Content Freshness

1. What content freshness is

Content freshness is the property that decides whether a page’s age and update history count for or against it when an engine retrieves a document and grounds an answer on it.

The word hides two properties that behave very differently:

• Recency — the timestamp. How long since the page was published or last changed. A machine-readable fact.
• Currency — the substance. Whether the claims the page makes are still true. A semantic fact that no timestamp can assert.

That distinction carries the rest of this entry. Every anti-pattern in §7 is an attempt to buy recency without paying for currency, and every legitimate tactic in §6 is the reverse: fix currency, and let recency follow as a side effect.

Freshness is also not a single-step signal. It acts twice in the answer loop: once at retrieval, where the question is whether the changed version of your page is in the index at all — a recrawl problem — and again at grounding and reranking, where the question is which of several candidate passages gets selected. Treating freshness purely as a publishing-cadence issue misses the first half, which is why pages that were genuinely updated sometimes behave as though they weren’t. Freshness is one signal inside GEO, not a strategy on its own.

2. Recency vs currency — and what each one can prove

The asymmetry drives everything downstream: recency is publisher-assertable, currency is not. A publisher can set dateModified to any value. No publisher can set “still true”.

Engines therefore treat stated dates as claims to be corroborated rather than as facts. That is why the failures in §7 are trust-shaped rather than simply ineffective. Currency is the “accuracy, transparency, currency” leg of Trustworthiness in E-E-A-T §4.4; it is orthogonal to whether the page is structurally liftable at all, which is citability’s question. A current page still has to be extractable, and an extractable page still has to be right.

3. What the evidence actually says — and what it does not

Two independent bodies of evidence bear on “AI systems prefer recent content”. One is observational and one is counterfactual, and they support different claims.

3.1 The correlational evidence — AI cites fresher, and “fresher” is older than you think

The largest public dataset is Ahrefs’ analysis of 16.975 million cited URLs across ChatGPT, Perplexity, Gemini, Copilot, AI Overviews and organic Google results (Law & Guan, July 2025). The headline: AI-cited URLs are 25.7% “fresher” than organic results.

The number underneath the headline is the one worth carrying. That 25.7% is the gap between an average cited-URL age of ~1,064 days and an average organic age of ~1,432 days. The average AI-cited page is roughly 2.9 years old.

So fresher here is a comparative against a very old baseline, not a synonym for new. Advice of the form “publish weekly or lose your citations” does not follow from this data — at the aggregate level it is contradicted by it. The study’s per-engine ordering is consistent, though: ChatGPT skews newest of the major assistants, and Google AI Overviews behaves essentially like organic search rather than like the other AI surfaces.

One honesty note. This is observational data. Fresher pages may be cited more because they are fresher — or because newer pages also tend to be better structured, better linked, and written about whatever is currently being asked. Correlation cannot separate those. That is what §3.2 is for.

3.2 The causal evidence — recency bias is in the model, not just the corpus

Fang, Tao, Chen, Chang and Sakai (2025) supply the counterfactual companion, and it is the strongest single piece of evidence here.

The method is what makes it load-bearing: the authors prepend artificial publication dates to passages from the TREC Deep Learning collections (DL21 and DL22) and re-run LLM reranking. The passage text never changes. Only the stated date moves. Any change in ranking is therefore caused by the date and nothing else.

There is an uncomfortable corollary, and it is better named than left implicit: the bias attaches to the stated date, not to any observed change in the content. Mechanically, that is an exploit. §7 explains why running it is still a bad trade.

3.3 The bounded reading

A sourcing caution specific to this topic: its secondary literature is unusually saturated with vendor marketing quoting other vendors’ numbers. Three source classes are load-bearing — original-data studies, peer-reviewed or preprint research, and first-party engine documentation. Everything else is worth reading for direction and nothing else.

And one adjacent finding, since it bounds how much recency can ever buy: when a model chooses between conflicting sources, topical relevance dominates, while stylistic credibility markers barely move it (Wan et al., ACL 2024). Freshness is a feature inside that preference, not a substitute for topical fit.

4. Freshness is query-conditional

§3.1 says the average cited page is ~2.9 years old. Vendor studies report that half of AI citations are under 13 weeks. Both can be honestly measured, because freshness weight is a function of the query, and the two samples asked different questions.

The mechanism is inherited from classical search and survives into generative retrieval: recency is weighted by how much the answer is expected to have changed since it was written. Google names it directly. Its ranking-systems guide lists Freshness systems as a category and describes “various ‘query deserves freshness’ systems designed to show fresher content for queries where it would be expected” (A Guide to Google Search Ranking Systems). The phrase to hold onto is where it would be expected — the weighting is conditional by design, not a blanket preference for new pages.

The practical consequence cuts both ways. A definitional page rewritten monthly buys almost nothing and exposes itself to §7’s failure modes. A pricing page left unrevised for a year is losing on currency no matter how well it is structured.

Freshness sensitivity also varies by surface:

These are directional. The underlying per-platform day-counts come from a single vendor study with an undisclosed query mix, so they should not be treated as stable coefficients — the same discipline citability §5 applies to the GEO benchmark numbers.

5. What actually signals freshness to an engine

Given that freshness matters, what is the engine reading? Five channels — and the publisher controls only three.

The load-bearing point: the two channels the publisher does not control are the ones that decide whether the three it does control are believed.

This is documented, not inferred. Google states it uses the sitemap lastmod value “if it’s consistently and verifiably (for example by comparing to the last modification of the page) accurate”, and that a significant update means a change to main content, structured data, or links — explicitly not a copyright-date bump (Build and Submit a Sitemap). On the byline side, the guidance is to label a prominent visible date, mark it up with datePublished / dateModified, keep the visible and structured values consistent, and avoid both future dates and the date of the event being described (Add a Byline Date).

Worth stating plainly, because much freshness advice implies otherwise: Google’s own AI-features documentation does not describe freshness as a requirement, and is explicit that “there are no additional requirements to appear in AI Overviews or AI Mode, nor other special optimizations necessary”, with no special schema.org markup needed (AI features and your website; Optimizing for generative AI features). Date markup is a disambiguation aid — it helps an engine choose correctly when a page carries several dates — not a freshness lever.

6. Decay triggers and the update cadence model

6.1 The five decay triggers

A page needs revision when one of these fires — not when a calendar interval elapses.

Triggers two and three are the ones you cannot detect by reading your own page; they need the measurement layer in GEO metrics.

6.2 What counts as an update

6.3 The cadence, assembled

Set a review interval by query class, using §4’s table. Revise on trigger, using §6.1’s.

The interval exists to force a look, not to force an edit. This is the distinction most cadence advice collapses, and collapsing it is what produces the date-bump habit: if a quarterly review is treated as a quarterly update, the only way to satisfy it on a page where nothing has drifted is to change the date and nothing else. A review that concludes “nothing has drifted” is a successful review, and it should end with no edit at all.

Entries on this site carry an internal nextReviewDue field for exactly that purpose — it schedules the look, and most reviews end without a change to lastUpdated.

7. Anti-patterns — the date-bump trap

§3.2 showed that injected dates move rerankers. That makes this section load-bearing rather than decorative: the exploit is real, and the reason not to run it has to be stated rather than assumed.

The first row is not a GEO Wiki inference. Google lists it among its own examples of search-engine-first content, as a self-assessment question publishers are told to ask themselves: “Are you changing the date of pages to make them seem fresh when the content has not substantially changed?” (Creating Helpful, Reliable, People-First Content).

The load-bearing line: freshness is earned by currency, and recency is its receipt — not the other way round.

One honest caveat, because the opposite claim would be easy to disprove and would discredit the rest. The reason not to bump dates is not that it never works — it demonstrably moves lab rerankers. It is that the gain is small, temporary, and asymmetric against a trust penalty that is neither. The Fang et al. result is best read as a description of a bias that engines have an active interest in mitigating, not as a durable lever; and deliberate date-signal manipulation is a recognised spam pattern in its own right — see GEO Spam and Manipulation. The single-actor picture also degrades once competitors optimize the same surface (C-SEO Bench, Puerto et al., NeurIPS ‘25 D&B).

8. Why this matters for GEO + how to act

Freshness is the cheapest grounding signal to fake and one of the more expensive to actually hold — which is precisely what makes holding it worth something. It is also the only major signal that degrades passively. Citability and E-E-A-T decay slowly if at all; a page’s currency decays on a clock somebody else is setting.

References

Academic:

• Fang, H., Tao, S., Chen, N., Chang, K.-X. & Sakai, T. (2025). Do Large Language Models Favor Recent Content? A Study on Recency Bias in LLM-Based Reranking. SIGIR-AP ‘25. arXiv:2509.11353 · ACM DL · paper summary
• Wan, A., Wallace, E. & Klein, D. (2024). What Evidence Do Language Models Find Convincing? ACL 2024 Main. arXiv:2402.11782 · paper summary
• Puerto, H., Gubri, M., Green, C., Oh, S. J. & Yun, S. (2025). C-SEO Bench: Does Conversational SEO Work? NeurIPS ‘25 Datasets & Benchmarks. arXiv:2506.11097

Original-data studies:

• Law, R. & Guan, X. (2025). New Study: AI Assistants Prefer to Cite ‘Fresher’ Content (17 Million Citations Analyzed). Ahrefs

Official platform documentation (as of 2026-07):

• Google Search Central — Add a Byline Date to Google Search Results · Help Google Search know the best date for your web page
• Google Search Central — Build and Submit a Sitemap
• Google Search Central — AI features and your website · Optimizing your website for generative AI features on Google Search
• Google Search Central — A Guide to Google Search Ranking Systems (the “Freshness systems” entry) · Creating Helpful, Reliable, People-First Content