Content freshness isn't a quality signal to an LLM, it's a date the machine reacts to. What the research shows, why nobody refreshes, and which pages to update first.
Key Takeaways
- Researchers changed nothing about a passage except the publication date stuck to the front of it, then re-ran retrieval across seven models. Fresh passages got promoted every time. Content freshness isn’t a quality signal to an LLM. It’s a number the machine reacts to.
- Ahrefs analyzed 17 million AI citations and found AI-cited URLs run about 25.7% fresher than top-10 organic results. They also found cited pages still average 2.9 years old, which means freshness amplifies authority rather than replacing it.
- Content decay in AI search runs on a different clock than Google. A page can hold position 4 for three years and be missing from AI answers the entire time, and your ranking report won’t tell you.
- Refreshing has always lost to publishing because refreshing burns senior hours and returns nothing you can put in a slide. That’s a cost problem, not a value problem, and cost is the part that just changed.
AI crawlers show a preference to content that was more recently published, even when the content remains the same.
Researchers at Waseda University did exactly this and published the results at SIGIR-AP 2025 in a paper called Do Large Language Models Favor Recent Content? They took passages from the TREC Deep Learning collections, prepended artificial publication dates, and ran the retrieval again across GPT-3.5-turbo, GPT-4o, GPT-4, LLaMA-3 at 8B and 70B, and Qwen-2.5 at 7B and 72B. The passages were identical. The only variable was the date.
Every model promoted the fresher-looking passages.
What does that mean for your website, which likely has multiple pages that were published several years ago? Heck, even content published several months ago looks stale to an LLM at this stage.
All of this is giving rise to “content freshness seo”, a new strategy that focuses on keeping great content ranked and cited.
What we know about content receny and GEO (Generative Engine Optimization)
The Waseda result doesn’t stand alone. Ahrefs’ analysis of 17 million AI citations, published in April 2026. AI-cited URLs averaged 1,064 days old. Top-10 organic Google URLs averaged 1,432 days. That’s a gap of roughly 25.7%, and it points the same direction as the lab work. AI assistants reach for newer things than Google does.
Now the part most people quoting that study leave out.
Ryan Law, who ran it, published a list of caveats alongside it, and they’re inconvenient. The average age of a cited page is still 2.9 years, so AI assistants clearly aren’t allergic to old content. Google is the engine least influenced by freshness, and Google is still where most searches happen. And his conclusion is blunt, which is that most brands will probably get better results from creating new, high-quality content than from pouring resources into refreshing what they have.
The most credible source on this topic partly disagrees with the article's premise. Worth saying so out loud, because you’ll find it the moment you click the link.
Here’s where the two findings actually meet. Freshness doesn’t beat authority, it compounds with it. A page that’s ranked for two years has backlinks, indexed history, and a track record of relevance to its topic. A new URL has none of that and has to earn it. Update the old page and you keep every one of those signals while resetting the clock the model is watching. Publish a new one and you start at zero on everything except the date.
That’s why refreshing outdated content tends to return faster than the same effort spent on a new asset. Not because the writing is better, but because the date at the top of the page really is doing a lot of heavy lifting. In the eyes of an LLM, at least.
Why content freshness is so hard to stay on top of
Everyone in SEO agrees updating old content works. Almost nobody does it consistently.
Because it’s boring.
Nobody joined a marketing team to swap a 2023 statistic for a 2026 one across 40 posts.
Another issue is, the success is hard to attribute. Did it start ranking again because Tom updated it, or because Jessie initially wrote it? Sounds messy already.
It also takes time and effort to tell which pages actually need refreshing, and how. If you have 400 published posts, finding the twelve with dead statistics or a reference to a tool that got acquired last year means scanning all 400. That’s a lot of hours, or a lot of AI credits!
What a genuine refresh changes
A refresh needs to doa few things in order to look fresh to Google and AI crawlers. Including:
- Dated statistics swapped for current-year sources, with the source name sitting next to the claim rather than buried in a link at the bottom.
- Tool names, pricing, and version numbers corrected. Half the SaaS content published in 2024 references products that have since been renamed, acquired, or shut down.
- New sections covering developments that postdate the original. If your piece on AI coding predates agentic PR review, it isn’t stale, it’s incomplete.
- Terminology brought current, because the words your audience searches with move faster than your content does.
- dateModified, sitemap lastmod, and the visible last-updated line all matching the edit that actually happened.
That last one sounds like housekeeping. It’s the part that makes the other four legible to a crawler.
How Fimo’s automated content freshness agent works
Fimo runs a purpose-built agent against your live site whose only job is finding and fixing decay.
It reads your library and flags staleness triggers. Statistics past an age threshold you set. References to tools, versions, or pricing that no longer match reality. Gaps between what a page covers and what’s happened in the topic since it was published. It doesn’t wait for you to hand it a list, because the “which of my 400 posts is stale” problem is the one that stops most teams before they start.
Then it proposes changes. It doesn’t make them.
Every edit the agent produces lands as a proposal in the same Git PR review flow that governs everything else on a Fimo site. Someone reads the diff. Someone approves it, rejects it, or edits it first. The agent works inside an isolated environment, so nothing it touches reaches production until a person signs off. If it swaps a statistic you don’t like, that costs you a rejected pull request, not a correction post.
You can configure what it watches for and how often it runs, which matters because a legal page and a comparison page decay at completely different speeds.
One thing worth being blunt about. None of this helps if AI crawlers can’t read the page in the first place. GPTBot, ClaudeBot, and PerplexityBot don’t execute JavaScript, so a client-side rendered site isn’t stale to them, it’s blank. Fimo renders server-side by default, which is the prerequisite that makes a freshness strategy mean anything at all.
The freshness agent is one of several autonomous agents running against a Fimo site, alongside SEO and translations, and you can build your own for jobs specific to your business. That’s the autonomous website idea in practice. The site maintains itself. People keep the approvals.
Which pages to update first
Four filters, in order.
Existing authority plus declining impressions. A page with backlinks and a downward trend is the highest-return update on your site. You’re adding one missing signal to a page that has the rest. A page with no links and no traffic doesn’t need a refresh, it needs deleting.
Statistics older than 18 months. This is the cheapest and most mechanical filter, and it’s where content decay starts on almost every B2B blog. A 2023 stat in a 2026 article tells a model the whole page is 2023.
Dead references. Any page naming a tool that’s been renamed, acquired, or shut down. These are worse than stale, they’re wrong, and being wrong is the one thing retrieval systems are actively trying to avoid.
Ranking on Google, absent from AI answers. The signal almost nobody tracks and the clearest evidence of the two-clock problem. If a page holds a Google position and never appears in an AI response for the same query, age is the most likely culprit.
Run those four and the twelve pages worth touching fall out of four hundred in an afternoon. That’s the entire job that used to be the reason nobody did the job.
Content freshness is a maintenance budget, not a campaign
Most content teams are funded to publish. Almost none are funded to maintain. You can see it in the org chart, where somebody owns the editorial calendar and nobody owns the archive.
That made sense when the archive kept working on its own. It doesn’t anymore, because the machine now reads a date before it reads a word, and it does that on every query, not just the ones where recency matters.
Publishing is how you get into the index. Updating is how you stay in the answer. Budget for one and you’re paying for the other in visibility you’ll never see leave.
Join the waitlist or book a demo and put the agent to work on the four hundred posts you’ve been meaning to get to.
FAQ
What is content decay?
Content decay is the gradual loss of traffic, rankings, and citations on a page that used to perform. Nothing breaks. The page just stops being the best available answer as competitors publish, facts change, and, in AI search, as the page’s age counts against it in retrieval.
How do you identify content decay?
Compare impressions and clicks over rolling 6-month windows in Search Console and look for pages trending down while the rest of the site holds. Then check AI citation presence separately, because a page can look stable in Google and have disappeared from AI answers entirely. Age plus declining impressions plus existing backlinks is the profile worth acting on.
How do you fix content decay?
Update the page rather than replacing it, so you keep the links and the indexed history. Swap dated statistics for current sources, correct anything factually out of date, add sections covering what’s happened since publication, and make sure dateModified reflects an edit that genuinely occurred. How to fix content decay comes down to changing the page enough that the new date is honest.
Content decay vs content pruning, what’s the difference?
Decay is the problem. Pruning is one response to it. You refresh pages worth saving, which means pages with links, rankings, or topical relevance. You prune pages that have none of those and are diluting your site’s focus. Most libraries need both, applied to different pages.
What’s the difference between a content refresh and a rewrite?
A refresh keeps the structure and argument, updating facts, examples, and gaps. A rewrite starts from the keyword again and rebuilds the piece on the same URL, usually because search intent moved or the original was thin. Refreshes take hours. Rewrites take days. Most pages need the first one.
How often should you refresh content for AI search?
It depends on how fast the underlying facts move, not on a calendar. Pages about pricing, tools, or fast-moving categories go stale in months. Conceptual explainers can hold for years. The useful trigger isn’t elapsed time, it’s whether anything on the page is now untrue or incomplete, which is exactly what an agent can watch for continuously and a human can’t.
Does changing the publish date improve rankings?
Not on its own, and Google’s John Mueller has warned about this specifically. The date has to correspond to a change that actually happened. What the Waseda research shows is that models react to recency signals, not that you can fabricate one and get away with it. The signal works when it’s true. Faking it is the fastest way to teach a crawler your dates mean nothing.
