Automated lifecycle recommendations
Keeping a large documentation set healthy means constantly asking the same questions: What has gone stale? What needs re-verifying? Are two topics saying the same thing? The lifecycle agent asks these questions for you every night and turns the answers into a short list of concrete, actionable recommendations. In the app it is called the health scan, and its results are labeled Health scan recommendations.
The agent recommends, it never edits. Every suggestion is something a human accepts, acts on, or dismisses. Topicary does not change your content automatically.
What the agent scans for
Each nightly run checks every topic against seven signals:
Signal | What it looks for | Suggested action |
|---|---|---|
Staleness | Topics not updated within your staleness threshold, adjusted for the topic's type | Update |
Verification expiry | Published topics never verified, or verified longer ago than your verification cycle | Verify |
Near-duplicates | Pairs of topics whose content is more similar than your duplicate threshold | Merge |
Low AI-readiness score | Published pages whose GEO (AI search readiness) score falls below your minimum | Update, Expand, or Split, depending on the weakest factor |
Orphans | Topics not in any map that are also stale (30+ days) | Archive |
Ambiguous topics | Topics in a gray zone (somewhat stale) where AI judges that action would help | Update, Verify, Split, or Expand |
Reader feedback | Published pages where 2 or more readers answered "was this helpful?" with no in the last 30 days | Update |
Each recommendation carries the evidence behind it (the number of days, the similarity percentage, the weakest AI-readiness factor, the number of readers) so you can judge it before acting.
Two of these signals behave in ways worth knowing:
Staleness scales with topic type. Reference, task, troubleshooting, tutorial, and how-to topics use your threshold as set. FAQ topics get 1.25 times longer, concept and overview topics twice as long, and untyped topics 1.5 times longer, because an evergreen explainer does not go out of date on the same clock as a reference page.
Low AI-readiness picks the action from the weakest factor. Most weak factors (schema potential, structural clarity, internal linking, answer density, concrete evidence) produce Update. Weak semantic completeness produces Expand, and a page large enough that AI agents truncate it mid-read produces Split.
Where recommendations appear
Scan results appear in two places, and they do different jobs:
Analytics ▸ Content health, in the Health scan recommendations section: a summary of the latest scan. It tells you how many topics were flagged and what kinds of work they need, as a count next to each action type. This is where you observe; there are no per-item controls here.
Demand, your worklist: the recommendations themselves, ranked against every other signal Topicary collects, including reader gaps, failing doc tests, and watched code changes. This is where you act on them.
Select Review in Demand in the summary to go straight to the worklist.
Act on a recommendation
Open Demand and work the list. Each item names the topic, the reason it was flagged, and the action Topicary recommends, and each carries the controls to act on it or clear it.
Because the scan never edits, acting on a recommendation is an invitation to do the work, not an automatic fix. See Prioritize what to write next for how items are ranked and resolved.
Recommendations are advisory. Resolving or dismissing one only updates the recommendation's status; it never publishes, deletes, merges, or rewrites content on its own. The actual edit is always yours to make.
Run a scan on demand
You do not have to wait for the nightly run. In the Health scan recommendations section, select the refresh icon (Scan now) to queue a scan immediately. The scan runs in the background and new recommendations appear when it completes.
If the Content health tab is empty, select Run analysis now. That starts a health scan and a structure analysis together.
When a scan is incomplete
A scan runs several independent checks. When some of them fail, the section says so above the summary:
Message | What it means |
|---|---|
Last scan completed partially — some checks could not run. | Some checks failed. The recommendations you see are real, but signals from the failed checks are missing |
Last scan failed — checks could not run. Results may be incomplete. | Every check failed. Treat the current list as out of date |
Both messages describe the most recent completed scan, so a scan still in progress never shows a premature verdict. A degraded scan keeps the section visible even when it produced no recommendations at all, so a scan that silently found nothing is never mistaken for a clean bill of health.
Run a scan again to clear the message. If it persists, see Troubleshooting.
Tune the scan thresholds
The scan's sensitivity is controlled per project by the Content health scans settings, in Settings ▸ Content & AI. Only people can change these rules; the scan reads them but never writes them.
Setting | What it controls | Default |
|---|---|---|
Enable automated scans | Turns nightly scans on or off for the project | Enabled |
Staleness threshold (days) | How old a topic must be before it is flagged for update | 90 |
Verification cycle (days) | How often published topics should be re-verified | 90 |
Minimum AI-readiness score | Pages whose GEO score (from Analytics, AI search readiness) falls below this (0 to 10) are flagged | 5.0 |
Duplicate detection sensitivity | How much two topics must overlap to be flagged as duplicates. Higher is stricter and produces fewer flags | 0.92 |
Adjust a value, then click Save settings. The new thresholds take effect on the next scan.
If the scan produces too many duplicate flags, raise the duplicate detection sensitivity toward 1.0 so that only very close matches will be flagged. If staleness flags feel premature for content that genuinely changes slowly, raise the staleness threshold to match your real update cadence.
Turning off automated scans stops new recommendations from being generated but does not remove ones already created. Re-enabling resumes scanning on the next nightly run.
See also
Prioritize what to write next: The worklist where you act on every recommendation the scan produces
Content debt score: The 0-10 score built from the same staleness, findings, and verification signals
Review structure suggestions: The other AI proposals on the Content health tab, which you accept or dismiss in place
Track content freshness: Staleness indicators and human verification that the scan acts on
Content health and governance: How automatic health tracking surfaces decay before readers notice