“We Can’t Double The Content Budget, So What Then?”
Every conversation about AI search arrives at the same instruction: if you want to show up in the answers, publish more.
Almost nobody can. Headcount is what it was last year and so is the budget, so “publish more” is the one lever in the room that nobody can actually pull. I spend a lot of my week in exactly that room.
But someone went and checked the dates on the pages AI assistants are citing, and there’s a second door standing wide open next to the one everybody’s shoving at.
7,683 pages’ worth of dates. Let’s go look at the dates.
How New Are The Pages AI Actually Cites?
Three quarters of them are less than a year old. Of 7,683 cited pages, 75% had been updated within the past year — 47,097 citations, collected from ChatGPT, Gemini and Perplexity between March and June 2026 R.
The work is Seer Interactive’s, a US digital marketing agency.
They ran non-branded questions across four brands — a pet supplies retailer, a vacation rental company, a retail energy supplier and a commercial bank — and kept only the pages that got cited three or more times. Dates came from whatever the page itself declared, from the , and from the server’s own response headers.
A last-modified date is just a record of the last time anyone touched the page.
Seer also sells content-update work. A firm that sells refreshes telling you refreshes work is a baker recommending bread for breakfast (the underlying counts are first-party data, but the direction of the conclusion has some colour in it, and it’s worth holding that in view).
Almost A Third Of The “New” Pages Were Old Posts, Rewritten
This is the part that changes what you do on Monday. The authors took the pages updated within the past year and separated out the ones first published two or more years ago.
That rewritten group came to 27% on Gemini, 28% on ChatGPT and 28% on Perplexity. Across verticals it never dropped under 26%.
So more than one in four of the “new” pages AI is picking up isn’t a new article at all. It’s an old one somebody kept alive.
The same shape shows up from a second angle. On the 4,124 pages where both the publication date and the last-modified date were available:
- By last-modified date, 72% fell inside the past year
- By original publication date, 42% did
Same pages, one choice of date, 30 points of movement. That gap is the refreshed material, sitting on the new side of the line because somebody rewrote it.
The authors’ own framing is that the freshness LLMs reward is produced by updating, not by publishing. An LLM (large language model) is the machinery inside ChatGPT and Gemini, and what it’s reading is the current version of the page, whenever that version was first posted.
Which means moving the date on a page you already own puts you in the same pool as a brand-new article. If your production line is capped, that’s a far more realistic door.
Where To Start: The Pages That Haven’t Moved In Three Years
The distribution has a steep tail:
- Updated within one year: 75%
- Within two years: 88%
- Within three years: 93%
Three quarters inside a year, past ninety percent by three. Pages older than that have essentially fallen out of the distribution the citations come from.
So the priority orders itself. Start with whatever hasn’t been touched in three years and rewrite the substance.
Open your own sitemap, sort your main pages oldest-first by last-modified date, and you’ve got a shortlist by lunchtime.
There’s variation underneath, too. The share of citations coming from pages updated within a year was 78% for Gemini, 73% for ChatGPT and 65% for Perplexity — so Perplexity leaves 13 points more room for older material than Gemini does. Freshness isn’t one switch that all three assistants flip at the same voltage. It’s one of the conditions for being back inside the distribution at all.
How Far Can You Lean On These Numbers?
This is a correlation. Nobody randomised a set of pages, updated half of them and watched what happened, so the effect size of the update itself is out of reach here.
Last-modified dates were only recoverable for about two thirds of the set, and the other third could be anything. Retail energy is over-represented in the sample, the whole thing rests on four brands, and none of it is a measurement of your category.
Seer didn’t write a limitations section, so all of those reservations are mine rather than theirs.
The thing I’d most want you to avoid is the version where the displayed date moves and the words don’t. What this study is looking at is a set of pages whose dates moved because the body was rewritten. Touch up the timestamp on its own and the model still reads last year’s argument in this year’s clothes.
List Your Main Pages By Last-Modified Date Before You Plan New Ones
There’s less to do here than it sounds like. Pick the ten pages at the centre of the topics you want to be named in, put their last-modified dates in a list, and work down from the oldest — anything past three years first.
The second thing is a framing one. Treat updating your own pages as the part you can actually move, because the study also reports that only around 2% of cited content sits on brand-owned pages. The rest belongs to third parties, so your own refresh cycle was never going to fill the citation slot by itself. Move the half you control, and treat the third-party half as a separate campaign with a separate plan.
Maybe I’m reading too much into four brands and one agency’s dataset. What I’d still bet on is the sort: it costs an hour and it tells you which pages have quietly aged out.
The pull toward recent pages showed up in Seer’s 2025 numbers too, at 65% within the year — I went through those in ChatGPT mostly cites pages from the last twelve months. What this year’s work adds is the escape hatch: a rewrite gets you onto the new side of the line without a new article.
So before you fill in next quarter’s calendar, go and sort the old one.
