How long do AI assistants keep citing an article, and when should you update or replace it?
Citations do not come with a shelf life. Here is how we track decay page by page and decide whether an older article gets updated, merged or retired.
Clients ask us this early, usually after one old post that used to get cited stops showing up. The honest answer disappoints people: there is no fixed lifespan. A page can be cited for a long time on a stable topic and lose its place in weeks on a fast-moving one. We do not quote a number, because we have not seen one that survives contact with a real prompt set.
What we do instead is measure decay per page and set a rule for what happens next.
Treat citations like a time series
A citation is not something you earn once. Assistants that search the web choose sources again for each answer, so every page has a history: when it was first cited, how long it took to reach its normal rate, where it peaked, and how quickly it slid afterwards. We track those four things for the pages that matter to the client's pipeline.
The tool we run this on is Promptwatch. Its Citation Trends view reports citation rate over time and gives first cited, ramp-up days, peak rate and decay half-life for each page. It also classifies each citation by content type and source type, which helps when you find that a client's comparison pages hold up while their opinion posts fade quickly. Without that per-page history, you are guessing from a few screenshots.
The checks we run before touching any copy
Teams rewrite articles when the real problem is access. So the order matters.
1. Is the page still being fetched? In Agent Analytics we look at the crawler logs for the URL. They are IP-verified, so spoofed user agents are filtered out. We look for GPTBot, OAI-SearchBot, ChatGPT-User, ClaudeBot and PerplexityBot, and for errors with failure explanations. If a firewall rule or a robots change started blocking a bot, a content refresh will do nothing. Crawler logs start on the Professional plan, so we flag that early in scoping.
2. Did the prompt move? We compare the tracked prompts against the current query fan-outs. Sometimes the article still answers the old question while the model now asks something slightly different.
3. Who took the slot? With Page Tracker we watch the competing URL. If a rival page took over, we read it and decide what it does better: fresher figures, a clearer opening answer, better sourcing.
4. Are the claims still true? An outdated claim that is still being cited is the worst case, because it keeps repeating a wrong answer. That one jumps the queue.
Our decision rule
We keep it simple enough to explain to a client in a call.
- Update when the page is still cited but its rate is sliding, or when it is cited for claims that have aged. Refresh the facts, rewrite the first paragraph so it answers the question outright, and date the change.
- Merge when two pages split the same prompts. We keep the stronger URL, fold the best material in, and redirect the other.
- Replace when the buyer's question has changed so much that the old page's job is gone. The old URL redirects to the new page.
- Leave alone when the rate is stable. Fresh dates on a healthy page add risk and no value.
- Retire or rethink when a page never ramped up after two honest attempts. Either the topic is wrong or the page is thin.
Updating beats replacing most of the time. An existing URL keeps its links and crawl history, and that is worth protecting.
How we run the refresh itself
Promptwatch has Content Refresh, which produces diff-style update suggestions for an existing page. We use it as a starting draft, never as a final one. A human editor checks every fact, because the suggestions are only as good as the data behind them. Content gap analysis gives coverage scores against the client's tracked prompts, so we can see which sections a refreshed article still lacks.
After the change we use the A/B period comparison in a dashboard to check whether the citation rate moved. We usually wait a couple of weeks before judging. Moving too quickly produces false conclusions in both directions.
The Action Board keeps the work visible. Its Discovery agent investigates the data weekly and suggests up to five evidence-backed actions per run, and we assign them to named people on the client side. A refresh that lives in a spreadsheet nobody opens is the usual way this work dies.
What we tell clients about expectations
A few things we say every time:
- A refresh is a test, not a promise. Tracking tells you what changed. It cannot guarantee a recommendation.
- Time-sensitive topics need a calendar. If a page decays fast because the subject moves fast, put recurring updates on the content calendar rather than treating each decay as a surprise.
- Do not refresh everything. Pick the pages tied to revenue, measure them, and ignore the long tail until the top is healthy.
- Third-party mentions matter too. If other sites stop mentioning the client, the assistant has fewer reasons to prefer the client's page. The offsite mentions report shows where the brand is named without a link.
Where we would start
If you only have an afternoon, do this. List your twenty most valuable articles. Pull the citation history for each. Check the crawler logs for any that have faded. Update the ones where the content is the problem, and fix access for the ones where it is not.
That sequence is cheap, and it stops you from rewriting pages that were never broken. It is also the part of our process that clients find most useful, because it replaces an argument about opinions with a chart.