How old are the pages AI engines cite?
Freshness is the most repeated advice in AI search and the least measured. We read the URL and the headline of every page cited in 12,408 ChatGPT and Perplexity answers, 9,015 distinct URLs on ChatGPT and 4,709 on Perplexity's first 10 sources, and recorded the year each one claimed. Most claimed nothing.
About 18% of citations carry a date in the URL or headline. Of those, the current year dominates: 62.6% on ChatGPT and 81.1% on Perplexity. ChatGPT cites more old pages, and nearly all of that gap is PDFs and uploaded files, which it cites 3.6 times as often. Take files out and the two engines cite earlier-year pages at almost the same low rate.
Four in five cited pages carry no date in the URL or title
| Share of citation rows | ChatGPT | Perplexity, first 10 |
|---|---|---|
| A date in the URL | 5.5% [4.7, 6.3] | 2.7% [2.1, 3.3] |
| A year in the headline only | 12.0% [10.3, 14.1] | 16.0% [13.8, 18.3] |
| Any date | 17.5% [15.5, 19.8] | 18.6% [16.3, 20.8] |
| Dated the current year | 11.0% [9.2, 13.1] | 15.1% [13.0, 17.3] |
| Dated an earlier year | 6.6% [6.0, 7.2] | 3.4% [2.8, 4.1] |
| Dated 2 or more years back | 4.6% [4.1, 5.1] | 1.6% [1.2, 2.0] |
Prompt-clustered 95% intervals (2,000 draws). Any date is a date in the URL or a year in the headline. On Perplexity's full September lists of 20 sources the any-date share is 19.3% [17.1, 21.5].
Counted per distinct page instead of per citation row, 22.2% of ChatGPT's pages and 18.5% of Perplexity's carry a date; a date in the URL, 8.4% and 3.6%. The paired per-page difference between the engines is +2.0 points [−0.2, +4.0], so neither engine prefers dated pages. Quote this as no date in the URL or title: a page can carry a date in its body that we did not read.
Where there is a date, it is usually this year
Table view: age of dated citations
| Dated rows | ChatGPT, % | Perplexity first 10, % |
|---|---|---|
| Current year | 62.6% | 81.1% |
| 1 year back | 11.3% | 9.8% |
| 2 years back | 8.6% | 2.5% |
| 3 to 5 years back | 8.5% | 3.0% |
| 6 to 10 years back | 6.2% | 2.4% |
| 11 or more years back | 2.7% | 0.6% |
| A future year | 0.0% | 0.5% |
| Dated rows, count | 4,829 | 11,559 |
The two kinds of date behave differently. Headline years are the current year 75.0% of the time on ChatGPT and 87.7% on Perplexity. URL dates are the current year only 35.4% and 41.8% of the time, and their median is one year back on both engines. A year in a headline is mostly a label; a year in a URL is mostly a publication date.
About three quarters of the current-year headline labels (73% to 78%, depending on the engine) sit on pages whose title also carries a best, top, guide or list style word; the same word test matches 16.6% of undated pages. That is a keyword test, not a hand check, so it is directional, but it points the same way as the list-page finding in Where the sources come from: a 2026 in a cited title is usually a freshness label on a list page.
The old pages are files
ChatGPT's citations are 3.6 times as likely to be PDFs or other files as Perplexity's (9.9% against 2.7% of citations), and 75.4% of ChatGPT's earlier-year citations are files or uploaded documents, against 30.3% of Perplexity's. Among files, 49.5% of ChatGPT's and 37.7% of Perplexity's carry an earlier year; among ordinary pages, 1.8% and 2.5%. Files are also dated far more often than pages (60.3% and 46.4% against 12.8% and 17.9%), which is most of why ChatGPT's dated share looks like Perplexity's while its old share does not.
Paired on the 453 prompt-months (297 prompts) both engines answered with citations, ChatGPT's earlier-year share is +3.0 points higher [2.2, 3.9] (6.6% against 3.5%). Without files it is −0.7 points [−1.3, −0.1] (1.8% against 2.5%): not higher. Per distinct page the gap is +5.5 [4.6, 6.4] with files and +0.3 [−0.3, +0.9] without. Prompt-month by prompt-month, without files, ChatGPT's earlier-year share is higher in 30%, lower in 20% and tied in 50%, and the no-file gap is above zero in only 2 of 7 industries. A file is a URL whose path ends in a document extension or sits in an upload or media path; a hand check confirmed the class on 79 of 80 sampled URLs. The year on a file is the year of the document, not of a web page, so this is a finding about what each engine retrieves, not about how fresh its sources are.
Pages with a month in the URL
| URL dates with a month | ChatGPT | Perplexity, first 10 | Pooled |
|---|---|---|---|
| Rows with a month in the URL | 1,082 | 1,364 | 2,446 |
| Median age, all of them | 20 months | 13 months | 15 months |
| Rows that are ordinary pages | 246 | 843 | 1,089 |
| Median age, ordinary pages | 4 months | 8 months | 6 months |
| Rows that are files or uploads | 836 | 521 | 1,357 |
| Median age, files or uploads | 26 months | 29 months | 27 months |
Age is counted from the fill month. The share of first-10 citations with a month in the URL is 3.9% on ChatGPT and 2.2% on Perplexity: a small slice, about 2% to 4% of citations, so quote these medians only as ages of cited pages whose URL carries a month.
The earlier draft of this study reported one pooled median of about 14 months. That number was a mixture of two populations: ordinary pages about half a year old and files about two years old. Directional, because it rests on the 2% to 4% of citations whose URL carries a month.
Host type does not explain the gap
Standardizing each engine's current-year share to the other's host-type mix (a direct standardization over 16 types holding 99.1% of rows) narrows the raw gap of −4.7 points [−6.5, −2.9] to −1.6 [−3.6, +0.4] on Perplexity's mix and −1.4 [−2.9, +0.2] on ChatGPT's. Host type moves the current-year gap but does not remove it, and the types leave 2,298 hosts unclassified, so this is directional.
Table view: dated share by host type
| Host type | Rows, ChatGPT / Perplexity | Dated, % | Date in the URL, % | Labelled the current year, % | Labelled an earlier year, % |
|---|---|---|---|---|---|
| Review and affiliate | 391 / 2,154 | 57.5 / 61.3 | 3.6 / 7.3 | 57.0 / 57.6 | 0.5 / 3.7 |
| News and trade | 364 / 2,439 | 51.4 / 32.8 | 23.6 / 7.7 | 31.6 / 23.5 | 19.8 / 8.9 |
| Unclassified | 3,086 / 6,147 | 27.8 / 22.4 | 8.7 / 5.8 | 16.8 / 14.2 | 11.0 / 8.0 |
| Directory and review | 971 / 3,171 | 21.9 / 32.0 | 0.5 / 0.0 | 20.2 / 27.3 | 1.8 / 4.7 |
| Data and research | 1,430 / 2,029 | 19.6 / 10.1 | 2.8 / 1.3 | 6.3 / 7.2 | 13.3 / 2.8 |
| Company site | 12,211 / 27,437 | 16.0 / 17.0 | 4.5 / 1.6 | 10.6 / 14.7 | 5.4 / 2.2 |
| Government | 3,530 / 1,535 | 15.3 / 17.9 | 11.8 / 13.3 | 6.0 / 3.8 | 9.3 / 14.1 |
| Information portal | 911 / 6,455 | 13.9 / 8.6 | 6.7 / 1.2 | 5.9 / 8.2 | 8.0 / 0.4 |
| Medical and academic | 1,326 / 1,333 | 7.5 / 11.2 | 2.1 / 5.2 | 2.1 / 2.8 | 5.4 / 8.4 |
| Reference (Wikipedia and similar) | 1,054 / 75 | 0.0 / 0.0 | 0.0 / 0.0 | 0.0 / 0.0 | 0.0 / 0.0 |
Tracked and partner hosts excluded. Each cell is ChatGPT / Perplexity first 10. Reference sites are the Wikipedia pattern: no dates at all.
When ChatGPT searches with an old year, it cites older pages
ChatGPT's own web searches sometimes carry an earlier year, as in a search for the best firms of 2025 run in 2026: 450 queries in 332 answers. Those sit in 142 prompt-months (114 prompts) where other answers to the same prompt in the same month searched without the stale year, which makes a within prompt-month comparison possible.
| Share of an answer's citations | Stale-search answers | Other answers, same prompt-month | Mean within difference |
|---|---|---|---|
| Labelled with a year that answer's own searches used | 6.2% | 2.4% | +3.8 [2.1, 5.8] |
| Labelled with a year the prompt-month's searches used | 6.7% | 2.4% | +4.3 [2.6, 6.3] |
| Labelled with any earlier year | 11.1% | 7.1% | +4.0 [1.9, 6.2] |
| 2 or more years earlier | 5.5% | 4.9% | +0.6 [−0.6, 1.8] |
| Labelled with the current year | 19.1% | 22.5% | −3.3 [−5.7, −0.9] |
| Dated at all | 30.2% | 29.5% | +0.7 [−1.8, 3.3] |
Within prompt-month comparison: 319 answers whose own searches used an earlier year against 1,104 answers to the same prompt in the same month whose searches did not. Intervals are prompt-clustered.
With files removed the gap on the first row is +3.4 points [1.9, 5.3] (4.2% against 0.8%), and a permutation placebo that shuffles the stale-search label within prompt-month reached at most +1.9 in 500 shuffles. On the per-answer version the ratio is about 2.6. By month: +4.4 [1.8, 7.0] in September (n 88) and +3.4 [0.5, 6.8] in October (n 54). This is an association that survives a no-file cut and a placebo, not a rate and not a mechanism: the stale-search rate itself moved from 8.2% to 5.0% of answers between the two months.
How the dates were read, and how the reading was checked
A date is a four-digit year in the URL path, with a month where the path carries one, or a four-digit year in the headline the engine returned. Headlines were present on 99.8% of citation rows. The extractor was hand-checked twice on fresh random draws: 111 of 113 dated rows and 60 of 60 undated rows agreed with the hand reading. An independent sample of 70 dated and 60 undated URLs, judged on the year the page claims, disagreed on 3.3% (date codes in forms the extractor does not read).
One instrument note. Perplexity's API returned 20 sources in September and 10 in October, and its first-10 URL-dated share moved from 3.6% to 1.6% across all sites, and from 2.3% to 1.5% on the 10 sites tracked in both months, while ChatGPT moved from 3.9% to 3.4%. Most of Perplexity's drop is one panel leaving the sample, so no month-to-month freshness claim is made here.
What this does not show
- It reads URLs and headlines, not page bodies. A page dated only in its body counts as undated.
- It cannot say whether a dated page was chosen because of its date. Neither engine prefers dated pages on the per-page comparison above.
- The age of AI Overview sources was measured and withheld: the result was null and the sample too small to publish.
- Host types come from the source-mix study and leave 2,298 hosts unclassified.
- Two fills only. The sample, the instrument and the statistics are on the method page; the house rule for every figure is on belief against measurement.
If a page of yours was cited with the wrong year, or you want a date pattern added to the reader before the November refresh, say so.
Quoting a figure? Link to this page as the source: https://www.bestseopodcast.com/research/ai-cited-page-age. Figures are dated 2026-10-25 and cover answers collected 12 to 26 September 2026 and 1 to 5 October 2026; the next refresh is after the November 2026 fill closes, about 15 November. The page is the citation: there is no download.
Who did this
Matthew Bertram runs EWR Digital, the SEO and AI-visibility agency whose monthly tracker produced this corpus, and he is the author of LLM Visibility: A Decision-Grade System for Winning AI-Mediated Discovery. The studies were built, re-derived from counts and reviewed before anything was written up; every figure on this page traces to a script in the project record. Questions about the method go to him through the contact page. More about Matthew Bertram.
Are your pages the ones being cited?
The audit reads which of your URLs ChatGPT and Perplexity return, and what they return instead.
Get a free AI visibility audit