Seer Interactive measured 47,097 citations from ChatGPT, Gemini, and Perplexity against the dates on the pages behind them, and found that 75% of cited pages had been updated within the past year (7,683 dated pages, March–June 2026). Read the dates the other way, by original publish date, and that share falls to 42%. The pages models cite are mostly old pages kept alive, and we have looked before at what those citations are actually worth.
Table of contents
Key takeaways
- 75% of cited pages were last updated within one year (7,683 dated pages, March–June 2026); 88% within two years.
- Freshness is manufactured primarily through updates to existing pages: more than a quarter of the “fresh” pages in the study were first published over two years ago.
- Publish date and update date are not interchangeable signals. Pages that look fresh by update date drop from 72% to 42% when measured by original publish date instead.
- The three engines diverge: Gemini cited 78% fresh content, ChatGPT 73%, Perplexity 65% — and the gap traces to content-type mix, not a blanket freshness preference.
- Pages cited every month across the four-month window were less recently updated (median ~six months) than one-month spikes (median ~two months), which suggests durable AI visibility is a maintenance problem, not a publishing-volume problem.
What this is about
Freshness in AI citations is manufactured through maintenance, not publishing. Across four verticals, about a quarter of what looks current by update stamp dates back two years or more. For anyone planning content work, that moves the budget from new articles toward a refresh queue for pages already earning citations.
The design
The dataset covers four client accounts in unrelated categories: pet retail, short-term rental, energy supply, and commercial banking. Coverage spanned three engines — ChatGPT, Gemini, and Perplexity — across a shared four-month window from March through June 2026.
The unit of analysis was unique conversations, deduplicated, keeping only pages cited three or more times. Last-modified dates were recovered from structured signals: schema markup, sitemaps, and HTTP headers. Dates were recoverable for roughly two-thirds of cited pages, yielding 7,683 pages carrying 47,097 citations. Every figure below refers to that dated subset only, counting each page once.
Retail energy was substantially larger in the dataset than the other three verticals — a function of collection volume for that client, not category importance. Seer reports percentages rather than raw counts when comparing across engines or industries, and checks headline figures both pooled and equal-weighted to guard against that imbalance.
No control group exists in the traditional sense. This is an observational study of citation patterns, not a controlled experiment. Causal claims about recency causing citation are not supported by the design; association is what the data shows.
The numbers
Overall freshness, 7,683 dated pages (March–June 2026)
| Update age | Share of cited pages |
|---|---|
| ≤1 year | 75% |
| ≤2 years | 88% |
Seer describes the 2–3 year band as a small residual and everything beyond three years as near zero, without publishing figures for those bands.
Source: Seer Interactive, 47,097 citations across ChatGPT, Gemini, Perplexity, March–June 2026
Within that one-year band, the majority had been revised inside the previous quarter.
Publish date vs. update date, 4,124 pages with both dates recoverable
| Measurement basis | Share ≤1 year | Share 1–2 years | Share 2–3 years |
|---|---|---|---|
| By last update | 72% | 15% | 10% |
| By publish date | 42% | 19% | 16% |
Roughly one page in four that reads as current by its update stamp is in fact an article from two or more years back.
Freshness by engine, 7,683 dated pages
| Engine | Updated ≤1 yr | Updated ≤2 yr | Top cited content types | Fresh-from-old* |
|---|---|---|---|---|
| Gemini | 78% | 90% | Marketplaces, comparison/reviews | 27% |
| ChatGPT | 73% | 87% | Blogs, guides, brand pages | 28% |
| Perplexity | 65% | 83% | Blogs, guides, older reference | not reported |
“Fresh-from-old”: pages updated within the last year but originally published two or more years ago.
Freshness by vertical, pages with both dates recoverable
| Vertical | Updated ≤1 yr | Published ≤1 yr | Gap |
|---|---|---|---|
| Retail energy | ~72% | ~37% | +35 pts |
| Pet retail | 67% | 44% | +24 pts |
| Commercial banking | 68% | 45% | +23 pts |
| Travel | 69% | 47% | +22 pts |
Values marked ~ are read from Seer’s charts; the study does not publish exact figures for those cells.
Citation consistency vs. freshness, 7,683 dated pages
| Months cited | Pages | Share of dated set | Median update age | Share updated ≤1 yr |
|---|---|---|---|---|
| All 4 (always-on) | 3,423 | 45% | ~0.47 yr (~6 mo) | 68% |
| 3 months | 1,986 | 26% | ~0.33 yr (~4 mo) | 77% |
| 2 months | 1,749 | 23% | ~0.26 yr (~3 mo) | not reported |
| 1 month (spike) | 525 | 7% | ~0.16 yr (~2 mo) | 86% |
What could still be wrong
Date recovery rate. Dates were recoverable for roughly two-thirds of cited pages. The undated third is not a random sample — sites with cleaner structured markup are more likely to surface a date, which may skew the dated set toward more professionally maintained properties. If older or less-maintained pages are systematically missing from the dated set, the freshness figures are overstated.
Vertical size imbalance. Retail energy is substantially larger in the dataset than the other three categories. Seer uses percentages and equal-weighting to compensate, but the pooled figures still reflect collection decisions, not equal market representation. The engine-level freshness numbers in particular should be read as “what Seer collected” before they are read as “what each engine does globally.”
Four-month window. March through June 2026 is a single season. Seasonal query patterns — travel planning, tax-related banking queries, energy-switching cycles — could inflate or deflate freshness figures for specific verticals in ways that a full-year window would smooth out.
Correlation, not causation. The study observes which pages get cited and how recently they were updated. It cannot isolate recency as the cause of citation from other correlated attributes: domain authority, content depth, structured data quality, or the simple fact that high-traffic pages tend to get updated more often. A page that is updated frequently may earn citations for reasons other than the update itself.
Engine behavior changes. Any shift in how one or more models weighted recency signals during the March–June 2026 window would appear in the data as a real effect when it is actually a measurement artifact.
Segment mix within verticals. Seer reports that Gemini’s freshness figure is heavily influenced by its skew toward energy citations (62% of Gemini’s sample) and toward marketplace content within that vertical. A different client mix would produce different engine-level numbers.
What to do on Monday
The study’s operational implication is specific enough to act on without waiting for a controlled test.
First, separate your update date from your publish date in whatever tracking you use. If your citation monitoring or Search Console work treats these as interchangeable, you are measuring the wrong thing. A page published several years ago and refreshed recently should be categorized as recently updated, not as old content.
Second, pull your own always-on pages. If you have citation tracking running, identify the URLs that show up across multiple months — not the spikes. Check when those pages were last substantively updated. Seer’s data suggests that consistently cited pages have a median update age around six months; if yours are running older, that is the refresh queue to start with.
Third, sort your content library by type before setting any refresh cadence. Comparison pages and marketplace listings are cited freshest and are mostly earned placements you cannot directly update. Your blogs, guides, and brand pages are the owned levers. A blanket refresh schedule applied equally to both is wasted effort on one side and insufficient pressure on the other.
Fourth, if you manage partnerships or earned placements, add update frequency to your vetting criteria. Seer’s data indicates that roughly 98% of cited pages are third-party. A publisher who lets articles age out is a worse bet for sustained AI visibility than one who maintains their content, independent of domain authority.
For evergreen pages you are considering leaving alone, three questions are sufficient: Have the underlying facts changed? Is the page still appearing in citation tracking at its historical rate? Would a model citing an outdated version of this page produce a factually wrong answer? If all three are clean, leave it alone and move to something that needs the attention.
FAQ
Does this study prove that updating a page causes it to get cited more often?
No. The Seer Interactive study is observational — it records which pages were cited and how recently they were updated, but the design has no control condition. Frequently updated pages may earn citations for correlated reasons: they tend to be on higher-authority domains, they tend to carry cleaner structured data, and high-traffic pages are more likely to receive regular maintenance regardless of any AI-visibility strategy. The association between recency and citation is real in the data; causation is not established.
The three engines show different freshness figures. Which one should I optimize for?
The engine-level gap in the Seer data traces primarily to content-type mix, not to each engine having a different built-in freshness preference. Gemini’s 78% figure is partly explained by its heavy citation of marketplace and comparison content, which are structurally the freshest page types. Perplexity’s 65% reflects a greater pull toward older reference material. Optimizing for a specific engine by adjusting update frequency is probably the wrong frame, the same trap as the GEO tactics that do not survive testing; optimizing by content type — keeping comparison pages and brand pages current regardless of which engine you are targeting — is more durable and applies across all three.
How do I run a version of this check on my own site without a dedicated citation tracking tool?
A basic version is possible with data you likely already have. Pull your sitemap’s curl -I command returns this). Do this monthly for two or three months and you will have enough to identify your own always-on pages. The threshold Seer found — always-on pages with a median update age around six months — gives you a benchmark to compare against.
