AI search engines do lean toward recently updated pages, but the date stamp is not the thing to change. Across 17 million citations, Ahrefs found AI assistants cite URLs that are on average 25.7% "fresher" than Google's organic results, and an Otterly test found that changing only the visible date on 10 articles lifted their AI citations 31.5% in 30 days. Google, meanwhile, lists "changing the date of pages to make them seem fresh when the content has not substantially changed" as a warning sign of search engine-first content. The approach that holds up under both facts: update pages when they are actually stale, change the substance, and then let every date signal tell the truth.
This guide gives you the evidence with its limits spelled out (sample sizes, controls and who sold what), Google's guidance quoted word for word, a definition of a substantive update, a refresh priority matrix, the exact technical signals to set (dateModified, visible dates, sitemap lastmod), and a measurement protocol that separates a real refresh effect from normal AI answer churn. Sources were checked on October 18, 2026.
Key takeaways
- Freshness is a real but modest tilt. AI-cited pages are newer than organic results on average, yet the typical cited page is still about 2.9 years old (Ahrefs, 2025). Google AI Overviews show almost no freshness preference over organic results.
- Date-only bumps "work" on some engines, short term. Otterly's 10-page test saw +31.5% citations, but two-thirds of the gain came from Claude, Google AI Mode fell 20% and Gemini fell 37%, and only two control pages were used.
- The likely mechanism is a known model bias. Researchers who injected fake dates into passages found LLM rerankers promote "fresh" text, shifting the top 10's mean publication year by up to 4.78 years. Exploiting a bias is not a durable strategy.
- Google's line is clear. Dates must describe a real publication or significant update; sitemap
lastmodis used only when it is "consistently and verifiably" accurate. - Measure against noise. In our own tracking, 42.9% of prompt-engine pairs flipped between mentioning and not mentioning the brand at least once. A refresh test needs a baseline, holdout pages and repeated runs.
Do AI search engines prefer fresh content?
Yes, on average, and the size of the preference depends heavily on the engine and the query. The best large-scale evidence is correlational: AI assistants cite newer pages than search engines rank, but they do not ignore old ones.
The Ahrefs study published July 28, 2025 analyzed 16.975 million cited URLs from its Brand Radar dataset. Measured by the date Ahrefs' crawler first saw each page, AI-cited URLs averaged 1,064 days old versus 1,432 days for organic Google results, which Ahrefs reports as 25.7% fresher. Measured by last-updated date, the gap shrinks to 13.1% (909 versus 1,047 days). ChatGPT showed the strongest preference for newer content; Google AI Overviews' top three citations were actually 16 days older than organic results on the publish-date measure. The author also notes that the average cited page is still about 2.9 years old.
Two caveats matter. First, "age" here is crawler discovery date, not a verified publish date. Second, it is a correlation: pages that are popular tend to get updated, and popular pages tend to get cited, so freshness and quality travel together.
Why do AI assistants favor recently updated pages?
Four mechanisms are documented or strongly supported, and they explain why the effect varies so much:
- Retrieval comes from a live index. Google says generative AI features use retrieval-augmented generation, "a technique (also known as grounding) used to improve the quality, accuracy, and freshness of AI responses by relying on our core Search ranking systems to retrieve relevant, up-to-date web pages from our Search index" (Google's AI optimization guide). Freshness is a stated goal of grounding, but the pages still come through core ranking.
- Some queries deserve freshness. Google's ranking systems guide describes "various 'query deserves freshness' systems designed to show fresher content for queries where it would be expected." A question about this year's prices triggers that; a question about how photosynthesis works does not.
- Fan-out queries often add the current year. When an engine expands a prompt into sub-searches (see how query fan-out works), recency-sensitive sub-queries tend to surface newer pages. Otterly proposes this as the mechanism behind its result but says it "did not capture the fan-out queries," so treat it as a hypothesis.
- Language models read dates and like recent ones. In a September 2025 paper by Fang, Tao, Chen, Chang and Sakai, researchers prepended artificial publication dates to passages from the TREC Deep Learning 2021 and 2022 collections and asked seven models (GPT-3.5-turbo, GPT-4, GPT-4o, LLaMA-3 8B and 70B, Qwen-2.5 7B and 72B) to rerank them. "Fresh" passages were consistently promoted: the mean publication year of the top 10 shifted forward by up to 4.78 years, individual passages moved by up to 95 ranks, and pairwise preferences between equally relevant passages reversed by up to 25%. Larger models reduced the effect, but none eliminated it.
That last finding is the key to the whole debate. If an engine's reranking step reads a date string and favors recent ones regardless of content, then changing the date alone can move citations. That is a bias the authors argue needs mitigation, not a signal of quality, which is why building a strategy on it is fragile.
Which AI engines care most about freshness?
The two sources that report per-engine numbers point in the same direction for ChatGPT and Perplexity and disagree or stay silent on others. Treat this as a directional map, not a ranking.
| Engine | Ahrefs 2025 (age of cited pages vs organic results) | Otterly 2026 date-only test (change in citations) |
|---|---|---|
| ChatGPT | Strongest preference for newer pages: citations 458 days newer, in-text references 393 days newer | 37 to 293 citations (+692%) |
| Perplexity | Newer than organic; newer sources tend to be cited first | +217% (+674 citations); no page declined |
| Claude | Not in the dataset | 1,135 to 2,118 (+87%) |
| Gemini | Newer than organic | -37% |
| Microsoft Copilot | Newer than organic | Not reported |
| Google AI Overviews | Top 3 citations 16 days older than organic | +15% (+177 citations) |
| Google AI Mode | Not in the dataset | -20% |
The pattern across both: ChatGPT, Perplexity and Claude responded to freshness most, while Google AI Overviews, AI Mode and Gemini responded least or went the other way. That fits Google's statement that its AI features retrieve pages through core Search ranking rather than a separate freshness-first system.
What the freshness experiments actually show
The published experiments are small, mostly run by vendors on their own sites, and honest about their limits when you read past the headline. Here is every study cited in this debate, with what it can and cannot tell you.
| Study | Publisher and status | Sample and design | Headline finding | What it cannot tell you |
|---|---|---|---|---|
| Date-only update test (Sep 2026) | Otterly, AI visibility vendor, own blog pages | 10 treated pages, 2 controls, 320 prompts, 7 engines, two 30-day windows | Citations +31.5%; controls -10.6% | Whether it lasts; why engines split; author calls it "directional rather than conclusive" |
| Year in title test (Jul 2026) | Otterly, own blog pages | 11 pages, 3 unmatched reference pages, 15-day windows | Citations +56% and +61% in two waves | Anything: the untouched reference page rose 64% |
| Freshness of cited URLs (Jul 2025) | Ahrefs, SEO tool vendor, own dataset | 16.975M cited URLs, observational | AI-cited URLs 25.7% fresher than organic | Causation; exact publish dates |
| Stale content and ChatGPT (Aug 2025) | AirOps, content workflow vendor | 4,000+ ChatGPT-cited pages, 900 queries, 15 industries | 35.2% of cited pages updated in the last 3 months | Causation; the report gives two conflicting figures for stale pages |
| Citation half-life (Sep 2026) | Profound, sells a citation decay product | 883,000 pages, 1.19M page-engine lifecycles, 7 engines, 1-in-128 sample | Median page loses half its citation share 11 days after peaking | Whether updates extend a page's life (not analyzed) |
| Recency bias in LLM reranking (Sep 2025) | Academic preprint on arXiv, no product to sell | 7 models, TREC DL21 and DL22, injected dates | Top-10 mean year shifted up to 4.78 years | How production engines weight dates today |
Does changing only the publish date increase AI citations?
In one small test, yes, on some engines, over 30 days. In Otterly's experiment (last updated September 24, 2026), the team changed the visible "last published" date on 10 existing articles to August 4, 2026, without touching the content, and compared July 5 to August 3 against August 5 to September 3 across 320 prompts on ChatGPT, Claude, Perplexity, Google AI Overviews, Google AI Mode, Gemini and Copilot. Citations to the updated pages rose from 4,800 to 6,312 (+31.5%) while two unchanged control pages fell from 358 to 320 (-10.6%). Prompts citing at least one updated page rose from 205 (64%) to 246 (76%).
Read the numbers underneath and the picture gets more specific:
- One engine did most of the work. Claude added 983 citations, which is 65% of the 1,512 net gain (our arithmetic from Otterly's figures). ChatGPT and Perplexity added 256 and 674.
- Google's AI surfaces barely moved or went down. AI Overviews rose 15%, AI Mode fell 20% and Gemini fell 37%.
- One page dominated volume. The highest-volume page went from 3,624 to 4,778 citations, so it was about three-quarters of the baseline. Excluding it, the other nine rose 30.4%, and 7 of the 10 pages gained.
- The trend started earlier. The author notes citation trends were already moving before the change, and two controls is a thin comparison.
Otterly's own conclusion is the right one: its takeaway is "not to change publication dates on stale content without reviewing it," and to treat freshness as ongoing maintenance where the visible date reflects real work.
Does adding the current year to a title help AI citations?
The only public test found no effect you can separate from noise. In Otterly's year-in-title test, 11 pages had "2026" added to the title tag and H1 with nothing else changed. Citations rose 56% and 61% in the two waves, which sounds like a win until you see that an untouched reference page that already had the year in its title rose 63-64% over the same periods. One page ("Best AI Search Monitoring Tools") produced 86% of the 896 net gained citations; without it, the remaining Group B pages rose 12.5%. Six pages rose, three fell and two stayed at zero.
The author concludes the year in a title is "low-cost hygiene, not a citation lever" and should go on content that is genuinely current. One practical warning from the same test: titles said 2026 while body text and in one case the URL slug still said an older year. Mixed date signals are worse than none, so never put a year in a slug.
How long does an AI citation last before it decays?
Not long for most pages. Profound's September 30, 2026 study of 883,000 pages across seven engines (September 2025 to September 2026) found the median page loses half its citation share within 11 days of peaking. Of page-engine pairings observed at least two weeks past their peak, 78% had a half-life of 14 days or less; the other 22% held on for a median of 27 to 29 days depending on engine. Event-tied "report" content faded fastest. Profound did not analyze whether updates extend a page's life, so this tells you citations decay quickly, not that refreshing reverses it.
Our own data adds a complication every refresh test has to deal with: citations move even when nobody touches the page. Across 968 repeated runs of 42 prompt-engine pairs tracked in AI Rank Lab between February 13 and October 14, 2026 (pairs with at least five successful runs), back-to-back runs of the same prompt shared a median of only 41% of their cited domains (Jaccard 0.41), ranging from 0.59 on Perplexity to 0.26 on Gemini. We cover the full method in our AI answer stability study. The implication: a citation that disappears after a week may be churn, not staleness, so check before you rewrite.
What does Google say about changing dates on pages?
Google says dates should describe real publication or significant updates, and it names fake freshness as a warning sign. It does not list date changes as a named spam policy, but it is explicit that re-dating unchanged pages is search engine-first behavior.
Google's helpful content guidance (last updated October 5, 2026) lists questions where "Answering yes to some or all of the questions below is a warning sign that you should reevaluate how you're creating content." Two of them are about freshness:
"Are you changing the date of pages to make them seem fresh when the content has not substantially changed?"
"Are you adding a lot of new content or removing a lot of older content primarily because you believe it will help your search rankings overall by somehow making your site seem "fresh?" (No, it won't)"
Google's byline date documentation explains why simple re-dating is unreliable on its side: "Google doesn't depend on a single date factor because all factors can be prone to issues. That's why our systems look at several factors to determine our best estimate of when a page was published or significantly updated." It also asks you to "Make your dates and times consistent" between the visible date and structured data, and says "Don't specify future dates, or the date of the action described on the page."
On sitemaps, the sitemap documentation (last updated July 8, 2026) is the most precise statement Google has made on what counts as an update:
"Google uses the
value if it's consistently and verifiably (for example by comparing to the last modification of the page) accurate. The value should reflect the date and time of the last significant update to the page. For example, an update to the main content, the structured data, or links on the page is generally considered significant, however an update to the copyright date is not."
Google's John Mueller put it plainly in 2021, as reported by Search Engine Journal: "only update the date when you actually make significant changes on a page," adding that for trivial edits, "I don't think it would change anything with regards to search."
Bing says the same about sitemaps, with AI answers named explicitly. Its July 2025 webmaster post says "freshness signals directly influence how quickly updates are reflected in search results and AI generated answers" and advises: "Avoid setting lastmod to the time your sitemap was generated unless the content on that URL was actually updated." The same post frames this around "AI powered search engines like Bing Copilot," so it applies to Microsoft's AI answers directly.
Is updating the date without changing content considered spam?
Not by name. Google's spam policies (last updated August 28, 2026) do not mention dates. The practical risks are different: Google's systems estimate the real update date from several factors, so a fake date may simply be ignored; a sitemap whose lastmod values are not "consistently and verifiably" accurate stops being trusted for crawl scheduling, which hurts you when you do make a real update; and readers who see "Updated this month" over last year's prices lose trust in the page, which is the opposite of what E-E-A-T trust signals are for.
Why a tactic that works can still be a bad bet
The Otterly result and Google's guidance are not really in conflict. The test shows that some engines' rerankers reward a recent date string, which is exactly the bias Fang and colleagues measured. Google's guidance describes what a date is supposed to mean. A tactic that depends on a bias tends to stop working when engines correct it, and in the meantime the same test showed AI Mode and Gemini moving the other way. A real update captures whatever freshness benefit exists and survives the correction.
What counts as a substantive content update?
A substantive update changes what the page tells the reader: facts, figures, steps, recommendations or coverage. Google's sitemap guidance gives the floor (main content, structured data or links count; a copyright date does not). For a visible "Updated" date, use a stricter reader test: would someone who read the old version learn something new or be corrected?
Google does not publish a percentage threshold. Claims that Google wants "20-30% of the text" changed, which appear on some ranking pages, have no Google source that we could find.
| Change | Substantive? | Change visible date and dateModified? | Change sitemap lastmod? |
|---|---|---|---|
| Re-verified and corrected prices, plan limits or statistics | Yes | Yes | Yes |
| Replaced a superseded study with the newer edition | Yes | Yes | Yes |
| Rewrote steps after a product or API change, new screenshots | Yes | Yes | Yes |
| Added a section answering a question the page missed | Yes | Yes | Yes |
| Fixed broken links or updated internal links only | Minor | Optional; do not headline it | Yes (Google counts link changes) |
| Updated structured data only | Minor for readers | No | Yes (Google counts structured data) |
| Typos, formatting, image swaps | No | No | No |
| Changed the copyright year or sidebar | No | No | No |
| Added the year to the title only | No | No | No |
On this blog we enforce the same rule in our tooling: an article's dateModified is bumped only when the article is rewritten with fresh research, and a re-stamp without re-research is blocked. It slows us down slightly and keeps every date honest.
Which pages should you refresh first?
Refresh pages where facts have decayed and the page matters for prompts you want to win; leave accurate, stable pages alone. Freshness work is expensive, so triage by two questions: is the information wrong or incomplete now, and is the page losing citations or traffic it used to earn?
| Situation | What it usually means | Action |
|---|---|---|
| Was cited, citations falling over 2-4 weeks, facts now outdated | Genuinely stale | Refresh now, substantively; update all date signals |
| Was cited, citations falling, facts still correct | Churn, or a better page replaced you | Read the pages now cited; close coverage gaps; change the date only if you change content |
| Never cited for a high-value prompt | Not in the answer pool | Check indexing and snippet eligibility first, then rewrite or build a better-matched page |
| Cited and stable, facts correct | Working | Leave it; review on its cadence |
| Outdated, no traffic, no citations, overlaps another page | Dead weight | Merge into the stronger page and redirect |
| Event or news content after the event | Naturally historical | Do not re-date; add a dated note or link a new follow-up |
How often should you update content for AI search?
Match the cadence to how fast the page's facts decay, not to a calendar. No public study has tested review cadences, so the defaults below are our editorial recommendation, with triggers that override them. AirOps' data supports putting commercial pages first: over 60% of commercial pages cited by ChatGPT in its sample had been updated in the past six months, against informational pages where nearly a third were over a year old.
| Page type | Goes stale when | Default review (our recommendation) | What a refresh usually changes |
|---|---|---|---|
| Pricing, "best tools", comparisons | A vendor changes price or features; a new entrant | Every 60-90 days, plus on trigger | Re-verified prices and limits, added or removed products, a dated "checked on" line |
| Statistics and data roundups | A source publishes a newer edition | Every 90 days | Superseded numbers replaced, dead sources removed |
| Software how-tos and documentation | A UI or API release | On each relevant release | New steps and screenshots, version note |
| Regulated topics (health, finance, legal) | A rule or guideline changes | On trigger, plus a 6-month review | Updated requirements, reviewer and review date |
| Evergreen explainers | A mechanism or definition changes | Every 12 months | Facts, examples and links checked; often no date change needed |
How to run a substantive refresh
- Re-run the prompts and queries the page should win and read what is cited now. The newly cited pages show you what the engines consider current.
- Re-verify every number, price, version and claim against the primary source. Replace, correct or delete; do not leave a stale figure with a new date on top.
- Close coverage gaps. Add answers to sub-questions the cited pages handle and yours does not.
- Fix links and media. Replace dead or superseded sources and outdated screenshots.
- Write a short "What changed" note at the top or bottom with the date and a one-line summary.
- Update every date signal together: visible "Updated" date,
dateModified, sitemaplastmod. KeepdatePublishedand the URL unchanged. - Ask for a recrawl: URL Inspection in Search Console for Google, IndexNow or Bing Webmaster Tools for Bing.
If the refresh turns into a rewrite, AI Rank Lab's Content Suite builds a competitor-researched outline before writing, which helps with step 3. Whatever tool you use, step 2 needs a human checking primary sources. A site audit report is a good way to find pages with no modification date at all; our free SEO audit flags pages with no dateModified or article:modified_time and those last modified more than 12 months ago.
How to signal an update correctly: dateModified, visible dates and lastmod
Make three signals agree: a labeled visible date, dateModified in structured data, and sitemap lastmod, all set to the time of the last significant change. Google says it looks at several factors and that dates should be consistent, so a mismatch weakens all of them.
- Visible date. Google recommends adding "a user-visible date to the page" and labeling it with text like "Publish" or "Last updated." Show both if the page was updated.
- Structured data. Google recommends an
ArticleorBlogPostingtype withdatePublishedanddateModified, matching the visible dates, ideally with a time and the correct timezone. Keep the rest of the markup matching the visible page too. - Sitemap lastmod. Set per URL from the content's real modification time, never from the sitemap build time. Google ignores
- Push notifications. IndexNow tells Bing and participating engines that a URL changed; Bing recommends it alongside accurate sitemaps.
- Changelog. Not a Google requirement, but a short dated note helps readers and makes the update verifiable.
Should you change the publish date or the updated date?
Change the updated date and keep the original publish date. datePublished records when the page first appeared; dateModified records the last significant change. Re-dating the publish date of an old URL creates exactly the conflicting signals Google's byline documentation warns about. The exception is a true replacement, where you rewrite the page from scratch on a new topic; that is usually better as a new URL.
Does a CMS that auto-updates dates cause problems?
It can. Many platforms change the modified date on any save, including typo fixes and plugin updates, and some sitemap generators stamp every URL with the generation time. Both make your dates unverifiable. Check a few URLs: if the sitemap lastmod changes when the visible content has not, fix the generator before you start a refresh program.
How do you measure whether a content refresh improved AI citations?
Compare the refreshed pages against untouched holdout pages over equal before and after windows, per engine, using repeated runs of a fixed prompt set. Anything less cannot separate a refresh effect from normal churn, seasonality or a competitor's change.
- Fix the prompt set. List 10-30 prompts the page should answer and do not change them during the test.
- Record a baseline of 2-4 weeks, ideally with daily runs on each engine. In our data, 18 of 42 prompt-engine pairs (42.9%) flipped between mentioning and not mentioning the tracked brand at least once, so single runs mislead.
- Pick holdout pages similar in topic, format and citation volume that you will not touch. Otterly's tests show why: unmatched reference pages moved by 60% or more on their own.
- Refresh and log it: the date, what changed and every signal you updated.
- Confirm the recrawl before you start the post window. Google notes crawling "can take anywhere from several days to several months" depending on how often its systems revisit a page.
- Measure the same window length afterwards and compare the change in refreshed pages with the change in holdouts, per engine. Report the run count next to every rate.
- Check Google separately. Search Console's Generative AI performance report, rolled out to all sites on August 31, 2026, shows AI Overviews and AI Mode impressions by page and date, which is Google's own count. Our guide to the report walks through it.
AI Rank Lab can automate steps 2 and 6 (disclosure: it is our product). Brand visibility tracking runs a fixed prompt set on ChatGPT, Claude, Gemini and Perplexity each time you re-check it and records the cited URLs in every run, the citation tracker shows which of your URLs are cited and how that changes, and the AI Overview tracker covers Google, so you can see the week a page stops being cited. It cannot import Search Console's Generative AI report, because Google offers no API for it, so read that report in Search Console itself.
| Metric | Where to get it | What a refresh should move |
|---|---|---|
| Citation rate per prompt and engine | Repeated prompt runs (any tracker, or manual logs) | Up relative to holdouts |
| Share of prompts citing the page | Same runs | Up; more robust than raw counts |
| AI Overviews and AI Mode impressions | Search Console Generative AI performance report | Up for the refreshed URLs |
| Organic clicks and impressions | Search Console Performance report | Up or stable |
| AI referral sessions and conversions | GA4 (referrers such as chatgpt.com, perplexity.ai) | Up for the refreshed URLs |
| Recrawl after the update | Server logs, URL Inspection | A crawl after the change date |
Common content freshness mistakes
- Re-dating without re-checking. The page says "Updated October 2026" over last year's prices. This is the case Google's guidance describes, and readers notice first.
- Changing the URL to include a year. It throws away link and indexing history and forces another change next year.
- Sitemap lastmod set to build time. Bing specifically warns against it, and Google only uses lastmod when it is verifiably accurate.
- Calling one week's drop a decay. With median consecutive-run citation overlap of 41% in our tracking, a single week can drop for no reason. Look at a 2-4 week window.
- Refreshing everything on a schedule. Mass edits made to look fresh are, in Google's words, something that "won't" help. Spend the time on pages where facts decayed.
- Trusting vendor multipliers. Figures like "3.2x more citations within 30 days" circulate without published baselines or controls. Ask for the sample, the control and who sells the fix.
The bottom line
Content freshness in AI search is real but narrower than the headlines: ChatGPT, Perplexity and Claude lean toward recent pages, Google's AI surfaces barely do, and the strongest short-term "freshness" effect on record came from a date-string bias that engines have every reason to correct. Refresh the pages where facts have decayed and citations are slipping, make the update substantive, keep every date signal truthful, and measure against holdouts. If you want to see which of your pages AI engines stop citing, and when, AI Rank Lab's citation tracker records it run by run.
Get a free AI ranking consultation
Want your brand to show up when ChatGPT, Gemini and Perplexity answer questions in your market? Tell us about your site and our team will send a personalized strategy.
