with up to about 1,500 characters of the post. Languages written without spaces between words, such as Chinese, Japanese and Thai, are not matched.

Uninstalling now deletes your data only when the delete-on-uninstall opt-in is clearly switched on.

What's new in 1.1.12

• Fixed: uninstall deletes data only when the opt-in is clearly on (1, true, yes or on). A value set by hand to "false", "no" or "off" used to count as a yes and now keeps your data.

Link Health keeps reporting links to removed posts, and suggestions find more relevant links.

What's new in 1.1.11

• Fixed: links to trashed, deleted or draft posts disappeared from Link Health after the next scan. They are now reported on every scan until fixed.
• Suggestions come from each post's distinctive topic phrases, not only its exact title.
• Links to media files are no longer reported as broken.
• Retired AI models move to current ones with a notice, and AI errors show on the Settings screen.
• Titles with punctuation or non-Latin letters get suggestions, and scan warnings stay on screen.
• Multisite: tables are created and removed per site.

More accurate orphan reports and link suggestions.

What's new in 1.1.10

• Unpublished, trashed and deleted posts leave the link index.
• Dismissed suggestions stay dismissed, and the Exclude Categories and Scan Frequency settings now take effect.
• Suggestions are only offered where they can be applied, and links are never placed inside words, shortcodes or captions.

Translation-ready throughout.

What's new in 1.1.9

• Every screen, email and alert is now translatable, and translations bundled in the plugin's languages folder now load. Counts use proper plural forms, and numbers and dates follow your site's language.
• Priorities, post types and link statuses show readable labels.

Quieter debug logs: scheduled tasks now wait until WordPress has finished loading.

• Fixed: when one of Dragon Internal Links's scheduled tasks needed re-creating, it was scheduled before WordPress had finished loading. WordPress then logged "translation loading was triggered too early" notices that named other plugins, which could send you looking in the wrong place. Scheduling now waits until WordPress is ready.

The notices only appeared with debug logging switched on. Nothing else changes: no settings, no stored data, nothing to do after updating.

Applying a suggestion no longer breaks image or block markup, save failures are reported, and subdirectory installs now detect internal links correctly.

What's fixed in 1.1.7

• Links go into visible text only. Applying a suggestion could insert the link inside an image's alt text, a caption, a block's settings or an existing link, breaking the tag or the block. Only visible text is linked now, and the anchor keeps the text exactly as written in the post (capitalisation preserved).
• A failed save is reported as a failure. If WordPress refuses to write the updated post, the suggestion is no longer marked as applied. The save also no longer strips embeds or other HTML from the rest of the post for users without the unfiltered_html capability.
• Subdirectory installs detect internal links. On a site at https://example.com/blog, links written as /blog/page/ were not recognised as internal, so every post looked orphaned. Relative links (page/, ../page/, ?page_id=12) and links whose host differs only by letter case or a www. prefix are now recognised too.
• Database write failures are reported. Generating, applying or dismissing suggestions no longer counts a failed write as done.
• Long or unusual titles. Very long post titles are shortened at a word boundary when used as a keyword, titles with unusual characters no longer produce empty suggestions, and the Context column on the Suggestions tab no longer goes blank for them.
• The Context column no longer goes blank for accented or damaged text. An excerpt containing a byte that is not valid UTF-8 made the whole column empty. The offending bytes are dropped and the rest of the excerpt is shown, still with the keyword highlighted.
• Sites whose address is stored with capital letters find their own links. If the WordPress Address is recorded as, for example, https://Example.test, internal links were compared in lower case and WordPress then refused to match them, so every post looked like an orphan. The site's own spelling is now used for the lookup.
• A rescan cannot wipe a post's links. The existing links were deleted before the new ones were written, so anything that went wrong in between left the post with no links at all while the scan reported the links it had found. The replacement now happens in one step that either completes or leaves the previous links untouched, and the screen reports any post that could not be indexed.
• Dismissing a suggestion that no longer exists is reported honestly. A suggestion deleted in the meantime was reported as dismissed. The stored status is now checked.
• Suggestions that could not be saved are reported. A generation pass that could store nothing reported only how many posts it had analysed, which read as "no suggestions found". The number that could not be saved is now shown, as is a pass that could not clear the previous one first.

No settings or data changes. Run a scan after updating so subdirectory sites pick up the newly recognised links.

Readme correction: the AI re-ranking disclosure now states that your API key is sent in a request header, never in the URL. No functional change.

= 1.1.6 =

• Readme: corrected the AI re-ranking disclosure (the Google API key is sent in a request header, not in the URL) and added the missing 1.1.5 changelog entry. No functional change.

Full coverage & security: suggestion generation now covers every post (not just 30) without losing prior results, and applying a suggestion requires edit rights.

= 1.1.5 =

• Improvement: generating suggestions now pages through all posts and no longer wipes previous results mid-run.
• Security: applying a suggestion now requires edit rights on the post; the AI model setting is constrained to the selected provider.
• Hardening: the BYO AI key uses authenticated encryption at rest.

Tested and verified on WordPress 7.1.

• Compatibility: tested up to WordPress 7.1.
• Fix: the first-run guidance panel now shows its intended styling.

Fixes a misleading message where a never-scanned site was told all posts have links, makes scheduled scans survive large sites by continuing across cron runs, and documents the optional AI re-ranking in full.

Your data now survives uninstall by default.

Uninstalling no longer deletes the plugin's data unless you explicitly opt in first, so removing it to debug — or coming back later — keeps everything intact. Where the plugin has a settings screen you'll find a new "Delete all data on uninstall" option; leave it off to keep your data.

A cleaner, consistent look: the Dragon design system arrives.

This release brings Dragon Core's design system to every screen: a consistent header with the Dragon mark, cleaner tables, and unified status colours across the whole plugin family. Purely visual — everything behaves exactly as before.

Smarter link suggestions: ranked by real document similarity, with optional AI re-ranking using your own API key.

Suggestions are now scored by TF-IDF document similarity instead of keyword overlap alone, so the genuinely related page ranks first — not just the one whose title happens to appear in your text. Optionally, add your own OpenAI, Anthropic, or Google API key under Settings and suggestions are re-ranked for editorial relevance: one small request per post when generating suggestions, your key stored encrypted, and no account with us required. Without a key, the built-in similarity scoring applies automatically.

• Initial release
• Orphan detection
• Link suggestions