The app is in early access — small, free, and read-only. There is no dedicated support team yet, but every question gets a real answer from the person who built it.
Get help
Bug reports, feature requests, and general questions all go to the same inbox. There is no ticket queue to route around and no form to fill in — you are writing to the person who built the app.
[email protected]FAQ
Eight kinds of finding, and two of them are the reason the app exists. Contradictions are two pages making claims that cannot both be true — one runbook requiring two deploy approvals and another requiring one, a retention policy given as 30 days in one place and 90 in another. Duplicates are the same procedure documented twice in different words, which is why only one of the two ever gets updated. Neither can be found by comparing dates, titles or keywords, because pages like that usually match on all three.
A third finding needs no model either, and is one of the more useful: pages that state their own expiry date and are now past it. A page saying "valid until Q3 2025" or "during the migration, which ends March 2026" announces its own expiry date in prose. Once that date has passed the page is out of date no matter how recent the last trivial edit looks — and a last-modified timestamp will never tell you so.
The remaining five — stale, orphaned, empty, broken-link and ownerless pages — are arithmetic over dates, word counts and the link graph. That split is the design: rules catch what rules can catch, and AI is reserved for the handful of page pairs that actually need judgment.
No. Content Coherence requests read-only permissions and nothing else — it cannot edit, archive, move, or delete a page, not as a policy we promise to keep, but because the permissions to do so were never requested. You can verify this yourself on the listing's permissions tab before installing. It reports what it found and stops there; every fix stays yours to make.
Every scan is bounded, and the 1,000 pages selected are your most recently updated ones — the pages people are actually reading and editing right now, not a random or alphabetical slice of your site. If your site is larger than 1,000 pages, the oldest, least-recently-touched pages are the ones left out, and the report tells you exactly how many that was.
Two different answers, and we would rather state both on day one than surprise you later. The six structural checks — stale, orphaned, empty, broken-link, ownerless and expired pages — are unlimited and included at no cost, on a site of any size. They find what they find without calling a model, so there is nothing to meter: they do not draw on the AI budget and they do not stop when it runs out.
The duplicate and contradiction checks are different. They are free during early access, and when a paid plan exists we expect them to be part of it. No paid plan exists today, and we are not naming dates or prices for one. If that changes, we will announce it here and in the Marketplace listing before it takes effect, not after — and anything you have already been shown stays yours.
Yes — they are on, and they are part of the free report. Like any AI-backed feature, they run behind safeguards: an emergency off switch if something goes wrong, and a monthly usage ceiling.
If either one stops a scan from comparing some pairs, the report says so directly on the scan summary — it will tell you how many similar pairs went uncompared and why. It will never quietly show you a clean result instead. In that situation the six structural checks are unaffected and run exactly as normal.
A lightweight, structural pass narrows your whole site down to a small number of page pairs that look similar. Only short excerpts from those pre-filtered pairs — not your whole site — are then compared, by an Atlassian-hosted Claude model running on Atlassian's own first-party Forge LLMs API — not on a model vendor's own API. The comparison never leaves Atlassian's own infrastructure — no third-party server ever receives it — the model provider retains nothing, and neither the excerpts nor the answers are ever used to train a model. For the full policy, see our privacy policy.
Content Coherence runs on a scheduled interval configured by the app, plus you can trigger a manual rescan at any time from the dashboard. A pair of pages that hasn't been edited since the previous scan is never re-compared by the model — the previous verdict is reused — so repeat scans only recheck what actually changed.
No. Install it, open the Content Coherence page from your Confluence sidebar, and the first scan starts. There's nothing to set up and no credentials to enter — it's free while the app is in early access.
Permissions, restated
This is the single most important thing to know before you install: Content Coherence has no write scopes at all. It cannot edit, archive, move, or delete a page. Everything it does is read, analyze, and report.
Known limitations
No real page-view counts. The page cap ranks by most-recently-updated, not by how many people actually view a page, because real view counts require an API scope this app does not request.
Broken-link findings can't always tell why a link is broken. A link that points at a genuinely deleted page and a link that points at a page just outside the 1,000-page cap both surface as "broken link." On a large site scanned under the free-tier cap, a small number of these may be false positives.
No external error monitoring yet. If a scan fails, that failure is recorded in your report and shown on the admin page — but nothing proactively alerts us. If a scan looks stuck or wrong, please email us.
Findings can be exported, but not managed. You can export what's on screen as CSV or copy the list of page links, and you can filter by kind and severity. There is no way yet to dismiss a finding, assign it to someone, or scope a scan to particular spaces.