---
title: "How fresh are the pages AI engines cite? | Underneath"
description: "Held to the same questions, AI assistants cited pages first published half as long ago as Google’s top 10. AI Overviews showed no pull toward recent pages."
canonical: "https://underneath.agency/research/ai-source-freshness-study"
published: 2026-09-26
updated: 2026-10-08
publisher: "Underneath (https://underneath.agency/agent)"
entity: "https://underneath.agency/.well-known/entity.json"
---
Research · AI search

# How fresh are the pages AI engines cite?

Does AI prefer new content? We measured the age of the pages six AI engines cite and compared them with the pages Google ranks. Version 1.0 found that the assistants (ChatGPT, Gemini, Perplexity and Claude) cited clearly fresher pages than Google ranks, while Google’s AI Overviews and AI Mode did not. Version 1.1 asks why. It compares each assistant with Google’s results for the same question, looks at which pages win within the same results, and separates a page that is new from a page that only carries a new date. The main finding holds in a sharper form: the assistants’ edge is mostly pages first published recently, not old pages with recent dates.

## The short version

1. Against Google’s top 10 for the same 80 questions, the four assistants together cited pages first published about half as long ago (ratio 0.50, 95% interval 0.38 to 0.65). Each engine on its own shows the same direction, with intervals below 1.
2. Pages published in the last 90 days made up 17.4% to 22.6% of each assistant’s dated citations, against 6.9% of Google’s top 10. Within the same question, the assistants cited 11.9 points more new pages (interval 6.4 to 16.9).
3. On the newest date a page declares, the version 1.0 measure, the gap is smaller once the question is held fixed. Google’s pages for the assistants’ questions had a median age of 92 days, not 111. Only ChatGPT stays clearly fresher on its own (0.55 times the age, interval 0.36 to 0.80).
4. Within Google’s top 10 for a question, the pages an assistant cited were more often dated within 90 days: +6.2 points (interval 1.0 to 11.6) on a base rate of 15.5%.
5. AI Overviews showed no such pull. On the same results page, a dated page was more likely to be cited than an undated one, but among dated pages a recent date made no difference (−4.1 points, interval −9.6 to 1.3).
6. Most recent dates on the web are old pages with a new modified date: 66.4% of the recently dated pages in Google’s top 10, against 47.2% for ChatGPT.
7. None of this is a causal effect. We changed no pages. The test that could show whether updating a page changes its citations is set out below. We have not run it yet.

## Why this version changes the question

Version 1.0 described what the engines cite. A difference in age between cited pages and Google’s pages has at least three explanations, and they call for different responses.

- **Selection preference.** The engine favors recent pages over older ones that answer the same question. If so, fresher pages should win *within* the same question or results page.
- **Query-driven recency.** The questions asked call for recent information. If so, the gap should shrink when cited pages are compared with Google’s pages for the same question, and it should be larger for time-sensitive queries.
- **Source composition.** The engines cite kinds of pages (articles, dated editorial content) that are recent anyway. If so, the gap should shrink when each engine’s pages are reweighted to Google’s mix.

A fourth question cuts across all three: is a “fresh” page new, or an old page with a new date? A declared date can change without the content changing, so we read each page’s publish date and modified date separately.

Freshness can act at more than one stage. A page has to be found (retrieval), then chosen as a source (citation), then used (absorption) and represented correctly (fidelity). This study sees only the citation stage, with Google’s top 10 as a stand-in for what could have been found. It says nothing about how cited pages were used.

| Analysis | Unit | Compared with | Tests |
|---|---|---|---|
| Same question | Page cited for a question | Google top 10, same question | B |
| Within results | Page in a top 10 | Other pages, same results | A |
| Reweighting | Dated cited page | Google’s page mix | C |
| Date anatomy | Dated page | Publish vs modified date | New or re-dated |

## What we measured

The cited pages are the version 1.0 sample: up to 400 pages per engine drawn at random from the citations collected on 26 September 2026, AI Overviews and AI Mode for 800 US searches across eight industries, and ChatGPT, Gemini, Perplexity and Claude for 80 buyer questions. Platform pages (YouTube, Reddit, social networks, Google, Amazon, Yelp) and PDFs are excluded. New in version 1.1, each page is linked back to the queries that cited it.

Two Google baselines. The first is the version 1.0 baseline, the 3,096 top-10 pages for the searches in our [study of pages cited by AI Overviews](https://underneath.agency/research/ai-overview-cited-pages-study). The second is new: Google’s top 10 for the assistants’ own 80 questions, from our [study of AI citations and Google rankings](https://underneath.agency/research/ai-citations-google-rankings-study) (26 September). We fetched those 547 non-platform pages by plain request on 28 September.

Age is the time from a page’s newest machine-readable date (structured data, article meta tags, time elements) to 26 September, as in version 1.0. Publish age uses the earliest declared publish date. A new page is one published in the last 90 days. Intervals come from resampling whole queries, not single pages, since pages cited for the same question are not independent.

## Finding 1: held to the same questions, the gap narrows on dates and holds on new pages

Google’s top 10 for the assistants’ 80 questions is younger than the version 1.0 baseline: a median of 92 days (interval 67 to 121), against 111 days for the 485 searches of the baseline. Part of the gap in version 1.0 therefore came from the questions, not from the engines.

The table compares each assistant with Google’s pages for the same question. An age ratio below 1 means the assistant’s cited pages are younger; the difference in pages dated within 90 days is in percentage points.

| Engine | Age ratio | 95% interval | Within 90 days |
|---|---|---|---|
| ChatGPT | 0.55 | 0.36 to 0.80 | +13.5 |
| Claude | 0.76 | 0.52 to 1.09 | +7.0 |
| Perplexity | 0.78 | 0.53 to 1.15 | +8.0 |
| Gemini | 0.78 | 0.50 to 1.15 | +6.8 |
| All four | 0.70 | 0.51 to 0.94 | +8.8 |

On the newest declared date, only ChatGPT is clearly fresher on its own; for Claude, Perplexity and Gemini the direction is the same but the intervals include no difference. The within-90-day intervals follow the same pattern (ChatGPT 4.2 to 23.5 points; all four 0.7 to 17.2).

The publish date tells a clearer story. Measured by when a page was first published, every assistant cites younger pages than Google ranks for the same question.

| Engine | Publish-age ratio | 95% interval | New pages |
|---|---|---|---|
| ChatGPT | 0.39 | 0.24 to 0.61 | +16.1 |
| Claude | 0.55 | 0.38 to 0.77 | +9.5 |
| Perplexity | 0.66 | 0.48 to 0.94 | +10.9 |
| Gemini | 0.49 | 0.35 to 0.69 | +12.1 |
| All four | 0.50 | 0.38 to 0.65 | +11.9 |

New pages is the difference, in points, in the share of pages published in the last 90 days. The intervals are 7.0 to 25.8 (ChatGPT), 2.6 to 17.3 (Claude), 4.2 to 16.9 (Perplexity), 3.8 to 20.4 (Gemini) and 6.4 to 16.9 (all four).

Why the two measures differ: Google’s top pages are often old pages with a recent modified date, which makes them look fresh on the newest date. The assistants cite more pages that are actually new.

## Finding 2: new pages or new dates

We sorted every dated page by what its dates say.

| Engine | New pages | 95% interval | Median publish age (days) |
|---|---|---|---|
| Gemini | 22.6% | 14.8 to 31.5 | 284 |
| ChatGPT | 22.2% | 14.8 to 30.3 | 294 |
| Perplexity | 19.2% | 13.7 to 24.0 | 537 |
| Claude | 17.4% | 11.8 to 23.8 | 311 |
| AI Overviews | 10.8% | 6.0 to 16.1 | 575 |
| Google top 10, same questions | 9.0% | 5.7 to 12.5 | 733 |
| AI Mode | 7.8% | 4.0 to 12.6 | 736.5 |
| Google top 10 (baseline) | 6.9% | 5.5 to 8.3 | 801 |

Among pages dated within the last 90 days, the share that are older pages with a recent modified date was 66.2% for AI Overviews, 68.4% for AI Mode and 66.4% for Google’s top 10, against 47.2% for ChatGPT, 54.9% for Perplexity, 56.1% for Gemini and 59.3% for Claude. A further 2.3% to 11.0% of recently dated pages carried a date within a day of our fetch, which usually means the site stamps the current date on every page.

So a recent date on the web usually marks a revision, not a new page, and Google’s results reflect that. The assistants lean further toward pages that are new.

## Finding 3: within the same results, who picks the fresher page

For each of the 80 questions we took the dated pages in Google’s top 10 and asked whether the pages an assistant cited were more often recent than the ones it passed over, with the organic position held fixed. Across 320 engine and question pairs, 15.5% of these pages were cited. A page dated within 90 days was 6.2 points more likely to be cited (interval 1.0 to 11.6). This is the pattern a selection preference would produce, though a recent page may also differ in ways we did not measure.

AI Overviews behave differently. On the same Google results page and at the same position, pages with any machine-readable date were more likely to be cited than pages without one: +5.7 points for a date within 90 days, +8.3 for 91 to 365 days and +12.1 for over a year (intervals 1.0 to 10.3, 2.6 to 13.6 and 7.1 to 17.2). Among the 1,480 dated pages, a recent date made no measurable difference (−4.1 points, interval −9.6 to 1.3). What goes with citation in AI Overviews is declaring a date, not having a recent one; this matches [the date finding in our cited-pages study](https://underneath.agency/research/ai-overview-cited-pages-study).

## Finding 4: the questions matter, but we could test that only on Google

If freshness depends on the question, it should matter most for time-sensitive queries (a year, “latest”, prices, rates) and least for evergreen ones. We classed every query by its wording. Of the 800 Google searches, 105 were time-sensitive, 16 were recommendations and 679 were other searches (informational, navigational or local). Of the 80 assistant questions, 71 were recommendations (“What is the best…”) and only 9 were time-sensitive, too few to test.

On Google’s side the test shows no clear effect. In AI Overviews, a recent date made no measurable difference on time-sensitive searches (+2.0 points, interval −11.0 to 13.0, 86 searches) or on other searches (−5.3, interval −11.7 to 0.9). By Google’s own intent label, recent pages were less likely to be cited on commercial searches (−9.3, interval −16.0 to −2.4) and not clearly more likely on informational ones (+6.3, interval −2.9 to 16.0). The query effect we can see is the one in finding 1: the assistants’ questions have younger Google results to begin with.

## Finding 5: the kind of page does not explain the gap

The assistants cite pages with article markup somewhat more often than Google ranks them (63.6% for ChatGPT to 90.4% for Gemini, against 59.0%), and articles are more often recent. Reweighting each engine’s pages to Google’s mix of page kinds (article markup or not, by reference, review, retail or other site) barely moves the results.

| Engine | Median age | Reweighted | Within 90 days, reweighted |
|---|---|---|---|
| ChatGPT | 50.5 | 43 | 66.5% |
| Claude | 54 | 54 | 60.4% |
| Perplexity | 60 | 60 | 57.8% |
| Gemini | 65 | 65 | 56.4% |
| AI Mode | 111 | 103 | 46.1% |
| AI Overviews | 113 | 117 | 42.9% |

Ages in days. At this level of detail, source composition does not account for the difference. A finer classification (topic, publisher, commercial versus editorial intent) could still find one.

## The three explanations, weighed

- **Selection preference: supported for the assistants, not for Google’s AI.** Within Google’s top 10 for a question, the assistants picked recent pages more often, and they cited younger pages than Google ranks for the same question. AI Overviews showed no preference for recent pages among dated ones.
- **Query-driven recency: part of the version 1.0 gap.** Google’s own results for the assistants’ questions are younger than for the searches in the version 1.0 baseline, which narrows the gap on the newest date. Whether freshness matters more for time-sensitive questions is not settled: the assistants’ questions barely vary, and on Google the intervals are wide.
- **Source composition: not supported at the level measured.** Reweighting to Google’s page mix leaves the ages almost unchanged.

## Version 1.0 results, kept for comparison

The version 1.0 figures are unchanged. The intervals now come from resampling whole queries.

| Engine | Dated | Median (days) | 95% interval | Within 90 days |
|---|---|---|---|---|
| ChatGPT | 162 | 50.5 | 36 to 69.5 | 65.4% |
| Claude | 149 | 54 | 40 to 75.5 | 61.1% |
| Perplexity | 219 | 60 | 42 to 99 | 55.7% |
| Gemini | 115 | 65 | 45 to 102 | 57.4% |
| AI Mode | 167 | 111 | 74.5 to 158 | 45.5% |
| Google top 10 | 1,480 | 111 | 93 to 127 | 46.2% |
| AI Overviews | 166 | 113 | 75 to 157 | 44.6% |

Between 4.0% (Claude) and 7.8% (Perplexity) of the assistants’ dated citations were more than two years old, against 13.9% of Google’s top 10. The share of fetched pages with a machine-readable date ranged from 47.8% (Google’s top 10) to 81.0% (Gemini).

## What this means

These are inferences from observational data, not measured effects.

- **For the assistants, new pages have an edge that re-dated pages do not show as clearly.** The assistants’ advantage is largest on the publish date. A new, substantive page on a topic is the thing the data points to; changing the date on an old one is not.
- **Google’s AI follows Google’s results.** AI Overviews and AI Mode cite pages as old as Google’s top 10, and in AI Overviews a recent date does not help a page that already ranks. Declaring a date, accurately, does go with citation there.
- **Compare like with like.** Much of the freshness advice in circulation compares AI citations with the web at large. Held to the same questions, the gap is smaller and depends on the engine.

## What a causal test would need

Only an experiment can say whether updating a page causes more citations. The design we plan is below; it needs pages we control and budget for repeated runs, and we have not run it.

- Four versions of comparable pages, assigned at random: unchanged, date changed only, a minor edit, and a substantive update (new facts, figures and comparisons).
- Queries in two groups, time-sensitive and evergreen, so the effect can differ by what is asked.
- Several engines, each query run several times per date and on several dates, since answers vary from run to run.
- Outcomes beyond citation: whether the page is cited at all, how prominently, whether its content is used in the answer, and whether the answer states what the page says.
- A condition where competing pages are updated too, to tell an absolute gain from one that only holds while competitors stay still.

The hypotheses, stated before any data: a substantive update raises the chance of citation more than a date change alone; a date change alone has little or no effect; the effect is larger for time-sensitive queries; and it differs by engine.

## Methodology

- **Cited pages:** the version 1.0 sample, up to 400 per engine drawn at random (seed 20260926) from citations collected on 26 September 2026 for our AI Overview frequency, AI Overview citation, AI Mode and four-assistant studies; platform pages and PDFs excluded; linked back to every query that cited them.
- **Baselines:** the 3,096 top-10 pages of our study of pages cited by AI Overviews; and the Google top 10 for the assistants’ 80 questions (26 September), 547 non-platform pages fetched by plain request on 28 September 2026.
- **Dates:** newest valid date among JSON-LD dateModified, datePublished and uploadDate, article meta tags, og:updated_time and time elements; publish date is the earliest declared datePublished, uploadDate or article:published_time; ages counted to 26 September 2026.
- **Models:** linear probability models with question or search fixed effects and dummies for organic positions; comparisons of cited pages with Google pages for the same question use question fixed effects and log age.
- **Query classes:** rules on the query text (time-sensitive, recommendation, other), and DataForSEO’s intent label for Google searches.
- **Intervals:** bootstrap resampling whole queries, 2,000 resamples for descriptive figures and 1,000 for models (seed 20260928).
- **Code:** s22_v11.py (collect, analyze, package), and s22_fresh.py for version 1.0.
- **Update schedule:** quarterly, with a second collection date to measure stability.

## Limitations

- Observational only: nothing was changed on any page, so no difference here is a causal effect of freshness.
- Retrieval is not observed. Google’s top 10 is a stand-in for what could have been found; the assistants run their own searches.
- Declared dates are what a page says, not what changed on it. We did not compare page versions over time.
- One collection date; run-to-run and month-to-month stability are not measured.
- 71 of the 80 assistant questions are recommendations, so query temporality could not be tested for the assistants.
- The same-question Google pages were fetched two days after the cited pages. Any update in between makes them look slightly fresher, which, if anything, narrows the gap.
- 23.0% to 32.5% of sampled cited pages per engine could not be fetched by plain request, and only pages with a machine-readable date are aged.
- Query classes come from simple word rules and were not checked by hand.

## Data and downloads

- Every cited and same-question page with its query, dates, date anatomy and query class: [s22_v11_pages.csv](https://underneath.agency/research-data/ai-source-freshness-study/s22_v11_pages.csv)
- The version 1.0 sample with ages: [s22_cited_page_dates.csv](https://underneath.agency/research-data/ai-source-freshness-study/s22_cited_page_dates.csv) and [JSON](https://underneath.agency/research-data/ai-source-freshness-study/s22_cited_page_dates.json)
- Every statistic on this page, version 1.0 and 1.1: [stats.json](https://underneath.agency/research-data/ai-source-freshness-study/stats.json)
- Machine-readable methodology: [methodology.json](https://underneath.agency/research-data/ai-source-freshness-study/methodology.json)

The data is free to reuse with attribution (CC BY 4.0).

To cite: Underneath. (2026). *How fresh are the pages AI engines cite?* (Version 1.1). Underneath Research. https://underneath.agency/research/ai-source-freshness-study

## Frequently asked questions

### Does ChatGPT prefer recent content?

The pages ChatGPT cited were younger than the pages Google ranks for the same questions: 0.55 times the age on the newest date, and 0.39 times on the publish date. That is a pattern in what it cites, not proof that recency causes citation.

### Do Google AI Overviews prefer fresh content?

Not in our data. Cited pages were as old as Google’s top 10, and among dated pages on the same results page a recent date made no measurable difference. Pages that declared a date at all were cited more often.

### Which AI engine cites the freshest sources?

ChatGPT, by the newest declared date (median 50.5 days). By publish date, ChatGPT and Gemini cite the most new pages (22.2% and 22.6% published in the last 90 days).

### Does updating the date on a page help it get cited by AI?

Our data does not support it. Most recently dated pages on the web are old pages with a new modified date, and the assistants’ edge is in pages that are actually new. Whether a substantive update causes more citations is the experiment described above, which we have not yet run.

## Related research

- [What pages cited by AI Overviews have in common](https://underneath.agency/research/ai-overview-cited-pages-study)
- [The hidden searches AI assistants run before they answer](https://underneath.agency/research/ai-hidden-searches-study)
- [Do ChatGPT, Gemini, Perplexity and Claude cite pages that rank?](https://underneath.agency/research/ai-citations-google-rankings-study)
- [How often do AI Overviews appear?](https://underneath.agency/research/ai-overviews-frequency-study)

## Related guides

- [What on-page signals are linked to citations in Google AI Overviews and Perplexity?](https://underneath.agency/resources/on-page-signals-linked-to-ai-citations)
- [Why doesn’t ChatGPT mention our newly launched product?](https://underneath.agency/resources/why-chatgpt-misses-new-products)
- [Do AI search engines cite the same websites as Google?](https://underneath.agency/resources/do-ai-search-engines-cite-the-same-sites-as-google)
- [Will optimizing content for ChatGPT hurt our Google rankings?](https://underneath.agency/resources/chatgpt-optimization-google-rankings)
- [What decides whether an AI engine cites my page over a competitor’s?](https://underneath.agency/resources/why-ai-cites-competitor-page-first)

---

This is the Markdown twin of https://underneath.agency/research/ai-source-freshness-study. The HTML page is canonical. Publisher: Underneath, https://underneath.agency/agent. Site index: https://underneath.agency/llms.txt.