How old are the pages AI cites?

Server-reported content age is the number of days between a cited page's Last-Modified date, as reported by its own server, and the day the citation appeared in an AI answer. Across 116,379 measurable citations on 56,408 pages, the median age was 3 days. A quarter of citations pointed to pages modified the same day: 30,068 citations, or 25.8%, had an age of zero days.

The distribution has a long tail: the 75th percentile is 16 days, the 90th percentile is 147 days, and the mean is 102.6 days. The typical cited page is very recently modified; a small minority of old pages stretches the average.

75,995 of 116,379 measurable citations, or 65.3%, pointed to pages modified within 7 days of the answer. 93,852, or 80.6%, were within 30 days, and 101,800, or 87.5%, were within 90 days.

Only 7,392 citations, or 6.4%, pointed to pages last modified more than a year before the answer. In this measurable slice, content that goes untouched for a year almost disappears from AI citations.

Bucketing every measurable citation by age shows how front-loaded the distribution is. The first week holds 65.3% of citations. Each later bucket is smaller than the one before it, and every bucket past 90 days combined holds 12.5%.

The old end is thin but real: 2.2% of citations pointed to pages 2 to 5 years old and 1.4% to pages more than 5 years old.

Ahrefs analyzed 17.0 million AI-cited URLs and found they averaged 1,064 days since publication, about 2.9 years, measured from when a page first appeared. Our grain is different: the server-reported last modification. On that grain the median measurable citation is 3 days old.

Both results can be true at once. The pages AI cites are often years old as URLs, and their content is re-touched constantly. The page earns the citation over years; the content behind it is days old. This comparison describes two observed datasets with different methods, not a contradiction.

The median measurable citation on ChatGPT Search pointed to a page modified 5 days before the answer; on Google AI Mode, 2 days. 9.5% of ChatGPT Search citations pointed to pages older than a year versus 5.6% on Google AI Mode, and 72.5% of ChatGPT Search citations were within 30 days versus 82.7%.

Both engines lean heavily on recently modified pages; Google AI Mode leans harder. This is an observed difference in the measurable slice, and it does not identify why the engines differ.

Median content age varies by two orders of magnitude across source categories. Academic sources sit at 311 days, news at 46 days, and government at 45.5 days. At the other end, SaaS pages sit at 3 days and health pages at 1 day.

The pattern follows how each kind of page is maintained. A paper or a news story is written once and left alone; product, documentation, and marketing pages are re-touched continuously. Compare a site with its own category, not with the overall 3-day median.

Split by the industry of the prompt instead of the kind of source, and the freshness pattern is uniform. All 17 industries with at least 1,000 measurable citations have a median content age between 2 and 5 days.

The only visible outlier is on the old tail: 11.7% of measurable citations on Financial Services prompts pointed to pages older than a year, roughly double the rate of most other industries.

Among the 15 domains with the most measurable citations, arxiv.org has the oldest content: a median of 327 days, with 48.6% of its measurable citations older than a year. blog.hubspot.com sits at 133 days and businessinsider.com at 68 days. At the fresh end, techradar.com, en.wikipedia.org, and garagegymreviews.com all have medians of 4 days or less.

The count identifies where measurable content is old or fresh; it does not show that age caused any citation.

Of the 15 most-cited websites in the corpus, 11 had zero measurable citations. youtube.com (383,058 citations), reddit.com (310,792), linkedin.com (71,348), facebook.com, medium.com, instagram.com, forbes.com, amazon.com, nerdwallet.com, and quora.com either return no Last-Modified header on cited pages or block the crawl entirely. google.com was measurable for 0.13% of its 440,653 citations, and nytimes.com for 0.01%.

This is the central limit of the study. The measurable slice under-represents the platforms that dominate AI citations, and over-represents SaaS, blog, and documentation sites that serve real headers.

A serve-time stamp is a Last-Modified value within 2 minutes of the moment our crawler fetched the page; it records when the response was generated, not when the content changed. 67,854 of 213,300 cited pages with a parseable header, or 31.8%, returned a serve-time stamp on their latest observation, and 32,480, or 15.2%, were within 5 seconds of the fetch.

Widen the window and the overlap grows: 139,673 pages, or 65.5%, reported a Last-Modified within one day of the fetch. Some of those are genuinely fresh pages, but the overlap with response-generation time means a single Last-Modified reading is weak evidence on its own. We excluded all serve-time stamps from age measurement.

For 308,821 of 1,183,499 citations on header pages, or 26.1%, the first crawl at or after the citation day found a Last-Modified that postdated the citation. The page had already been modified again in the gap between the answer and the crawl, a median of 1 day for measured citations.

Those citations are unmeasurable, and they are unmeasurable because the page changes so often. Pages that re-modify within days of a citation are unlikely to have sat untouched for long before it, so including them would most plausibly push the age distribution younger; the 3-day median is measured on the more stable pages, not the churning ones.

The oldest end of the leaderboard mixes two kinds of pages. Genuinely old evidence is there: arXiv preprints from 2019 to 2021 still earning citations, and a university boarding-facility guide modified 8.1 years before its citations. But the two oldest entries are visible header artifacts: a creditkarma.com roundup titled for August 2026 whose server reports a 9.4-year-old modification date, and a sourcegraph.com product page reporting 26.6 years, older than the company.

The far tail of any Last-Modified study is contaminated by servers that report stale dates for maintained pages. The 1.4% older-than-5-years share is best read as an upper bound on genuinely abandoned content in the measurable slice.

The exclusions are large and every one is deliberate. Of 8,215,946 citations, 1,183,499, or 14.4%, pointed to pages that returned a parseable Last-Modified header when crawled. From those we excluded serve-time stamps (192,746), pages modified again before the crawl (308,821), pages with no crawl at or after the citation day (542,234), crawl gaps over 45 days (23,119), and pre-1995 or future-dated values (200), leaving 116,379 measured citations, or 1.4% of the corpus.

The headline survived three checks. A month-bucketed reconstruction reproduced the older-than-1-year share exactly (6.4% both ways). Restricting to pages with a single header observation gave a median of 6 days. A strict variant requiring the header to be at least a day older than the crawl gave a median of 4 days and an older-than-1-year share of 7.1%. The measurable slice is not a random sample: it excludes the major platforms and over-represents sites that serve accurate headers, and Last-Modified remains a server claim, not a verified content change.

In the slice of AI citations that can be aged at all, recently modified content dominates: a 3-day median, 80.6% within 30 days, and 6.4% older than a year. Academic, news, and government sources are the exception, and the biggest platforms do not reveal content age at all.

Three actions follow. First, re-touch citation-critical pages on a regular cadence with substantive updates; in this data, pages untouched for a year rarely appear. Second, serve an accurate Last-Modified header; a serve-time stamp tells crawlers and engines nothing, and a stale one makes a maintained page look abandoned. Third, benchmark your category, not the global median: a 46-day-old news story is normal, and a 311-day-old paper is normal, but a year-old product page is not. Rerun the same study next quarter over a window of the same length to track drift.