How often should you refresh content to keep AI citations?
Last updated October 14, 2026. Written by the Third Coast AI team.
Refresh a page when its facts, figures or sources have changed, review every page that earns citations at least quarterly, and leave the date alone when nothing else moved, because the engines reward recency unevenly, with half of Perplexity's citations going to content under 13 weeks old while 29% of ChatGPT's go to pages from 2022 or earlier, and because a dateModified that moves while the text stays still is a signal Google's documentation warns against rather than a signal that helps. A real refresh adds or corrects something a machine could quote; a cosmetic one adds a date.
This page sets out what the evidence says about freshness by engine, what Google asks for in dates and markup, what belongs in a refresh, and how often to re-run the prompts that tell you whether the refresh worked. It is the maintenance half of our GEO content program.
Which engines reward recency?
Perplexity most, ChatGPT when the question is time-sensitive, and Google's AI features in proportion to how fresh the ordinary search results for the query are. Half of Perplexity's citations are to content under 13 weeks old, which makes it the engine where a current page can displace an established one quickly and where a stale page falls out fastest. On ChatGPT the picture splits by query: for freshness-sensitive queries 76% of its top-cited pages are under 30 days old, while across all citations 29% are to pages from 2022 or earlier, so an evergreen buyer question is still answered from established pages and a question with a date in it is answered from new ones. Google's AI Overviews and AI Mode are fed by the normal index, and Google's publication-date guidance says it does not depend on any single date factor because every factor can be prone to issues, which is a polite way of saying that changing the date alone will not move anything.
| Engine | Recency signal | What it means for a refresh schedule |
|---|---|---|
| Perplexity | Half of citations under 13 weeks old | Cited pages reviewed quarterly at minimum; new pages on time-sensitive topics get cited fast |
| ChatGPT (time-sensitive queries) | 76% of top-cited pages under 30 days old | Pricing, availability, regulation and news-adjacent pages kept current monthly |
| ChatGPT (all queries) | 29% of citations to pages from 2022 or earlier | Evergreen pages hold their place; refresh for substance, not for the date |
| Google AI Overviews and AI Mode | No single date factor; visible date and markup should agree | Refresh when the search results for the query show fresh content winning |
| Copilot and Bing | Microsoft's Fabrice Canel recommends adopting IndexNow to tell search engines about changes as they happen | Submit updated URLs through IndexNow so Bing re-reads them promptly |
Does changing the date on a page help?
Only when the content changed with it, and it can hurt when it did not. Google's publication-date guidance asks for a visible date with a clear label such as published or last updated, for the same date in structured data through datePublished and dateModified on an Article or similar type, for the two to agree, and for no future dates or dates of the events described on the page; Google's Article markup page says dateModified is the way to give Google more accurate date information, which only works if it is accurate. Header-only refreshes are detected, and a site whose every page claims to have been modified this week is telling the engines that its dates mean nothing. The rule we apply is that dateModified changes when a human changed something a reader would notice, and the visible date changes with it.
A refresh is not a date. It is the sentence you added because the number changed, the source you swapped because the old one went stale, and the paragraph you cut because it was no longer true. If none of those happened, leave the page alone and let the engine keep trusting its date.
Third Coast AI
What belongs in a real refresh?
- Updated figures with updated sources. Any statistic more than a year old is checked against the original source and replaced if a newer edition exists, with the source named in the same sentence.
- Corrections from the fact-check log. Anything the engines are currently getting wrong about you that this page could settle, stated plainly near the top.
- New decision factors. The quarterly re-run of the factor prompts usually turns up one criterion that was not on the list last time, and the relevant page gets a passage for it.
- A passage for a new hidden search. When the captures show a sub-query the page does not answer, a headed passage is added rather than a sentence squeezed into an existing paragraph.
- Removal of what is no longer true. Discontinued services, old prices, former staff and expired certifications come out, because the engines cross-check and a stale claim is worse than no claim.
- Matching schema. Every value in the markup checked against the visible text after the edit, and dateModified set to the edit date.
How often should you re-run the prompts?
Monthly for the prompts a recent batch of pages was aimed at, quarterly for the full bank, and two to three weeks after any indexed change rather than the day it ships. Google's recrawl page says crawling can take a few days to a few weeks after a request, and the other engines publish no schedule, so a re-run the morning after a refresh measures the old page. The monthly run tells you whether the refresh moved the citation; the quarterly run tells you which pages have started losing citations and are due for review; and the fact-check prompts run on the same quarterly cycle, since an engine that described your pricing correctly in March can be wrong again by June after reading a stale third-party page. Our timeline page covers why the runs have to be repeated to mean anything.
Which pages should be refreshed first?
The ones that are already cited, because a citation is a position to defend and the engines re-read cited pages more often than uncited ones. Then the pages that carry decision-factor proof, since a factor proven with a two-year-old figure is weaker than the same factor proven by a competitor with this year's. Then pricing, service-area and process pages, which are where fact-check errors come from and where a stale page does the most damage. New pages come after all of that, which surprises people who think of content as a volume game, and the structure guide explains why a few well-maintained passages beat a stream of new ones. Glenn Gabe documented a site that published 850,000 AI-generated URLs and received a manual action from Google, and refreshing is the opposite strategy: fewer pages, kept true.
How do you know the refresh worked?
The same way you know anything in this field, by re-running the prompts and reporting the rate. The report we give clients each month lists each refreshed page, the prompts it was aimed at, whether the page is now cited for them, whether the company is named, and whether the fact-check log changed, with the misses alongside the hits. If a refreshed page has not moved after two cycles, the brief was wrong rather than the date, and the fix is to read the pages the engines are citing instead and rewrite to their shape. The audit is where that reading starts for a new client, and the content program keeps it running.
Frequently asked questions
No. Google's publication-date guidance asks that the visible date and the dateModified markup be accurate and agree, and header-only refreshes are detected. Change the date when a reader would notice what changed.
It depends on the engine and the query. Half of Perplexity's citations are to content under 13 weeks old; 76% of ChatGPT's top-cited pages for time-sensitive queries are under 30 days old, but 29% of all its citations go to pages from 2022 or earlier.
Quarterly at minimum, monthly for pricing, availability and anything regulatory. Review means checking every figure, source and claim, not re-saving the page.
Two to three weeks after the page is indexed. Google says crawling can take days to weeks after a request, and a run the next morning measures the old version.