Content freshness and AI citations: what "last updated" is really doing

Yes, but less than the headline numbers suggest, and not in the direction most people assume. Cited pages skew fresh by update date rather than publish date. The freshness gap against ordinary search results is around 25%, not a cliff. And the pages cited month after month are consistently less fresh than the ones that appear once and vanish. Recency buys a spike. Maintenance buys a habit.

The primary source is Seer Interactive's 2026 analysis: 7,683 pages carrying 47,097 citations, across four brands in pet retail, vacation rentals, retail energy and commercial banking, observed on ChatGPT, Gemini and Perplexity between March and June 2026. Only pages cited three or more times in non-branded answers were counted (Seer Interactive).

Two headline numbers:

• 75% of cited pages had been updated within the past year
• 88% had been updated within two years

Note the verb. Both are about the last modified date. Seer pulled dates from structured signals, meaning schema, sitemaps and HTTP headers, and could only date about two thirds of the pages they collected.

The split they published next is the interesting part. Among the 4,124 pages where both dates were readable:

And 27 to 28% of cited pages were what Seer calls "fresh from old": updated recently, but originally published two or more years ago. A page from 2019 that was properly rewritten in March sits comfortably inside the 88%.

So the freshness engines appear to reward is being manufactured by maintenance, not by new publishing. That is genuinely useful, and it is the opposite of the "publish more, publish faster" conclusion people draw from the same number.

Here is the problem nobody quoting this study mentions. Seer only looked at pages that got cited. There is no sample of uncited pages to compare against.

"75% of cited pages were updated in the past year" is only evidence that freshness matters if uncited pages were updated less often. The study cannot tell you that, because it never looked at any. If 75% of all indexed commercial pages get updated within a year anyway, the finding says nothing at all.

Ahrefs ran the comparison Seer did not, using organic search results as the control group. Across 16.975 million cited URLs on seven platforms (Ahrefs):

Measured on last updated rather than publish date, the gap narrows to 13.1%. And Google AI Overviews cites content 16 days older than organic results, which is the wrong direction entirely.

So: a real preference, worth roughly a quarter, not a wall. The average page cited by an AI assistant is nearly three years old. Anyone telling you content has a shelf life of twelve months is describing a tendency as though it were a rule.

Worth being straight about one more thing. Ahrefs finds the freshness gap larger on publish date than on update date. Seer's data implies the reverse. Two vendor studies, two different reads, both from companies with an interest in the answer. Treat the direction as solid and the size as unsettled.

No, and this is the finding worth building a content plan around. Seer split cited pages by how many of the four study months they appeared in, and the pattern inverts what you would expect.

The freshest pages are the most disposable. Pages last touched a couple of months ago flare up and disappear. Pages sitting at around six months since their last meaningful update are the ones that hold their position month after month.

It makes sense once you say it plainly. A very recent update is often a news hook or a seasonal refresh, and it wins the answer while it is topical. A page cited in all four months is one the engines have settled on as the reference for a question, and settling takes time.

The practical read: chasing recency for its own sake buys visibility with a short half-life. If you want to be the page an engine returns to, you want one that has been around long enough to be trusted and is maintained often enough not to look abandoned. Those are different jobs, and most refresh calendars only do the first.

The bar moves a long way depending on what the page is. From the same dataset, share of cited pages updated within a year:

And by engine:

By sector, retail energy ran hottest at 80% and pet retail coolest at 69%, with commercial banking at 75% and travel at 72%. Claude, Copilot and AI Mode were not in this study.

If you sell into a comparison-heavy category, or you are chasing Gemini, your maintenance bar is genuinely higher. If you publish evergreen guides in a slow-moving sector, a two-year-old page that is still correct is not the liability someone is about to sell you a subscription to fix.

This is where most advice goes vague, so here it is engine by engine.

Google. Uses the lastmod value in your sitemap, but only "if it's consistently and verifiably accurate", checked against the actual last modification of the page. The value should reflect a significant update to the main content, structured data or links. An update to the copyright date is explicitly not significant (Google Search Central). Schema dateModified and the visible byline date are additional signals rather than the source of truth, and Google may ignore all of them if they disagree with each other (Search Engine Land).

Bing and Copilot. Bing says accurate lastmod values help it prioritise updated content for crawling, and recommends complete sitemaps plus IndexNow. IndexNow currently has seven participants: Bing, Yandex, Seznam, Naver, Yep, Internet Archive and AmazonBot (indexnow.org). Google is not one of them. Neither are OpenAI, Perplexity or Anthropic.

ChatGPT. This one surprises people. ChatGPT serves pages from a shared cache, converted from HTML to Markdown at fetch time. Cached copies have been observed served more than 90 days after the original fetch, with no eviction cap, and the recrawl schedule is driven by one thing: how often users ask about the page (Search Engine Land). Your dateModified does not refresh you there. Demand does.

Perplexity, Claude and Gemini. All operate documented crawlers, and none publish a mechanism for telling them a page has changed. No submission endpoint, no IndexNow support, no equivalent of Search Console.

So for exactly one of the major engines is your last-modified date a lever you can pull directly, and even there it is a crawl prioritisation hint rather than a ranking input.

Neither, on the current evidence.

On schema, Ahrefs ran a matched difference-in-differences on 1,885 pages that added JSON-LD between August 2025 and March 2026, against 4,000 control pages matched on prior citation levels (Ahrefs):

The correlation everyone repeats is in the same study: 53% of AI-cited pages carry JSON-LD, roughly three times the rate of uncited pages. That is confounded, because sites investing in schema invest in everything else too. In a companion test, none of five major AI systems parsed JSON-LD when fetching pages live. The full null result and its caveats are in how brands actually earn citations in AI answers.

On dates alone, Google's John Mueller has been about as blunt as Google gets: changing the date without doing anything else is just noise and useless. Bumping lastmod on an unchanged page does not make it fresh, and because Google only trusts lastmod when it is verifiably accurate, doing it at scale degrades a signal you might actually want later.

There is a narrower argument for schema that does hold. A machine-readable dateModified is how anything, Google's crawler and Seer's researchers alike, can tell your page was updated at all. Seer failed to date a third of the pages they collected. If your date is not readable, you are invisible to the mechanism rather than judged badly by it. That is an infrastructure argument, not a ranking one, and it is worth about ten minutes of a developer's time rather than a project.

Google's own definition is the most useful one available: a significant change to the main content, the structured data, or the links on the page. That is a low bar to clear honestly and an impossible one to fake at volume.

An update with a chance of changing your citation position looks like:

• Replacing figures that have moved, and dating them
• Adding sections that answer sub-questions the page currently skips, which is what engines actually retrieve against
• Rewriting the answer paragraph under each heading so it stands alone
• Removing claims that have stopped being true, which is the part everyone skips
• Updating internal links to newer pages in the same cluster

An update that will not do anything: the date, the copyright year, a synonym swap, or an AI-generated paragraph appended to the bottom.

Refresh first, most of the time, and the reason is arithmetic rather than ideology. A page that already exists has already been crawled, already sits in somebody's index, and may already be in the consideration set for the question. A new page starts at zero on all three. The 27 to 28% "fresh from old" share of cited pages is the visible result: a large minority of everything AI engines cite is old content somebody maintained.

The exception is coverage. If the question you are losing is not one your page attempts to answer, no amount of updating will win it, and you need the new page. Refreshing is cheaper per point when the gap is quality. Publishing is unavoidable when the gap is coverage.

Which of the two you are facing is not really a judgement call. It is visible in the answers themselves: cited but ranked below a competitor is a quality gap, absent from every answer in a topic is a coverage gap. Answyn's briefs are built against whichever one the data shows, rather than against a publishing calendar.

Graded on the AEO Evidence Grades we use throughout the handbook:

That last row is worth a word. The half-life claim comes from a 194,077-source analysis that inferred content age from a year appearing in the URL, where 82.6% of sources carried no year at all. It is an interesting signal and a poor measurement, and it is the number most likely to be quoted at you.

Figures on this page come from published third-party research rather than Answyn's own tracking data, and each is attributed with its sample size. Two of the three main sources are companies selling AI visibility software, and they disagree with each other in places, which we have flagged where it happens.

If you find a figure that has moved, tell us at team@answyn.com and we will correct it and date the correction. This page is reviewed quarterly.