Yes - and there's now a controlled study that proves it rather than just observes it. Researchers fed seven large language models pairs of passages the models had already rated equally relevant, then changed nothing except the publication date attached to each one. Every model shifted its ranking toward the newer passage.
The study that tested it directly
The test ran across GPT-3.5-turbo, GPT-4o, GPT-4, LLaMA-3 (8B and 70B) and Qwen-2.5 (7B and 72B) in a September 2025 study on recency bias in LLM reranking. Injecting a fresher date into a passage shifted the mean publication year of a search result's top 10 forward by up to 4.78 years - years, not days - with individual passages moving by as many as 95 ranks purely because of the date attached to them.
The clearest signal was in direct comparisons. When two passages were rated equally relevant, swapping in a newer date flipped the model's preference by up to 25% on average. Larger models were less prone to it than smaller ones, but none of the seven tested eliminated the bias.
Why this happens
The paper doesn't pin down a single cause, but the likely explanation is straightforward. In the text these models learned from, recency correlates with relevance often enough - news, prices, product listings, sports results - that the association gets baked in as a general-purpose shortcut. Apply that shortcut to a topic where a five-year-old explanation is exactly as correct as one written last week, and it stops being useful and starts being a bias.
It shows up in real citations, not just the lab
Seer Interactive's tracking of 7,683 pages carrying over 47,000 real AI citations found the same pattern outside a controlled test: 75% of the pages models actually cited had been updated within the past year. The lab result and the field data point the same way.
Distribution slows the decay
A tracking study of more than 3 million citation events across six AI platforms over 26 weeks found a median citation half-life of 4.5 weeks for content that exists in one place - half its citation potential gone within about a month. Content referenced or republished across multiple trusted domains held on for roughly twice as long.
A single article ages out on its own. The same claim showing up in more than one place doesn't age out at the same rate, because a model has more than one place to find it. That's the same mechanism behind why branded mentions elsewhere correlate with AI visibility far more than backlinks do: a claim that exists in more than one location gives a model more places to draw on than a link count ever measures.
What this means for how you publish
An article you never touch again decays. Updating it, rather than replacing it, resets the clock without starting from zero, since the URL, the links pointing to it and its existing citations don't disappear when you edit the page.
Publishing on a defined schedule rather than in bursts also helps, simply because something recently touched sits in front of a model more often than a backlog of untouched pages does. Updating old pages is one of the ten things that moves a citation rate - on its own it won't fix a page nobody links to or a claim nobody repeats, but skipping it leaves a real, measurable gain unclaimed.
A few common follow-up questions
Does this mean old content is worthless?
No. Structured content built around specific, checkable claims still gets cited well after publication - the freshness signal favours it less over time, but a citation-worthy claim doesn't stop being true because the page is a year old. What decays is how often a model reaches for it relative to something newer saying the same thing. Update the page occasionally and the decay slows without needing a rewrite.
Isn't this just Google's "freshness" ranking factor by another name?
Related, not identical. Google's long-standing freshness signal - Query Deserves Freshness - applies mostly to queries where recency genuinely matters: news, trending topics, anything with rising search interest. The bias in the study above showed up across ordinary comparisons where nothing about the query itself demanded a recent answer. It's a broader, less selective version of the same instinct.
How often should I update an article to stay cited?
There's no single number that holds across every platform and topic, but the half-life data above gives a working range: weeks, not months. Check in on a citation-worthy article periodically, confirm the claims still hold, and update the page when something's moved on. Leave it untouched for a year and it's competing against everything published since.