Half of AI-cited content is under 13 weeks old

Roughly half the pages cited in AI answers were published or substantially updated within the previous 13 weeks. That single pattern reframes what a content refresh is for. It stopped being tidy-up work and became the cheapest way to stay visible in the answer layer.

The premium got stronger after Gemini 3 became the default model behind AI Overviews in January 2026. Not because the model prefers new things aesthetically, but because of how the retrieval step works.

The mechanism, not the correlation

Google's own May 2026 guide on generative AI features names two things that explain this: retrieval-augmented generation and query fan-out.

When someone asks a question, the system does not answer from what the model memorised during training. It fires a set of concurrent related queries, retrieves live pages, and grounds the answer in what it pulls back. That retrieval step runs against the current index, with the current freshness signals applied.

So the freshness premium is not a ranking factor bolted onto AI answers. It is inherited from the retrieval layer that feeds them. Anything that makes a page more likely to be retrieved for a fan-out query makes it more likely to be cited.

That distinction matters, because it tells you what a refresh must change. Retrieval is matching on the content of the page. Not the date stamp on it.

The refresh that does nothing

Changing dateModified in your schema and republishing is not a refresh. Neither is swapping 2025 for 2026 in the title and adding a paragraph at the top.

Both are visible. The page content hash barely moves, the substantive text is identical, and you have taught your own team that the job is done. Sites that run this pattern at scale end up with a library of pages that all claim to be current and none of which are.

A refresh qualifies when the substance changed. Concretely, at least one of these:

• A number, threshold or price in the piece is now different, and you have updated it and said what it used to be.
• A platform behaviour described in the piece changed, and you have flagged the change rather than quietly rewriting history.
• You have added a section covering something that did not exist when you first published.
• You have cut a section that is no longer true. Deletion is a refresh. It is also the one nobody does.

The flagging point is worth sitting with. A page that says this threshold was X until March 2026 and is now Y is more citable than one that only states Y, because it answers a comparison question as well as a factual one, and comparison questions are exactly what fan-out generates.

Which pages to refresh

Refreshing everything on a 13-week cycle is not possible for anyone with a real library, and it is not the right target anyway. Prioritise in this order.

1. 1.Pages already being cited. Check which URLs receive AI referrals in your custom channel grouping, and which pages appear in AI Overviews for your tracked queries. These have proven retrievable. Keeping them current protects a position you already hold.
2. 2.Pages ranking positions 5 to 20 on queries that trigger AI Overviews. These are close enough that improved substance can move them into the cited set, and the CTR difference between cited and uncited is the whole game.
3. 3.Pages covering anything with a date, a rate, a threshold or a platform in it. These decay on a schedule you can predict.
4. 4.Everything else, on an annual sweep, mostly to delete.

Stable reference content does not need this. A page explaining how a physical process works, or what a term means, does not gain from quarterly churn, and the freshness premium is weak on queries where the answer has not changed since 2019. Do not put a structural explainer on a 13-week cycle to satisfy a calendar.

The trap in the data

Be honest about what the 13-week figure can and cannot support. Newer pages are also more likely to target newer topics, and newer topics have thinner competition. Some of the freshness premium is survivorship, not preference. A brand new page about a January 2026 platform change gets cited because it is one of four pages that exist on the subject, not because it is fresh.

Which means the tactic is not publish more, faster. It is cover the change early, then keep the page honest.

Volume publishing on a 13-week cycle produces the exact content type Google's documentation warns against, and burns the budget you need for the pages that actually earn citations.

A schedule that works

For a team with an existing library of 100 to 400 pages, this is the version that survives contact with reality:

• Monthly: the 10 to 15 pages that carry numbers, prices, platform behaviours or regulatory detail. Owner named per page, not per batch.
• Quarterly: the pages already cited in AI answers, plus everything sitting at positions 5 to 20 on AI Overview queries.
• Twice a year: a deletion pass. Anything with no sessions, no citations and no strategic reason to exist comes down. Consolidate two thin pages into one good one and redirect.
• Never: the date-stamp-only update. Ban it explicitly, because it will creep back in the first quarter somebody is behind.

Budget it as a fixed share of content capacity rather than as spare time. Something like a third of the content hours going to refresh is defensible for a library over 100 pages. Below that, new coverage still wins.

What this costs

The real cost is not writing time. It is the research to know what changed, which means somebody has to actually be following the platforms, the regulators and the market rather than skimming a newsletter. That person is expensive and cannot be replaced by a tool that summarises other people's summaries.

The payoff is that your best pages stop quietly ageing out of the answer layer while your rankings look fine.

Refreshing is not maintenance. It is the distribution strategy now.