An AI assistant answering a question does not read your whole archive. It retrieves a handful of candidate pages, and one of the signals it can weigh is how recent each page looks. Two large studies now put numbers on that tilt. Ahrefs analysed almost 17 million cited URLs and found that ChatGPT, Perplexity, Copilot and Gemini cite pages roughly a quarter newer than those Google ranks. Seer Interactive found that three quarters of the pages cited in a four-month window had been updated within a year. Neither study shows that a new date alone wins citations.
What does the research say about content age and AI citations?
Ahrefs used its Brand Radar data to extract cited URLs from ChatGPT, Perplexity, Gemini, Copilot, Google AI Overviews and Google organic results. It measured two ages for each URL: days since publication and days since last update. The sample was 16.975 million URLs, and the study was first published on 28 July 2025 (the page was later modified, on 27 April 2026).
The headline result is that the four AI assistants cited URLs averaging 1,064 days since publication, against 1,432 days for Google organic results. Ahrefs describes that as 25.7 per cent fresher. On last update the gap was smaller: 909 days against 1,047, or 13.1 per cent.
| Source | Average days since publication | Average days since last update |
|---|---|---|
| Google AI Overviews (top three) | 1,432 | 1,067 |
| Google organic results | 1,416 | 1,047 |
| Perplexity | 1,166 | 993 |
| Gemini | 1,118 | 831 |
| Microsoft Copilot | 1,056 | 865 |
| ChatGPT (references) | 1,023 | 865 |
| ChatGPT (direct citations) | 958 | 989 |
These are the figures from Ahrefs' table. The article text quotes slightly different numbers in a couple of places, so treat the table as indicative and check the source before reusing a specific figure.
Two things stand out. First, Google AI Overviews cited content slightly older than Google's own organic results, so Google's AI layer is the least freshness-driven of the group. Second, even the most recency-friendly assistant cites pages that are, on average, well over two years old.
Are most cited pages actually recent?
Seer Interactive's study, published on 24 July 2026, looked at the other end of the distribution. It covered ChatGPT, Gemini and Perplexity (not Google AI Overviews) from March to June 2026, using non-branded answers for four brands across four verticals: pet retail, vacation rentals, retail energy and commercial banking. After deduplication and a filter for pages cited at least three times, it dated 7,683 pages carrying 47,097 citations, using last-modified dates from schema, sitemaps and headers. Only about two thirds of pages could be dated at all.
Of the dated pages, 75 per cent had been updated within a year, 13 per cent were one to two years old, and the remaining 12 per cent were older. By engine, the share updated within a year was 78 per cent for Gemini, 73 per cent for ChatGPT and 65 per cent for Perplexity.
The two studies are less contradictory than they look. Ahrefs measures averages across a very large, mixed set of queries, where a long tail of old reference pages pulls the mean up. Seer measures how many cited pages were touched recently within commercial verticals, where being current matters more. Both are consistent with the same picture: most citations go to pages that have been maintained, but a good number go to old pages that remain the best answer.
Which kinds of content need to be fresh?
Freshness is not uniform across content types. In Seer's data, the share of cited pages updated within a year was highest for marketplaces (78 per cent) and comparison and review pages (77 per cent), followed by reference pages (74 per cent) and brand and corporate pages (72 per cent). Blogs and guides came in at 67 per cent.
News and editorial content was the outlier at 45 per cent. That is the finding most relevant to publishers. A news article is rarely updated after the day it runs, yet it is still cited, because its value is that it reports a specific event on a specific date. The freshness signal rewards pages whose answer changes over time, such as prices, rankings, rules and product specifications. It does not penalise an accurate report of something that happened last year.
Seer also separated pages cited steadily across all four months from pages cited in only one. The always-on group had a median update age of 0.47 years, while the single-month spikes had a median of 0.16 years. Short bursts of citation tend to attach to recently touched pages, which fits a news-driven or trend-driven query pattern.
Why does an updated date not equal fresh content?
Seer found that 72 per cent of its cited pages looked fresh by last-updated date, but only 42 per cent had been first published within the previous year. More than a quarter of the "fresh" pages were first published over two years ago. In other words, much of the freshness that AI assistants reward is the product of maintenance on older URLs.
That cuts both ways. It supports updating good older pages instead of replacing them. It also creates a temptation to change the date without changing the page. Ahrefs notes that Google's John Mueller has warned against altering publication dates without a corresponding content change, and the Ahrefs authors say they suspect most brands would do better by producing new, high-quality content than by updating very frequently. Low-quality content that is refreshed often is still low-quality content.
Google's own documentation on byline dates sets out how its systems work out a page's date. It looks at a clearly labelled visible date, such as "Published" or "Last updated", and at the datePublished and dateModified fields in structured data on article-type markup. Google recommends keeping the visible and structured dates consistent, including a time zone where possible, and minimising other dates on the page that could confuse the estimate. Google's guidance is written for Search, so it does not tell us how ChatGPT or Perplexity read dates. It is, however, a sensible baseline, because those systems also draw on schema, sitemaps and HTTP headers, which are the sources Seer used to date pages.
How should publishers handle freshness for AI search?
The evidence supports a measured approach rather than a refresh sprint.
Update pages whose answer changes. Evergreen explainers, comparison pages, pricing and rules pages, statistics roundups and how-to guides are where recency has measurable weight. Make a real change - new data, a corrected claim, a revised recommendation - and record it.
Leave accurate news reporting alone. Editing a 2024 report to add a 2026 date helps neither readers nor retrieval systems, and it risks undermining trust in the byline.
Show the date, and make it machine-readable. Use a visible "Published" and "Last updated" label and matching datePublished and dateModified values. Only move dateModified when the substance has changed. Where a page has been materially revised, a short note on what changed gives readers a reason to trust the new date.
Keep sitemaps and headers honest. Because AI systems and their crawlers can read last-modified values from sitemaps and response headers, make sure these reflect real edits and not a nightly rebuild that stamps every URL with the current time. Pages that all claim to have changed today give no signal at all.
Measure the effect on your own pages. Server logs show which of your URLs AI crawlers and agents actually fetch, and how that changes after an update. That is a more reliable test than any industry average, and it shows whether a refreshed page is being retrieved more often. Where AI agents are fetching pages live at the moment a user asks, the version they receive is the current one, which is a further reason to keep the live page accurate.
What are the limits of this evidence?
Both studies are observational. They show that cited pages tend to be newer than a comparison set, not that refreshing a page causes more citations. A page may be cited more because it is actively maintained by a strong site, and that same site is also likely to have updated it recently.
Date detection is imperfect. Seer could date only about two thirds of its pages, and Ahrefs relied on its own crawler's first-seen date as a proxy for publication. Seer's sample covers four verticals and three assistants, with one vertical overrepresented, so its percentages should be read as directional. Seer's published text and one of its vertical tables also disagree on a couple of figures, which is why this article does not rely on the per-vertical numbers.
Finally, the studies describe how citations were distributed in the period they covered. Assistants change their retrieval behaviour regularly, so any freshness effect measured in one quarter may shift in the next.
Frequently asked questions
Do AI assistants prefer new content to old content?
On average, yes, but modestly. Ahrefs found that ChatGPT, Perplexity, Copilot and Gemini cited pages about 25.7 per cent newer than Google organic results by publication age. The typical cited page was still around three years old, so AI assistants favour recency without discarding older material that answers the question well.
Should publishers change the published date when they update an article?
Only the "last updated" date, and only when the content has materially changed. Keep the original publication date visible, and keep the visible dates and the datePublished and dateModified structured data consistent. Altering dates without real edits is discouraged, and it adds nothing to the page's actual usefulness.
Does Google AI Overviews weigh freshness as much as ChatGPT?
The Ahrefs data suggests not. Pages cited in the top three Google AI Overviews were slightly older than Google's organic results, while ChatGPT, Gemini and Copilot cited newer pages. Google's systems appear the least influenced by recency of the platforms studied. Seer's separate study did not cover AI Overviews.
Does freshness matter for news articles?
Less than for evergreen content. In Seer's data only 45 per cent of cited news and editorial pages had been updated within a year, compared with 67 per cent for blogs and guides and 77 per cent for comparison and review pages. News is cited because it reports a specific event, not because it has been refreshed.
How often should a publisher refresh evergreen pages?
The research does not give a safe interval. Seer found a median update age of about half a year for pages cited consistently over four months, which suggests that pages in competitive, fast-changing topics benefit from review at least that often. For stable topics, update when the facts change, not on a calendar.
Sources: Ahrefs, "New Study: AI Assistants Prefer to Cite Fresher Content" (Ryan Law and Xibeijia Guan, ahrefs.com/blog/do-ai-assistants-prefer-to-cite-fresh-content); Seer Interactive, "Content Recency's Impact on AI Visibility in 2026" (Sonny Vasquez, 24 July 2026); Google Search Central, "Influence your byline dates in Google Search".