Methodology
We ran 1,000 real searches across nine industries. Most of them are the kind of search you make when you're picking a product or a provider.
For each search we recorded what Google's AI Overview cited, then asked the same question to Perplexity, Claude, Gemini and ChatGPT and recorded what each of them cited. We also captured the date Google printed beside every result.
Then we compared the cited pages against the pages that ranked in Google's top 30 for the same search but didn't get cited, and read every date we could find on each page: the visible dates and their labels, schema datePublished and dateModified, the published and modified meta tags, and <time datetime> tags.
One thing to keep in mind the whole way through. We read the dates a page carries, and we never compared its content against an older copy.
So when this report says a page was updated, it means the page says so. A real rewrite and a changed dateModified look exactly the same to us.
How to read the numbers
- "% more citations" is the lift. So 27 citations per 100 pages instead of 9 is 186% more. The "out of 100" numbers are always next to it.
- Every comparison is within the same search. Some searches are just easier to get cited in, so we only ever compare pages competing for the same one.
- "Held in 40 of 40 re-tests" means we split the searches in half at random, ran the check on each half, and repeated that 40 times.
- "The displayed date" is the date Google printed beside the page in its results for that search, including relative ones like "3 days ago".
- We show the gap two ways. Rank free holds page length and site authority equal and lets rank vary, so it includes any harm that comes through ranking lower. Rank fixed holds rank equal too, which shows the part that isn't explained by rank.
- ChatGPT's numbers are tiny throughout. It cited about 2 pages per 100 here, because it mostly cites pages outside Google's top 30, so small differences in its column are just noise.
Finding 1: If Google can see a published date, that's the date it shows
This one surprised us most, and it's the easiest thing here to go and fix.
We took every page that was first published more than 12 months ago and claims an update within the last 12 months, where Google displayed a date. Then we grouped them by how the page shows its dates to a reader.
| How the page shows its dates | Google displayed the fresh updated date |
|---|---|
| A visible "Updated" date only | 87% |
| Dates in schema or meta tags only, nothing visible | 40% |
| A visible "Published" date only | 20% |
| A visible "Published" and a visible "Updated" date | 13% |
Read the bottom row again. You updated the page, you showed both dates because that's the honest thing to do, and Google showed the old one anyway, 87 times out of 100.
And the top row is the same kind of page. Published just as long ago, updated just as recently. The only difference is that the published date isn't sitting on the page where a reader can see it, and Google shows the fresh date on 87 of every 100.
So schema on its own doesn't rescue you either. Pages that kept their dates in schema and meta tags with nothing visible got the fresh date shown 40% of the time, which is better than showing a published date and worse than showing an updated one.
Google reads the page like a person does
Here's how often Google showed each kind of date exactly as written, across every page carrying one:
| Where the date lives | Google displayed exactly that date |
|---|---|
| Visible in the page text | 92% |
<time datetime> tag | 87% |
article:published_time meta tag | 74% |
Schema datePublished | 67% |
Schema dateModified | 52% |
article:modified_time or og:updated_time | 46% |
So the date a reader can see wins, and the modified fields lose. Whatever your markup says, Google is mostly reading what's printed on the page, and on a normal article that's the published date.
And it pays to print something. Pages with a visible date backed by matching schema got a date shown 83% of the time, against 12% for pages with no date anywhere.
Finding 2: The date Google displays is what moves citations
Pages showing a date from the last 12 months got nearly three times the AI Overview citations of pages showing one over two years old.
Comparing the two ends, within the same search:
| AI engine | Displayed within 12 months | Displayed over 2 years ago | Lift | Odds, rank free | Odds, rank fixed |
|---|---|---|---|---|---|
| Google AI Overviews | 26.8 in 100 | 9.3 in 100 | +186% | 3.30 | 2.38 |
| Claude | 22.0 in 100 | 8.7 in 100 | +153% | 2.68 | 2.30 |
| Gemini | 14.6 in 100 | 7.6 in 100 | +91% | 1.72 | 1.45 |
| Perplexity | 22.1 in 100 | 13.2 in 100 | +68% | 1.70 | 1.33 |
| ChatGPT | 1.7 in 100 | 0.9 in 100 | +97% | 2.53 | 2.62 |
Google and Claude both held in 40 of 40 re-tests, with rank free and with rank fixed.
Anything inside the last 12 months does well, and three to six months is the sweet spot on Google and Claude. Past two years, citations fall off a cliff.
And it's a stronger signal than the dates written into the page. Run the same comparison on the page's own dates and you get +146% on Google, against +186% for the date Google picked to show.
A stale date costs you rank as well
Within the same search, pages showing a date over two years old rank 3.3 positions lower than pages showing one from the last 12 months, at 17.8 against 14.5 across 579 searches.
That's why the two odds columns differ. Rank free gives Google 3.30, and holding rank equal brings it down to 2.38. So a stale date costs you citations directly, and then costs you again through the lower ranking that comes with it.
Finding 3: The same page, cited twice as often, because of which date showed
Take those updated pages again. Google displayed the fresh date on 59% of them and the old published date on the rest. Here's how the two groups did:
| Perplexity | Claude | Gemini | ||
|---|---|---|---|---|
| Cited per 100, fresh date displayed | 22.1 | 25.5 | 20.6 | 14.2 |
| Cited per 100, old date displayed | 9.0 | 10.6 | 11.5 | 9.3 |
| Lift | +147% | +139% | +78% | +52% |
These pages all claim an update inside the last 12 months. The only thing separating the two groups is which date Google put beside the result, and the gap is 147% on Google.
They rank differently too: 14.4 against 16.7, so 2.3 positions.
There's a caveat here, and it's a real one. Every engine points the same way and the rank-free result is statistically clear on all four, but this group is smaller and it didn't hold across the 40 re-tests on Google or Claude. So treat it as a strong signal rather than a settled one.
And we can't tell cause from this either. Google might show the old date on pages where the update was thin, rather than the date itself holding the page back. Either way, getting your update shown goes with better results, and Finding 1 says how to do it.
Finding 4: No date at all
44% of the pages in Google's top 30 had no date displayed beside them.
| No displayed date, compared with | Claude | Gemini | Perplexity | |
|---|---|---|---|---|
| a date displayed within 12 months | -25% | -38% | -36% | -20% |
| a date displayed over 2 years ago | +107% | +68% | +26% | +38% |
So a fresh date is best, no date is second, and a stale date is worst.
The no-date advantage over a stale date works mostly through rank. It holds with rank free and disappears with rank fixed, because pages showing a stale date rank lower, and that's most of the gap.
On rank, no date is neutral. Those pages rank about the same as pages showing a fresh date, at 14.4 against 14.7.
Finding 5: What a recent update date is worth
Pages claiming a recent update got cited more too, and this is the one place where what we measured and what you'd want to know come apart.
Among pages that publish a date, within the same search:
| AI engine | Update date 2025 or later | Older | Lift |
|---|---|---|---|
| Google AI Overviews | 22.3 in 100 | 10.3 in 100 | +117% |
| Claude | 19.2 in 100 | 7.5 in 100 | +158% |
| Gemini | 13.6 in 100 | 6.9 in 100 | +98% |
| Perplexity | 19.7 in 100 | 13.2 in 100 | +49% |
| ChatGPT | 1.6 in 100 | 1.1 in 100 | +47% |
Google held in all 40 re-tests here, and this is the one finding we committed to before we collected any data.
But here's what we can't tell you. We read the dates on the page and never compared the content against an older copy, so a page that was really rewritten and a page where somebody changed dateModified and nothing else are the same page to us.
So read this as a recent update date going with more citations. Whether the rewrite does the work or the date does, this data can't separate, and anyone telling you it can from numbers like these is guessing.
What we'd still do is both, and Finding 1 is the reason. The date only helps you if Google shows it, and Google shows what's printed on the page.
Finding 6: The year in your title is a freshness signal too
A year in the title tag is the strongest title signal we've found. You get 192% more Claude citations, 115% more Gemini and 48% more Perplexity. It works because the assistants put a year into their own searches, so your title matches what they typed.
On Google the year does very little on its own, at +20% with no re-tests held. A stale year still costs you there: a current year beats an older one by 113% on Google and 252% on Claude.
And the year isn't just standing in for a newer page. Hold how recently the page was updated constant and it keeps almost all of its power, so the two signals stack.
So the assistants reward the year in your title, and Google rewards the fresh date it shows. The title side is covered in full in Title Tags Can Increase AI Citations By 192%.
What to do
- Show a visible "Updated" date, and take the visible "Published" date off. Keep the original date in schema
datePublishedso the record stays honest, and setdateModifiedto the update. That setup got the fresh date shown 87% of the time. Showing both got it shown 13% of the time. - Go and check what Google actually shows for your important pages. Just search for them. If Google is showing an old date on a page you updated, that page is losing citations and rank right now.
- Don't let the displayed date on an important page pass two years. That's where citations fall off hardest.
- Refresh the pages that matter at least once a year. Three to six months was the sweet spot, but anything inside the year did well, so you don't need a monthly refresh cycle for this.
- If a page can't be kept current, no date beats a stale one. But a fresh date beats no date, so this is the fallback and not the plan.
- Put the current year in the title of pages that date, and change it when you genuinely update the page.
Then check it on your own pages. Take a group of updated pages where Google is showing the old date, switch half of them to a visible "Updated" date only, leave the rest alone, and compare what Google shows a month later.
Limitations
This shows patterns, so it can't prove cause and effect. Pages with fresh displayed dates might differ in other ways we didn't measure. The rank and authority checks narrow that down, and the experiment above is what settles it for your own pages.
We read dates, not content. Nothing here verifies that an updated page actually changed. Every "updated" page in this study is a page claiming an update, so a genuine refresh and a bumped dateModified are indistinguishable in the data. That limit sits hardest on Finding 5.
The displayed date is Google's, for that search. Relative dates like "3 days ago" were converted to an approximate date. Pages with no date shown count as "no displayed date", whether or not the page itself carries one.
AI answers move between runs. Asking the same engines the same questions hours apart, 49% to 65% of the cited pages came up again.
This covers part of what each engine cites, the part that also ranks in Google's top 30: 40% of Google's citations, 33% of Claude's, 29% of Perplexity's, 21% of Gemini's and 8% of ChatGPT's.
We only tested US, English, desktop results, collected September 15, 2026. Every age is measured from that day.
Scope: 1,000 searches across nine industries, 9,795 distinct pages, Google AI Overviews, Perplexity, Claude, Gemini and ChatGPT, September 2026. Every headline pattern re-tested on 40 random half-splits of the searches.
