Last updated: 21 June 2026 · Accurate to PageKeeper v10.20.0.
Plain-English summary
PageKeeper helps a Confluence admin find pages that have gone stale, duplicate or contradict each other, or are no longer linked — and archive the dead weight. It reads page metadata and (for the AI features) page text, acts only with your own Confluence permissions, sends none of your content to any service we operate or any third party, and stores no page content — only small settings and numeric/derived values needed for the features to work.
What the app accesses
- Page list + metadata (title, last-modified date, owner, web link, space) for pages your Confluence user can see.
- Page body text — read to produce the AI archive verdict, the content-staleness check, and the Conflict Finder comparison. Read as you; the body text itself is never stored.
- View counts — read from Confluence analytics to weigh the verdict, only after your own read of that page succeeds.
- Labels — to honor your "do-not-archive" label.
- Archive status — archiving changes a page's status to "archived" via the Confluence API. Reversible; nothing is deleted.
How the AI works
The AI verdict, content-staleness flag, and Conflict Finder send the relevant page text to
Atlassian's Forge-native AI (@forge/llm, running on Atlassian's infrastructure)
and return a short result to your screen. Processing happens inside Atlassian's own platform with no
data egress, under Atlassian's data-handling terms. PageKeeper operates no AI endpoint of
its own and sends your content to no third party. The page text is not retained after the result is produced.
What the app stores (in Atlassian's Forge storage — never page content)
- Weekly-digest settings you configure (space, threshold, recipient) and run markers (a date) so the digest doesn't double-post.
- A short-lived "safe to archive" list of page IDs (~15 minutes) so you can only archive what a complete scan actually flagged.
- A conflict-verdict cache keyed by page IDs + version numbers (the short verdict + evidence phrase, not page bodies), bounded in size, so re-checks of unchanged pages are instant.
- Per-user Wiki-Health snapshots — dated numeric counts and the 0–100 score. No titles or content — just numbers.
No database of your pages, no copies of page bodies, and no summaries of content are stored.
Permissions the app requests, and why
read:page,read:space,read:confluence-content.summary,read:analytics.content— list pages, read bodies for the AI features, read view counts.read:confluence-user— resolve page owners for display.write:page,write:confluence-content— archive (and so restore) pages at your request.write:comment— post the weekly digest's @mention comment.search:confluence— find pages by label and group candidates.storage:app— store the small operational data above.
The app runs as you, so it can never read, compare, or archive a page you couldn't access yourself.
Data residency & retention
- Residency: all processing happens within Atlassian's Forge platform and Atlassian's AI infrastructure. No external servers; eligible for Runs on Atlassian.
- Retention: no page content is retained. Settings, the short-lived safe-list (~15 min), the bounded verdict cache, and per-user health snapshots persist in Forge storage until they expire, are pruned, are overwritten, or you uninstall.
Your control
- Archiving requires you to select and confirm; bulk archiving shows a confirmation first and verifies each page with a CSV receipt.
- A "do-not-archive" label protects pages, enforced on the server.
- Every archived page is restorable from Confluence at any time.
- Uninstalling removes the app's access immediately and its stored operational data with it.
Contact
Questions about data handling: [set support email before publishing].