Content freshness

Ahrefs measured the publication age of 16.975 million cited URLs pulled from ChatGPT, Perplexity, Gemini, Copilot, AI Overviews and organic Google results^[3]. The per-engine averages separate the assistants from search: ChatGPT cites pages averaging 958 days old, Copilot 1,056, Gemini 1,118 and Perplexity 1,166, while AI Overviews at 1,432 days sits alongside the organic Google result set at 1,416^[3]. Google's AI surface, in other words, cites material about as old as its own index does. The study's headline compares 1,064 days against 1,432 for organic results, which is the figure its own per-engine table gives for AI Overviews rather than for organic Google^[1]. Ahrefs reads the gap as a preference for newer content rather than a causal mechanism, and stops short of claiming that updating a page earns a citation^[4].

The 1.4 million prompt study looked at the candidates one prompt retrieved rather than at citations in aggregate, and found the freshest pages were the ones discarded, with the median cited page around 500 days old and some past 2,700^[2]. Age still did work, but narrowly: in news, where relevance scores for cited and non-cited pages are nearly identical, page age acts as the tie-breaker and the younger page wins^[5]. The conclusion the authors draw is that a match against the model's fan-out sub-queries decides the citation: a new page that matches gets cited, a new page that does not is retrieved and ignored^[6]. See retrieval-augmented generation for the two-stage process this describes.

A controlled experiment isolates the effect the field studies can only infer. Injecting newer publication dates into otherwise identical passages, and reranking them with GPT-3.5-turbo, GPT-4o, GPT-4, LLaMA-3 and Qwen-2.5, moved the mean publication year of the top ten forward by up to 4.78 years, shifted individual passages by up to 95 ranks, and reversed the preference between equally relevant passages by up to 25% on average^[7]. Larger models reduced the bias without removing it^[7]. A date on a page is therefore an input to the ranker itself, not just a line of display metadata.

A second benchmark finds the opposite bias in the same component. FRESCO pairs recency-seeking queries with historical Wikipedia revisions and reports one failure mode common to the re-rankers it tests: a strong bias toward older, semantically rich documents even where those documents are factually obsolete^[8]. The two experiments disagree about direction rather than about mechanism. An injected date moves the ranking; which way it moves depends on how much semantic weight the older passage carries against the newer one. Optimising the instruction given to the re-ranker recovers up to 27% on the recency-sensitive tasks while holding performance on stable-knowledge ones^[9], so the bias is addressable at the prompt layer rather than fixed in the model.

Google describes query deserves freshness systems that surface newer content for queries where a user would expect it, such as a newly released film, rather than a general boost for recent pages^[10]. For AI surfaces it documents no freshness requirement at all: a page needs only to be indexed and eligible for Search to appear as a supporting link in AI Overviews or AI Mode, with no additional requirements or special optimisations^[11]. The Article structured data reference defines dateModified as the date and time the article was most recently modified, in ISO 8601 format, and lists it as a recommended property for date display in search results rather than a condition of citation^[12].

Whether declaring dates helps has been tested directly. Ahrefs tracked 1,885 pages that added JSON-LD markup against 4,000 control pages between August 2025 and March 2026 and found citation rates barely moved: AI Overviews down 4.6%, AI Mode up 2.4% and ChatGPT up 2.2%, the last two statistically indistinguishable from zero^[13]. Related: citation accuracy and content structure.