The Freshness Illusion: Why AI Cites Edited Pages, Not New Ones

What did Seer actually measure?

Sonny Vasquez, an SEO manager at Seer, pulled last-modified dates from structured signals — schema, sitemaps, HTTP headers — for pages cited in non-branded answers across three engines, for four client brands in four categories: a national pet retailer, a vacation rental marketplace, a retail energy provider and a commercial bank. The window was March to June 2026. Only pages cited three or more times were counted, and roughly two-thirds of candidate pages could be successfully dated, yielding the 7,683-page set.

Two things about that design are worth stating before any number gets quoted. This is four agency client accounts, not a random or stratified sample of the web, and Seer is a marketing agency publishing an insights piece — not peer-reviewed research. Vasquez is candid about the unevenness: “I lean on percentages, not raw counts, when I compare across engines or industries.” We would apply exactly the same discount here that we apply to vendor data from Ahrefs, Profound or SE Ranking. The direction of the finding is more trustworthy than its precise magnitude.

Is “fresh” the same thing as “new”?

No, and this is the whole story. Taken at face value, the headline numbers look like a straightforward endorsement of publishing velocity: 75% of all cited pages had been updated within the last year, and 88% within the last two. Read only that, and the obvious conclusion is that answer engines want a steady stream of new material.

The subset with both dates readable kills that reading. Among those 4,124 pages, 72% appeared fresh by update date while only 42% had been published in the last year — and more than a quarter of the pages that looked fresh were originally published more than two years ago. The pages being cited are, disproportionately, old pages that someone went back and touched.

“The freshness LLMs reward is being manufactured by updates, not by new publishing,” as Vasquez puts it. For anyone budgeting content, that is a materially different instruction than “publish more.” It points the money at maintenance of what already ranks and gets retrieved, not at net-new production.

Does freshness buy durable citations, or just spikes?

Here is the part of the dataset we would put ahead of the headline, because it complicates the tidy version of the advice.

Seer split cited pages by how many of the four study months they appeared in. The pattern runs backwards from what you would expect:

Cited in	Median page age	Updated within last year
All 4 months	5.6 months	68%
3 of 4 months	3.96 months	77%
2 of 4 months	3.12 months	82%
1 month (spike)	1.9 months	86%

The pages cited most consistently are the oldest and the least recently updated. The pages that look freshest are the ones that flare for a single month and vanish. Recency correlates with getting cited; it correlates negatively with staying cited.

That is not an argument against maintenance. It is an argument against reading a citation spike as a win. If your measurement window is one month and you refresh a batch of pages, you will see exactly the bump you were hoping for, and it will tell you very little about whether those pages are still being cited in November. Seer frames the single-month bucket as short-term spikes, and we would not stretch it further than that.

Where does this hold, and where does it not?

The effect is not uniform, and two breakdowns matter for anyone deciding where to spend.

By engine, Gemini’s cited pages were the freshest (78% updated within the last year, 90% within two years), then ChatGPT (73% and 87%), then Perplexity (65% and 83%). Perplexity is the most willing of the three to cite something that has not been touched recently.

By content type, the spread is wider. Marketplace pages were 78% freshly updated and comparison and review pages 77% — but News and Editorial pages were the least fresh of any category at 45%. An engine citing a news page is often reaching for something that has sat untouched since it was written, which is what you would expect from reference material. By vertical, retail energy pages led at 80%, then commercial banking at 75%, travel at 72% and pet retail at 69%.

One boundary is worth stating plainly because it is easy to miss: this study covers ChatGPT, Gemini and Perplexity only. It says nothing about Google AI Overviews or AI Mode. “AI search rewards fresh content” is not a claim this data licenses about Google’s surfaces, and we have written before about how differently those behave.

How much of this can you even control?

The uncomfortable number in the study is the one least likely to make it into a LinkedIn summary. Roughly 2% of all cited content in the dataset was owned, first-party brand content. About 98% was third-party — marketplaces, comparison sites, review platforms and other earned placements.

The distribution backs it up: blogs and guides made up 50% of cited pages, comparison and reviews 20%, marketplaces 13%, and brand and corporate pages just 6%.

So the freshness lever is real, and for roughly 98% of the surface that gets cited about you, it is not yours to pull. You can maintain your own pages. You cannot go and update a marketplace listing or a review roundup on someone else’s domain. This is the same structural problem behind ghost citations: the thing being cited is frequently not the thing you own, which makes “optimize your content for AI” an incomplete strategy no matter how well you execute it.

What follows from this?

Three things we would act on, stated at the confidence the data supports.

Move budget from production to maintenance. If old-but-edited pages are being cited at the rate this study suggests, a refresh cycle on pages that already get retrieved is a better use of money than another net-new post. This is the study’s clearest practical implication.

Make your update signals machine-readable. Seer read freshness from schema, sitemaps and headers. If your CMS does not emit a truthful dateModified, you are invisible to the signal being measured — regardless of how much you actually revise.

Do not confuse a refresh with a rewrite for its own sake. Vasquez’s summary is “Publish and forget loses. Publish and maintain wins.” Note what that does not say. It does not say that editing a page causes citations.

That distinction is the one caveat we would not let anyone drop. This is observational data: it matches citation snapshots against page timestamps. Nobody updated a controlled set of pages and measured the before and after. The finding is entirely consistent with a causal story where refreshing content earns citations — and equally consistent with a world where pages that are actively maintained are simply the pages that were already good enough to be cited. Correlation is what was measured, and correlation is what should be quoted.

The honest version of the takeaway is narrower than the headline and still worth having: among pages that answer engines cite, being recently edited is common, being recently published is not, and the pages that hold their citations longest are older than the ones that spike.