Content Freshness: How Often AI Drops Old Pages

A page that ranked well on Google for three years can disappear from ChatGPT's answers within weeks of going quiet. This is one of the sharpest differences between classic SEO and GEO, and it catches most content teams off guard, since Google rewarded evergreen content for over a decade while AI search engines are already treating the same pages as expired.

This guide covers exactly how fast AI platforms drop old content, why they behave this way, and the specific refresh schedule that keeps a page inside the citation window. For the broader mechanics of how models pick sources at all, see our guide on how LLM models decide which brands to mention.

Facts, figures and platform behavior in this article were checked in September 2026. Freshness research moves fast, so treat the numbers below as a current snapshot rather than a fixed rule.

The clearest evidence comes from Ahrefs, which analyzed 17 million citations across AI platforms and found that AI cited content is 25.7 percent fresher on average than content ranking organically on Google for the same query. On average, ChatGPT cites URLs that are 393 to 458 days newer than what ranks organically on Google for that same search.

That gap explains a pattern many brands notice but cannot name. A page can sit at position one on Google for years while quietly falling out of every AI answer on the same topic, simply because retrieval systems are comparing it against a much younger pool of competing pages.

Not every AI engine ages content at the same speed. Research from Quattr breaks content's citation lifecycle into five stages, spike, discovery and growth, peak, decay, and dormancy, with peak citation typically landing within 30 to 90 days of publication and near total disappearance by the one year mark without a refresh.

Source: Quattr, The Content Decay Cycle for AI Citation, 2026

A citation half life is the point at which a page has lost roughly half of its peak citation rate. ChatGPT's 3.4 week half life means a page can lose half its citation strength in under a month if nothing changes on it, which is a completely different maintenance rhythm than traditional SEO, where a well ranked page can coast for a year or more.

Separate research from Gander adds a longer view. Content loses roughly half of its total AI citation potential within twelve months of publication, and by the time a page is around five years old, it operates at only about 18 percent of its peak visibility. Rank.bot's tracking found an even steeper curve at the start: citation rates can fall from about 2 percent at peak right after publishing to as low as 0.2 percent within six months without a meaningful update.

This is not just a theory. Several independent studies converge on the same conclusion using different datasets.

• Pages updated within the last two months earn roughly 28 percent more AI citations than older pages on the same topic, based on Semrush data reanalyzed by Superlines in March 2026.
• Content under 30 days old earns an estimated 3.2 times more AI citations than older content, according to ConvertMate's analysis of more than 80 million citations.
• Pages not updated at least once every 13 weeks are three times more likely to lose their AI citations entirely, according to Quattr's 2026 research, and this holds even for genuinely evergreen topics.
• About half of all AI search citations now come from content less than 13 weeks old, per Amsive's 2026 citation freshness analysis.
• Content updated every 90 to 120 days maintains search rankings roughly 4.2 positions higher than static content, and earns a 47 percent higher click through rate on time sensitive keywords.

The consistent theme across all five data points is the same. Freshness is not a minor ranking bonus in AI search. It functions as a primary trust signal that AI systems use as a cheap proxy for accuracy, since retrieval pipelines re-rank sources on every single query rather than relying on a static index the way traditional search does.

Large language models are trained specifically to avoid presenting outdated or hallucinated information, and the cheapest, most reliable way to reduce that risk during retrieval is to lean toward content that looks and reads as current. A three year old article about pricing or software features is statistically far more likely to contain deprecated details than one updated last month, so models learn to treat age itself as a mild risk signal, independent of whether the specific page is actually still accurate.

This shows up directly in how models phrase their own searches. AI systems frequently add the current year to their internal retrieval queries even when the user never typed a year at all. Research from Gander tracking Q1 2026 retrieval behavior found that among URLs containing a year in the path, the drop off is steep and immediate: content from one year prior loses about 42 percent of its share compared to current year content, and the year before that loses another 43 percent on top of that.

Half life measures how fast citations decay, but each platform also applies different rules for what counts as fresh enough in the first place.

This is why a single republish date is not a universal fix. A page optimized purely for Perplexity's tight 30 day window may still perform fine on ChatGPT's more forgiving multi year window, but assuming one platform's rules apply everywhere is a common and costly mistake, one we cover in more general terms in our ChatGPT Search optimisation guide.

The single most common freshness mistake is updating a visible publish date without changing anything meaningful inside the content itself. AI systems are increasingly able to detect this gap. If a page displays a 2026 date but still lists 2024 pricing or references a product version that has since been discontinued, the inconsistency itself becomes a negative signal, sometimes described as a failure of entity freshness, since the system can tell the surface level date does not match the substance underneath it.

A genuine refresh needs to touch the actual facts: updated statistics, current pricing, corrected product details, and new examples, not just a new timestamp. The strongest freshness signals line up together: a visible on page date, a matching dateModified value in structured data, an updated XML sitemap lastmod entry, and real content changes that a crawler can verify. Submitting the updated URL through IndexNow, so Bing and the platforms built on it can see the change within hours rather than weeks, is a detail worth pairing with any refresh, a step we go into further in our guide on GEO vs SEO: What's Actually Different in 2026.

Not every page needs the same cadence. Blanket monthly rewrites waste effort on content that does not need it, while treating a comparison page as evergreen guarantees it will fall out of citation within weeks.

Prioritize the pages closest to actually ranking or being cited first, rather than starting an update sweep from your oldest content. A page that is already showing up occasionally in AI answers is usually a faster win than one that has never been cited at all.

The pattern shows up repeatedly in client audits: a brand assumes a well written page from a year or two ago is still doing its job, right up until it is tested directly against ChatGPT, Perplexity and Gemini and simply does not appear anymore, even for the exact keywords it still ranks for on Google. The fix is rarely a full rewrite. It is closer to a targeted pass that updates the specific numbers, dates and claims inside a page, refreshes the schema and sitemap signals, and gets the change indexed quickly.

Treating a republish as a rewrite. Changing the visible date without updating the facts inside the page is easy for AI systems to detect and can hurt more than doing nothing.

Applying one cadence to every page. A definitional explainer and a pricing page do not decay at the same speed, and treating them identically wastes effort in one direction or leaves citations on the table in the other.

Forgetting the technical signals. A meaningful content update that never gets picked up by a fresh sitemap lastmod entry or submitted through IndexNow can sit unseen for weeks, quietly losing the citation window it was meant to win back.

Refreshing once and stopping. Since ChatGPT's citation half life is measured in weeks, not months, a single refresh followed by a year of silence only delays the decay, it does not prevent it.

Our guide on tracking your brand across ChatGPT, Perplexity and Gemini walks through setting up this kind of ongoing, per platform monitoring.

How old is too old for a page to still get cited by AI?

It depends heavily on the platform and the topic. Perplexity's preference tightens sharply after about 30 days for time sensitive topics, while ChatGPT tolerates a broader window of two to three years for general subject matter. As a general rule, treat anything untouched for more than 13 weeks as at risk.

Does updating the date field alone help at all?

Very little, and it can actively hurt if the surrounding facts have not changed. AI systems can detect a mismatch between a fresh date and stale content inside the page.

Do I need to refresh every single page on my site?

No. Prioritize pages that are time sensitive, already receiving some AI citations, or covering a highly competitive topic. Purely conceptual, non time sensitive explainer content can run on a much slower six month cadence.

Is content freshness more important than the quality of the writing itself?

No, they work together. Freshness gets a page back into the pool of sources a model is willing to consider. Quality, clarity and factual accuracy are still what earns the citation once it is in that pool.

If you want to see which of your pages have already fallen out of AI citation due to freshness decay, you can check it directly.