Recency bias is the documented tendency for modern answer engines to cite newer pages over older ones, even when the older page has higher topical authority. The bias is strong enough that publication date functions as a tie-breaker on competitive queries, and severe enough that pages older than a year often lose their citations entirely. The result has been a structural change in how content investment is paced.
How strong the bias actually is
A 2024 Ahrefs study of ChatGPT browsing citations found that 60 percent of cited URLs were under 12 months old, despite the indexed corpus being far older. Public benchmarks of Perplexity show similar skew. The mechanism is not entirely clear from outside, but appears to be a combination of explicit recency reranking (the retriever weights date), training-data recency (the underlying LLM has seen more recent content) and corpus dynamics (newer content tends to score higher on entity density and definition-first patterns because those patterns have themselves become more common recently).
Why this exists
Three forces. Factual evolution: on most topics, more recent answers are more likely to be correct, so retrieval that biases recency is a defensible default. Quality drift: older content was written for SEO patterns that have since dated, while newer content is increasingly written for AI retrieval, raising the average quality of recent results. Engine economics: encouraging fresh content keeps the corpus refreshed, which is what the major engines need to differentiate from each other and from the static web.
What this means for content cadence
Two practical changes. First, the canonical asset of an AEO program is no longer a one-time flagship piece but a flagship piece on a refresh cycle — the cadence is part of the asset. Second, the optimal portfolio is a smaller number of relentlessly-maintained pieces rather than a large library of write-once content. A team running 30 evergreen pieces refreshed quarterly will out-cite a team running 100 pieces refreshed never, by a wide margin.
What does and does not count as recency
Updating dateModified in the schema without changing content is increasingly being detected and devalued — engines look at content diff, not just timestamps. A real edit (new section, new entities, refreshed numbers) is what restores citation odds. Adding a published-date stamp to the visible page also helps; many comparison engines parse the date out of the visible HTML rather than the schema.
How to plan around the bias
Pick your top 20 pages and put them on a 60 to 90 day refresh cadence. Each refresh is one real edit, one schema update, and one re-scan a week later. For non-evergreen content, accept that citations will decay over time and plan to either ship a successor (which carries forward by virtue of being newer) or sunset and merge into the flagship. Programs that organize around the curve rather than fighting it consistently outperform programs that publish-and-pray.