In Seer’s March-June 2026 sample of 4,124 cited pages with both dates available, 72% were updated within a year but only 42% were first published that recently. The sample covered four brands on ChatGPT, Gemini and Perplexity. This separates maintenance from new publishing; it does not prove that an edit caused more citations.
A content refresh should leave the reader with a more accurate or more complete answer. The research now lets us distinguish a page's original age from how recently it was maintained. It still does not give us a universal number of days after which content stops earning citations.
Newer evidence: modified recently does not mean published recently
Parse's page-age study examined 8,215,946 citations from ChatGPT Search and Google AI Mode between May 24 and August 28, 2026. Only 116,379 citations passed every Last-Modified reliability filter, about 1.4% of the total. Within that subset, the median reported modification age was three days.
That result adds a different measure from the published-age studies below. A five-year-old guide updated yesterday can have an old publication date and a recent modification date. A server can also emit an unreliable timestamp without any meaningful edit, which is why the study excludes suspicious headers.
For your own page, keep three dates straight: when it first appeared, when its substance changed and when the cited measurement was collected. A subtle line such as “Updated October 6 · Data collected May-July 2026” tells a reader more than replacing the year in the title.
What an age difference actually measures
Ahrefs' July 2025 analysis found an average publication age of 1,064 days across ChatGPT, Perplexity, Copilot and Gemini citations, compared with 1,432 days in its organic Google comparison. That is an observed difference between selected results. It does not identify a refresh threshold or isolate the effect of editing a page.
Ask which fact on your page may have changed. A current price matters to a pricing question. An accurate explanation of a stable concept may need no update at all. Neither case is well served by a blanket requirement to publish something new each month.
What the newer 2026 evidence adds
Seer's July 2026 follow-up examined March-June answers for four brands. It dated 7,683 cited pages; 75% had been updated within a year. The subset with both dates is the 4,124-page comparison in the opening.
An old URL can contain current evidence. Keep that distinction when deciding whether to maintain an existing page or publish another.
How the measured age mix differs by engine
Seer's dated-page sample, March-June 2026:
| Engine | Updated within a year | Within two years |
|---|---|---|
| Gemini | 78% | 90% |
| ChatGPT | 73% | 87% |
| Perplexity | 65% | 83% |
Content and category mixes differ. This is not a universal ordering of freshness preference.
Why cosmetic date changes do not work
A date change alone does not repair an obsolete price, a discontinued feature or an unsupported claim. Keep the publication date and the substantive update date distinct. If a research page now explains its method more clearly but still describes May-July data, label that as an editorial update and keep the measurement period visible.
We do not have evidence for a universal percentage of text that must change before AI will recognize a refresh. Rewriting 30% of an accurate article can be less useful than correcting one consequential sentence. Save the previous version and record the facts changed so the update can be reviewed.
How to decide what to refresh
Refresh budgets are finite, and a published library of 200-plus pages cannot all be touched in the same quarter. Parse tracks AI visibility across ChatGPT and Google AI Mode, and the consistent pattern in that data is that 80% of a brand's AI citations come from 15% to 20% of its pages. The shortlist that earns the refresh budget should be built from three signals stacked together: pages that already earn AI citations and could lose them to a fresher competitor (and AI citations churn from run to run even without a competitor's refresh); pages that rank in the top 10 organically on queries that now trigger AI Overviews; and pages whose factual content is anchored to a date, version, regulation, or product capability. A page that fails all three (a definitional explainer for an evergreen concept, no AI citations, no AI Overview overlap) does not need a quarterly refresh. A page that hits all three (a product comparison earning citations on a query with active AI Overviews) almost certainly does.
A practical refresh cadence by content type
Use review triggers that match the claim's lifetime. This is an editorial workflow, not a measured ranking schedule.
| Content | Trigger for review | Substantive update |
|---|---|---|
| Pricing | Plan or billing change | New amount, billing period and effective date |
| Product comparison | Release affecting the tradeoff | Revised capability table with source links |
| Original research | Comparable new observations | New sample, period, method and downloadable data |
| Implementation guide | Changed interface or documented behavior | Tested steps and current screenshots |
| Evergreen explanation | New evidence or a reader's unresolved question | Clearer example or corrected argument |
Set reminders to inspect high-value pages, but publish changes when there is something useful to change. Retain an older research result when it remains valid for its historical window; add a separate follow-up when the new sample is not directly comparable.
The minimum viable refresh program
Choose a manageable set of pages that already answer important buying questions. For each, record the claim to verify, its current source, the last substantive change and the next review trigger. Work through factual errors and missing explanations before adding length.
An example: a comparison says a plan costs $20 per user per month. If the price changes, update the figure, billing basis and linked source wherever the comparison relies on it. If the price is unchanged but readers keep asking whether contractors count as seats, add that explanation with a source. Both edits serve the reader; neither requires rewriting the rest of the article.
How to measure whether a refresh worked
Record repeated answers to a fixed set of questions before the edit, then repeat the same questions on the same engines afterward. Keep the page URL, answer text, citation presence and collection dates. Use comparable unedited pages as a reference for changes affecting the whole panel.
Report the number of eligible answers as well as the citation rate. A move from two citations in five answers to four in twenty is more citations but a lower appearance rate. Do not attribute the movement to the edit unless you have ruled out obvious changes in questions, crawl access and engine behavior. No movement may reflect limited exposure or ordinary variation; it does not prove the update lacked substance.
What freshness does not fix
A current date cannot supply missing evidence. Before refreshing a page, check whether it answers the intended question, supports important claims, and exposes its useful content in accessible text. A comparison table helps when readers need to compare actual dimensions; a table added only for extraction adds little.
Keep definitions and methods stable when they remain correct. Expand them with a worked example, a boundary case or an explanation of how to apply the result. The goal is a more useful answer, with an honest date attached to it.
Related Parse research: How long does AI take to notice a website change?.
If you want to see which sources shape AI answers about your brand, run a free brand check.