TL;DR
- Some AI citation studies show a preference for newer pages, but that does not prove that changing a date improves visibility.
- In one 81-page Semrush test, Google AI Mode cited more pages on day one; ChatGPT search cited more on day 30. These are results from one site.
- Fix stale facts that affect buyer decisions first. Keep a sound explanation when its content is still accurate.
- Measure first citation, continued citation and answer accuracy separately. Early pickup is only one part of the result.
Three different freshness questions
Three questions often get folded into one: do AI systems cite newer pages, how soon can they cite a new page, and does updating an existing page improve its results? The first two have useful published evidence. The third requires a different kind of test.
What a new-page test can tell us
Semrush published 81 FAQ-style pages on its own blog in September 2025 and tracked related queries for 30 days. On day one, Google AI Mode cited 29 pages, while ChatGPT search cited 8. By day 30, those counts were 21 and 34.
| Platform | Pages cited on day 1 | Pages cited on day 30 |
|---|---|---|
| Google AI Mode | 29 | 21 |
| ChatGPT search | 8 | 34 |
Source: Semrush, December 2025. Counts on each observation day, from 81 pages on one established SEO site. They are not cumulative totals or evidence that the same pages stayed cited throughout.
The useful lesson is about evaluation windows. A first-day result and a month-end result can lead to different conclusions. Fast pickup did not produce the higher ending count in this test. It would be a mistake to turn either pattern into a guaranteed schedule for your site.
The experiment concerns new pages. It does not compare old pages that received a real update with similar pages left unchanged. It therefore cannot tell a B2B team how much lift a refresh will produce.
One-site Semrush publication test, September 2025. Counts are not cumulative totals or proof that the same pages stayed cited. Semrush, December 2025.
Newer cited pages do not establish a refresh effect
Ahrefs's July 2025 analysis used 16.975 million cited URLs. Its platform table lists average publication ages of 958 days for ChatGPT's citation group and 1,432 days for Google's top-three AI Overview citations. Publication age used the date Ahrefs first saw a URL, rather than a guaranteed record of its first publication.
These are average ages across different source sets. The result does not show what happens when the same page gets updated. Topic mix, source mix and retrieval choices could help explain why the averages differ. A site with many frequently changing pages also faces a different task from a specialist supplier with a stable reference library.
Notice the timescale as well. Even the lower of those two averages spans years. A finding that one platform's cited pages were newer on average does not establish a short expiry date for all content.
Put the studies side by side
| Study | Design and selected result | Question it addresses | What it does not establish |
|---|---|---|---|
| Ahrefs, July 2025 | URL-age analysis; ChatGPT citation group 958 days, top-three AI Overview citations 1,432 days | How old were the pages cited in those source sets? | The causal effect of editing an existing page |
| Semrush, December 2025 | One-site publication test; day-30 counts were 21 and 34 of 81 pages | How did pickup differ across two platforms in this test? | A universal waiting period or a refresh return |
Both companies sell search tools and use their own data systems. The studies add distinct evidence, but neither is a randomized test of content updates. Read together, they support platform-specific observation and patient measurement. They leave the size of a refresh effect open.
Review the fact that can go stale
For a B2B team, a more useful starting point is the claim that could become wrong. A page may contain a stable definition and a fast-changing price. Those parts should not have the same review trigger.
| Content or claim | Reason to review it | Useful change | Evidence that the work is complete |
|---|---|---|---|
| Price or plan limit | Approved commercial terms change | Correct the value and its conditions | Page matches the current approved offer |
| Product compatibility | A release adds or removes support | Update version and support details | Product owner confirms the supported setup |
| Certificate or test result | Validity or scope changes | Replace the expired evidence and explain scope | Link and claim match the current document |
| Research statistic | A new study changes the answer | Compare methods before replacing the figure | The new conclusion follows the cited study |
| Stable definition | The field or explanation changes | Correct the specific gap | A reader can still get an accurate answer |
This is SemShift's editorial decision aid. The review triggers are tied to facts, rather than a claimed platform refresh formula.
For example, a fictional equipment supplier finds that a guide recommends a component no longer sold. The useful update replaces that advice, explains the current supported choice and checks linked specification sheets. Changing only the date would leave the buyer with the same wrong answer.
If the rest of the guide remains sound, it need not be rebuilt to make that correction. Keep a record of what changed so the next reviewer can see which facts were checked.
Dates should describe the work
Google's byline-date guidance says it uses several signals to estimate publication or significant-update dates. It asks for consistent visible and structured dates. This explains how to describe an update; it is not a promise of a citation gain.
Our recommendation is to reserve an updated date for real changes and keep an internal review record when a check finds no change needed. A review date and a substantial update date serve different purposes. Neither should make old facts look newly verified when they were not checked.
Content Freshness and AI Citations: Common Questions
How often should we review B2B content?
Set a schedule based on how quickly its claims can change and the cost of an error. Commercial terms and product support need different attention from a stable definition. No study cited here establishes one ideal interval for every page.
Is an old page a poor citation source?
Age alone does not answer that question. An older page may still contain the most relevant evidence, while a new page may repeat an error. Check the claim, its source and its current applicability.
Does the Semrush test mean we should wait exactly 30 days?
No. Thirty days was the study's observation window. It is useful evidence that results can change during a month, not a service-level promise from either platform.
How would we test whether a refresh helps?
Define the outcome first, then compare updated pages with suitable unchanged pages where possible. Keep buyer questions, platforms and observation methods stable, and record other changes. Without a fair comparison, report the result as an observation rather than a causal lift.
Discuss outdated answers
If outdated facts keep appearing in AI answers, bring a few examples and the pages they cite to SemShift. We can discuss which claims need correction and how to measure the result through an AI visibility audit. The starting point is an accurate answer with traceable evidence.