Content Health for Confluence
Confluence became your AI assistant's source of truth. When two pages make incompatible claims, it reads both and answers with exactly the same confidence either way — and nothing in Confluence marks either page as a problem.
What it finds
Two runbooks disagreeing on how many approvals a deploy needs. A retention policy that says 30 days in one place and 90 in another. When two pages make incompatible claims, one of them is wrong right now — not "probably outdated", actually wrong.
Dates, keywords and titles cannot find this, because contradictory pages usually share all three: both current, both linked, both edited recently. Finding them means comparing what the two pages actually claim, and that comparison is what this app is built around.
The same procedure documented twice in different words — which is exactly why nobody noticed, and why only one of the two ever got updated. Every duplicate is a contradiction waiting for someone to edit one copy.
Also included
Five more checks run across every page in the scan. They are arithmetic — dates, word counts, and the link graph — so no model is involved and they cost nothing to run. That is the whole design: spend AI only on the handful of page pairs a rule cannot judge, and the rest of the scan stays cheap enough to give away.
Not edited in over a year, still being read, still being cited in answers.
Nothing links to them. That used to mean nobody found them. Conversational search ended that.
Placeholders that quietly became permanent.
Pointing at pages that no longer exist.
Nobody's name is on them. When the answer turns out to be wrong, there is no one to ask.
Permissions
Content Health requests read-only permissions. It cannot edit, archive, move, or delete anything — not as a policy we promise to keep, but because the permissions required to do so were never requested. You can verify that on the listing's permissions tab before installing.
It reports what it found and stops. Export the findings as CSV or copy the list of page links, then work through them however your team already works. Every fix stays yours to make.
How it works
Age, links, word counts, and the link graph — across every page in the scan, costing nothing.
Text-similarity filtering narrows a thousand pages down to a handful of candidate pairs, and only those pairs are compared for contradictions and duplication.
A nightly pass reuses the previous verdict for any pair whose pages haven't moved. Set it up once, then check the dashboard when you want to.
Early access
Scans your 1,000 most-recently-updated pages, contradiction checks included, and shows you what it finds.
Built on Atlassian Forge. Your content is processed inside Atlassian's own infrastructure and is never sent to a third party's servers. Duplicate and contradiction detection compares short excerpts of pre-filtered candidate page pairs using an Atlassian-hosted Claude model on Atlassian's first-party Forge LLMs API — not a model vendor's own API. Nothing is retained by the model provider, and nothing is used to train models.