ChatGPT Citations Have a Half-Life: How to Schedule an Agent to Refresh Pages
Promptwatch tracked 90 days of ChatGPT citations by page type. Here is what the decay data suggests about refresh schedules, and how an agent can run them.
Product pages that ChatGPT cited on July 6, 2026 were 55% still being cited 60 days later. News pages were at 41% by day 60, and almost a quarter of them were never cited again after day 0. Those figures come from Promptwatch's ChatGPT content half-life report, which followed every classified page ChatGPT cited on that one day for 90 days, through October 4.
For anyone running an agent against a content library, the useful idea is that citations decay, and they decay at different speeds by page type. That makes refresh work schedulable.
The numbers
After 30 days, ChatGPT still cited 78% of landing pages and 76% of docs from day 0. After 60 days the figures were 69% and 67%. Comparisons held 72% at 30 days and 63% at 60. Listicles had 60% still cited at day 60.
Product pages, reviews and how-to guides sat together: about 55% still cited after 60 days, with all three reaching half-life after 74 to 77 days. Pricing pages are classified as landing or product pages in the report, so they are not a separate line.
News is the fast decayer. Half of news pages were gone after 36 days, and docs were cited again later 93% of the time versus 77% for news.
One caution on the long-lived types. Landing pages, docs, comparisons and listicles only cross the 50% line in the final week because the window ends there, so the report says their half-life is likely close to or beyond 90 days. Nobody measured it, so don't write 90 days into a runbook as fact.
What it does and does not tell you
It is one cohort, drawn from one day, in one engine, from one vendor's prompt set. It describes how long pages stayed cited, not why. A page that stops being cited may have gone stale, may have been outranked, or may simply have been swapped for a newer source. The data cannot separate those, so the report's advice is a starting point for a schedule, not a rule.
The report itself suggests treating landing, docs and pricing pages as long-term assets, not assuming product pages stay cited on their own (keep specs, availability and prices current), investing in comparisons and listicles because both keep more than 60% after two months, and treating news as short bursts.
A refresh schedule you can run without software
Sort your library into three bins using the report's shape.
- Bursts (news, announcements). Don't plan to refresh. Publish, promote, and accept the decay. If a news item points to a lasting page, link to it.
- Fragile mid-lifers (product pages, reviews, how-tos). Review on a roughly 60-day cycle, since that is where the report shows about 45% have already dropped out. Check prices, specs and availability first.
- Slow decayers (landing, docs, comparisons, listicles). Review quarterly and after any product change. Spend the effort on comparisons and listicles, where the report shows meaningful retention that you can influence.
Write the date of the last check on each page's internal record. A spreadsheet with URL, type, last reviewed and next due is enough to start.
Turning the schedule into an agent loop
The agentic angle is that the schedule runs itself and only asks for a human at the edit.
Monitor: per page, an agent reads when it was first cited, how fast citations ramped up, when they peaked and how they are decaying. Schedule: it runs on a fixed cadence and ranks pages by "was cited, now sliding." Fix: it drafts the update for review rather than pushing it live.
How Promptwatch maps to each step
Promptwatch has the pieces by name. Citation Trends classifies every citation by content type and source type and gives a citation rate over time, plus first cited, ramp-up days, peak rate and decay half-life per page. That is your own version of the report's chart, at page level. Page Tracker lets you follow specific URLs, yours or a competitor's, which suits the pages on your 60-day bin.
For the fix, Content Refresh produces diff-style update suggestions for existing pages, so the reviewer sees what would change rather than a rewritten page. Content Agent (Business and Enterprise) can also write against gaps on a schedule, with a review inbox or a hands-off mode with a quality check. Be careful with hands-off mode on pages that carry prices and specs. A human should see those.
For the cadence itself, Scheduled Chats run an Agent Chat prompt on a schedule and deliver the result in the app or in Slack. A prompt like "list pages whose citation rate has fallen since last month, grouped by type" fits. Findings feed the Action Board, where the Discovery agent suggests up to five evidence-backed actions per weekly run and you can assign each one.
The plan tiers for Citation Trends, Content Refresh and Scheduled Chats are not stated in the sources we work from, so check the plan matrix. Content Agent starts at Business ($579/mo). Essential is $95/mo and Professional is $245/mo.
Where this leaves you
Treat refresh as a calendar, not a reaction. Start with the 60-day bin, since that is where the report shows the steepest early loss for pages you can actually edit. Re-measure each page after the edit, and keep the before and after dates, because that is the only way to learn whether your own library behaves like the report's.
For adjacent loops, read Citation Trends and decay half-life: when to refresh a page, diff-style updates for cited pages, and how to set up Scheduled Chats.