Ahrefs analyzed 16,975,000 citations across ChatGPT, Perplexity, Gemini, Copilot, and Google's AI Overviews, then compared them against organic Google rankings for the same queries. The average AI-cited page was 1,064 days old. The average organic result was 1,432 days old. That's a 25.7% gap, and it holds whether you measure time since publication or time since the page was last updated. ChatGPT skewed freshest of the five platforms tested. Google's own AI Overviews barely differed from plain organic search. A separate 2025 academic study then found something less comfortable underneath that pattern: large language models can be fooled by a fake publish date almost as effectively as by a real one.
What 17 million AI citations reveal about content age
Ahrefs' analysis is the largest side-by-side comparison published so far of content age in AI citations versus organic search rankings, and it points the same direction across every platform tested. AI assistants pull from newer pages than Google's own top 10 results, though how much newer depends heavily on which engine is doing the answering.
The platform breakdown is wide. ChatGPT's direct citations averaged 958 days old and its inline references 1,023 days; Copilot averaged 1,056 days, Gemini 1,118, and Perplexity 1,166. Google's AI Overviews averaged 1,432 days, statistically close to the 1,416-day average for the organic results sitting just below them on the same page. Update recency followed a similar pattern: AI-cited pages had last been revised 909 days before the study ran, versus 1,047 days for organic results, roughly a 13% gap. Ahrefs also found that ChatGPT and Perplexity tend to list citations from newest to oldest within a single answer, which the researchers read as a deliberate ordering rather than chance.
Old content still wins most of the time
The freshness gap is real, but it's smaller than the headline number implies. Even the freshest citations in Ahrefs' data averaged 2.9 years old, which means most of what AI engines cite today was written long before anyone was optimizing for them. Freshness tilts the odds. It doesn't override everything else on the page, and that lines up with what we found when we mapped out how GEO differs from classic SEO.
Google's own documentation describes a "query deserves freshness" system that only promotes recent pages when the query itself calls for it: a movie that just opened, an earthquake from an hour ago, according to Google's ranking systems guide. A static explainer of what GEO actually means doesn't need a monthly rewrite to stay competitive. A piece about a fast-moving platform shift, like the twenty-point swing in ChatGPT's US prompt share we covered this week, is a different story. It goes stale within weeks, and AI engines seem to weight that difference. The distinction is more useful for planning a content calendar than any blanket "publish weekly" rule.
What happens when you fake the date
A study presented at the ACM SIGIR-AP conference in September 2025 tested exactly what happens when a passage's real publication date is swapped for a newer, fake one, with the underlying text left untouched. The bias it uncovered is large enough to worry anyone who treats a timestamp as a lever instead of a record.
Researchers Hanpei Fang, Sijie Tao, Nuo Chen, Kai-Xin Chang, and Tetsuya Sakai ran the experiment across seven models: GPT-3.5-turbo, GPT-4o, GPT-4, LLaMA-3 in two sizes, and Qwen-2.5 in two sizes. In listwise reranking tests, prepending a newer date shifted the mean publication year of the top 10 results forward by as much as 4.78 years, and individual passages jumped by as many as 95 ranking positions, purely on the strength of a changed date on identical text. In head-to-head comparisons between two equally relevant passages, the model's preference flipped toward the newer-dated one up to 25% of the time on average. Larger models resisted the bias somewhat better than smaller ones, but none of the seven models eliminated it.
The bias held even though the underlying passage never changed. An AI engine can prefer a page for no reason beyond a newer-looking date, whether or not anything on the page actually got better.
What a genuine freshness signal looks like
Touching a timestamp is cheap, and the recency-bias research shows it can work, at least within a single reranking pass. As a repeatable strategy it's close to worthless, because nothing about the content changed, and both search engines and the AI systems built on top of them are getting better at telling a real revision apart from a cosmetic one.
Sites that try to solve this by publishing a near-duplicate "refreshed" version next to the original usually make things worse. Near-duplicate pages get grouped by AI models and reduced to a single citation, and it's rarely the newest of the two that survives that grouping. The safer move is to update the actual page: replace stale statistics, correct a figure the original source has since revised, add a section covering something that happened after publication. That's the same discipline behind the research we've covered on why word count barely predicts AI citations and why topical coverage beats domain rank for AI citations. The metric an AI engine rewards is rarely the surface-level signal on its own; it's the work that signal is supposed to represent.
A freshness signal only means something when it's attached to a real change. Set it honestly and it becomes one more piece of evidence that a page is still worth citing. Fake it, and the only thing being tested is how long a model can be fooled by a number.
Frequently Asked Questions
Does content freshness actually affect AI citations?
Yes. Ahrefs' analysis of nearly 17 million citations found that pages cited by AI assistants average 1,064 days old, compared with 1,432 days for organic Google results, a 25.7% gap. The effect varies by platform: ChatGPT shows the strongest preference for newer content, while Google's AI Overviews behave almost like ordinary organic search.
Which AI engine cites the freshest content?
ChatGPT, according to Ahrefs' 2026 study. Its direct citations averaged 958 days old and its inline references 1,023 days, both noticeably newer than Perplexity (1,166 days), Gemini (1,118 days), or Google's AI Overviews (1,432 days, nearly identical to organic search).
Can you fake freshness by just changing the publish date?
Briefly, and only in a narrow sense. A September 2025 study tested seven large language models and found that swapping in a newer fake date, with no change to the actual text, shifted rankings and flipped model preferences up to 25% of the time. The underlying content never improved, so the trick doesn't hold up against a real revision and is getting easier to detect.
How often should content be updated for GEO purposes?
It depends on the topic, not a fixed schedule. Fast-moving subjects, like a platform's market share or a newly launched AI feature, go stale within weeks and benefit from frequent updates. Evergreen explainers of stable concepts don't need monthly rewrites; Google's own "query deserves freshness" system only rewards recency when the query itself calls for it.
Does freshness matter more than other GEO signals?
No. Even the freshest AI citations in Ahrefs' data averaged 2.9 years old, meaning most cited content predates any deliberate GEO effort. Freshness is one factor among several, alongside topical coverage and answer-first structure, not a signal that can outweigh thin or weak content on its own.
What counts as a real content update versus a cosmetic one?
A real update changes the substance: new statistics, a corrected figure, a section covering a development that happened after publication. A cosmetic update only changes the timestamp or system metadata while the text stays the same, which is the exact pattern the 2025 recency-bias study found models rewarding even though nothing on the page actually improved.