--- title: "Do AI Engines Cite Fresher Pages? A Cited Page Was Newer Than Another Page on the Same Website in 55% of Pairs, a Median Gap of Three Days" description: "We read the dates on 969 cited pages and paired 774 with another page from the same website. The cited page was newer in 55 percent of publish-date pairs, by a median of three days. The median cited page was 199 days old when last cited, and 31 percent exposed no machine-readable date." canonical: "https://machinerelations.ai/research/do-ai-engines-cite-fresher-pages-2026" --- # Do AI Engines Cite Fresher Pages? A Cited Page Was Newer Than Another Page on the Same Website in 55% of Pairs, a Median Gap of Three Days We read the dates on 969 cited pages and paired 774 with another page from the same website. The cited page was newer in 55 percent of publish-date pairs, by a median of three days. The median cited page was 199 days old when last cited, and 31 percent exposed no machine-readable date. Canonical URL: https://machinerelations.ai/research/do-ai-engines-cite-fresher-pages-2026 Published: 2026-11-11 Research type: Index Analysis Tags: citation-behavior, content-freshness, machine-relations-index ## Current Index data Updated from the Index release through **2026-11-11** (mri_score_v2.1); observation window 2026-05-10–2026-11-11. This card updates with the published dataset. The article below remains a dated analysis; these figures do not retroactively validate its claims. - [reddit.com](https://machinerelations.ai/index/domains/reddit.com): 6.95% citation rate; 2,673 of 38,449 observed runs, tier A. - [youtube.com](https://machinerelations.ai/index/domains/youtube.com): 5.92% citation rate; 2,277 of 38,449 observed runs, tier A. - [linkedin.com](https://machinerelations.ai/index/domains/linkedin.com): 4.51% citation rate; 1,734 of 38,449 observed runs, tier A. - Engines named here and observed in this dataset: ChatGPT, Claude, Gemini, Google AI Mode, Google AI Overviews, Perplexity. Coverage varies by engine; see the [release manifest](https://machinerelations.ai/data/mri-release-manifest.json) for returned runs and availability. Citation presence measures observed source selection, not endorsement or causation. [Explore the Index](https://machinerelations.ai/index) · [Read the latest weekly report](https://machinerelations.ai/research/machine-relations-index-report). ## Source Body AI engines cite pages that are a little newer than their neighbours, not new pages. In the Machine Relations Index answer ledger we drew 1,250 cited pages across five engines, read the publish and update dates in their HTML, and paired 774 of them with another page from the same website that no engine had cited. The cited page had the later publish date in 246 of 449 decided pairs (54.8 percent, 95 percent interval 50.2 to 59.3). The median gap was three days. The cited pages themselves were not young. Of 631 cited pages with a readable publish date, the median was 199 days old on the day of its most recent citation. 25.5 percent were under 90 days old and 20.6 percent were more than two years old. Across all 969 pages we could read, 30.9 percent exposed no machine-readable publish or update date at all. This is a description of what was cited, and a weak association. It does not show that refreshing a page earns a citation, and it does not show that engines prefer recent pages. ## What was counted - **Population.** Cited addresses in the Index's buying-question panel: successful runs from 2026-07-06 to 2026-11-09, five engines (ChatGPT, Claude, Gemini, Google AI Mode, Perplexity) through their APIs, runs that name a brand in the prompt excluded. The query string and fragment were dropped and the address lower-cased, as in [the page-counting study](/research/how-many-different-pages-do-ai-engines-cite-2026), without that study's full rule. This is not the consumer ChatGPT or Claude app, and Google AI Overviews are left out. - **Sample.** PDFs, Markdown and text files and twelve social and open publishing hosts (Reddit, YouTube, LinkedIn, Facebook, X, Quora, Medium, Substack, DEV and others) were removed. 250 pages were then drawn at random, with a fixed seed, from each engine's cited pages: 1,250 engine-page rows, 1,105 different addresses, because some pages were cited by more than one engine. A page cited by two engines is read once and appears in both engines' counts. Pages from a large publisher can appear more than once; the 774 pairs below cover 721 websites. - **Fetch.** On 2026-11-11 each address was requested once by plain HTTP with a research user agent and a 15 second limit. A page counts as readable if it answered 200 with HTML. 969 of the 1,250 rows (77.5 percent) were readable. The other 281 were unreadable, not zero: most returned 403, the rest a connection failure, 429, 404 or another status. A site that blocks automated reads may differ from one that does not, so these figures describe the readable pages. - **Dates.** The publish date is the earliest of schema.org `datePublished` and `dateCreated` in the page's JSON-LD and the usual publish meta tags (for example `article:published_time`). The update date is the latest of `dateModified` and the update meta tags. Only dates in ISO form or a written month-day-year were read. Visible bylines such as "Updated March 3" in the page body were not parsed, so a page that shows its date only to a human counts as exposing none here. - **Age.** A page's age is its most recent citing run date minus its publish date. A publish date later than the citing run (14 of 645 rows) is left out of age. - **Comparison.** For each readable cited page we read the website's sitemap, kept addresses the ledger has never recorded as cited by any engine or panel (16,391 cited addresses on the 988 sampled websites were excluded), preferred a page in the same first folder (for example `/blog/`), and drew one at random. 782 pairs were drawn, 774 with both pages readable. ## How old the cited pages are Of 631 cited rows with a publish date no later than the citing run: | Age on the last citing run | Share of pages | |---|---:| | 0 to 30 days | 9.5% | | 31 to 90 days | 16.0% | | 91 to 180 days | 20.8% | | 181 to 365 days | 21.9% | | 1 to 2 years | 11.3% | | 2 to 3 years | 5.1% | | More than 3 years | 15.5% | The median is 199 days. 68.1 percent were under a year old (64.4 to 71.7) and 20.6 percent over two years (17.6 to 23.9). By engine, median age and share under 90 days: | Engine | Rows with a publish date | Median age (days) | Under 90 days | Under 1 year | |---|---:|---:|---:|---:| | ChatGPT | 100 | 216 | 25.0% | 66.0% | | Claude | 135 | 217 | 17.8% | 72.6% | | Gemini | 120 | 153 | 30.8% | 70.0% | | Google AI Mode | 139 | 227 | 19.4% | 61.2% | | Perplexity | 137 | 145 | 35.0% | 70.8% | Gemini and Perplexity cited the youngest pages and AI Mode and Claude the oldest, in this sample. Each row has 100 to 139 pages and the study did not test whether the engines differ, so it claims no ranking. Perplexity's API has returned a fixed ten sources per answer since 2026-10-07, which changes which pages it cites; the Index treats that as its own collection regime. ## Cited page against a neighbour on the same website | Comparison (both pages readable) | Pairs | Cited page later | Cited page earlier | Same date | Cited page later, of decided | |---|---:|---:|---:|---:|---:| | Publish date | 485 | 246 | 203 | 36 | 54.8% (50.2 to 59.3) | | Update date | 467 | 237 | 163 | 67 | 59.2% (54.4 to 64.0) | The median gap was +3 days on publish date and +2 days on update date, the cited page being later. Measured on 2026-11-11, the median paired cited page was last updated 112 days earlier against 133 for its neighbour, and 85.0 percent of cited pages (the update-date pairs) had been updated within a year against 77.3 percent of neighbours. Two cautions apply to the update-date row. A page's update date is read now, not on the day the engine cited it. In 146 of the 625 cited rows that carry one, the update date is later than the last citing run, so some of the "later" cited pages were refreshed after they were cited. Templates also stamp an update date automatically. We therefore treat the publish date as the cleaner comparison, and its interval only just excludes an even split (two-sided sign test, p = 0.047, on decided pairs, one of two significance tests in this study and not corrected for that). ## Do cited pages show a date? Among the 774 pairs, a publish date was readable on 70.9 percent of cited pages (67.6 to 74.0) and 65.6 percent of neighbours (62.2 to 68.9); an update date on 68.2 and 63.8 percent. Across all 969 readable cited rows, 66.6 percent exposed a publish date and 69.1 percent either date. About three in ten pages engines cited, by this reading, give a crawler no machine-readable date, so engines can and do cite undated pages. ## What this changes - **Freshness is not the lever the advice suggests.** A median of 199 days and one cited page in five over two years old describe a body of sources that is not mostly recent. The study did not test whether changing a page or its date moves a citation. - **A real date costs nothing.** Cited pages were somewhat more likely to carry one, by 5 points, with overlapping intervals. Publishing an accurate `datePublished` and `dateModified` is hygiene, not a tactic. Do not change a date without changing the page. - **Stop reading "updated" as a ranking signal.** The association with update date is confounded by pages updated after citation, and the publish-date association is small and was one of two tests. - **Test your own pages, not an average.** The question is whether your page's date is readable and accurate, and whether it answers the question. The study does not tell you that a refresh will earn a citation. ## What this page does not show - **Any effect of freshness on citation.** The comparison is observational. A cited page can be newer because it was cited, because it was updated after being cited, or because the website publishes more often. - **That the neighbour is a fair control.** The comparison page is one sitemap address, in the same folder where possible. It is not matched on topic, length or links, and sitemap order and crawl scope may favour recent pages. 68 of the 799 websites had no readable sitemap and so contribute no pair. - **The dates engines see.** We read the live HTML on 2026-11-11, not the page as it was when the engine cited it. A page edited since carries its later date. - **Pages that block automated reads.** 22.5 percent of rows were unreadable, mostly 403, and are excluded. - **Visible-only dates and non-JSON-LD formats.** Microdata, `