Research

Does freshness get you cited by AI search?

I asked ChatGPT, Claude, Perplexity and Google AI Mode 60 questions and checked the dates on 2,500 pages they could have cited. Among Google's top 10, pages showing an old date were cited less often, but a recent date didn't clearly help.

Does freshness get you cited by AI search?

Key findings

Bottom line: a recent date didn't clearly help pages get cited by AI search, but an old one seemed to hurt. Pages with no date did about as well as recently dated ones. Changing only the date wasn't shown to work, and ChatGPT and Claude only searched the web for questions whose answers change over time.

  • A recent date didn't clearly help. Among Google's top 10 results, recently dated pages had 1.5 times the odds of being cited, but the result wasn't statistically reliable (p = 0.16).
  • An old date seemed to hurt. Pages showing a date more than a year old had less than half the odds of being cited compared with recently dated pages (p = 0.02). Pages with no date did about as well as recent ones.
  • Seer's 75% describes the web, not AI preferences. In July 2026, Seer Interactive found that 75% of the pages AI tools cite had been updated in the last year. 76% of the cited pages I checked showed a recent update too, but so did 77% of the top-10 pages no tool cited.
  • Changing only the date: inconclusive. Of 123 pages I could compare before and after their update, 37 had changed nothing but the date. They were cited a little less often than pages with real updates, but there were too few to say for sure.
  • ChatGPT and Claude only search when the answer changes. On default settings, both searched the web for all 30 time-sensitive questions and none of the 30 evergreen ones.
  • Question type made no difference to freshness. A recent date mattered no more for time-sensitive questions than for evergreen ones.

How I tested it

In July 2026, Seer Interactive looked at 7,683 pages cited by ChatGPT, Gemini and Perplexity and found that 75% had been updated in the last year. That shows cited pages tend to be recently dated, but not whether AI tools prefer them, since most pages on the web might be recently dated too. To find out, I compared the pages AI tools cited with the pages they could have cited and didn't.

I asked ChatGPT, Claude, Perplexity and Google AI Mode 60 questions, twice each, and recorded every page they cited. Half the questions were time-sensitive, like contribution limits and prices, and half were evergreen, like definitions and how things work. ChatGPT and Claude were each tested on their default settings and with web search switched on.

For each question I also kept Google's top 10 results and every page ChatGPT opened while answering, including the ones it didn't cite. I then visited about 2,500 pages and recorded the dates they show. Each page was only compared with other pages for the same question, taking Google rank and page length into account.

To see whether a new date meant real changes, I compared archived copies of pages from before and after their claimed update, and a reviewer sorted them by hand. The full method is in the Methodology section at the end.

Results

The chart shows odds ratios. An odds ratio of 1 means a factor made no difference to whether a page was cited, above 1 means it made a page more likely to be cited, and below 1 means less likely. Only the two highlighted results have 95% confidence intervals that don't cross 1.

Chart of odds ratios with 95% confidence intervals. Google top 10: dated in the last year vs not 1.48, dated over a year ago vs dated in the last year 0.44, no date vs dated in the last year 0.96. Pages ChatGPT read: dated in the last year vs not 0.74, dated over a year ago 1.65, no date 1.20, date-only change vs real update 1.49.
Odds of being cited compared with other pages for the same question, with 95% confidence intervals, controlling for Google rank and page length. From 720 answers collected on November 10, 2026.

Does a recent date help?

In Google's top 10, recently dated pages leaned toward being cited more (odds ratio 1.48, 95% CI 0.85 to 2.57), but the result fell short of the bar I set in advance. Splitting the other pages shows why. Pages dated more than a year ago were cited less than recently dated ones (0.44, 95% CI 0.21 to 0.90), while undated pages did about as well (0.96). Lumping old and undated pages together blurred the difference. The older a page's date, the lower its chance of being cited.

ChatGPT pointed the other way. Among the pages it read, it cited recently dated ones a little less often (0.74, p = 0.09), and it cited pages with an old date more often than recent ones (1.65). The gap came mostly from commercial blog posts with recent dates that ChatGPT read and passed over. Government and university pages showed no such pattern.

No single tool showed a reliable effect on its own. 79% of top-10 pages were cited by at least one tool, mostly because Perplexity and Google AI Mode list many sources, which leaves little room to see a difference. Counting only links inside the answer itself, the top-10 result was 1.24 (p = 0.35).

Seer's 75%, side by side

Most cited pages did show a recent update, just as Seer found. But the top-10 pages no tool cited showed one just as often, so the 75% describes how pages in these results are dated, not something the AI tools prefer.

PagesUpdated in the last yearFirst published in the last year
Cited by any tool (1,772)76%54%
Google top 10, never cited (97)77%35%
Seer, July 202675%42%

Shares are of pages showing that kind of date. Cited pages were more often first published in the last year, but only 65 uncited pages showed a publish date, so that gap is uncertain.

Does changing only the date work?

Of 708 pages claiming an update in the last year, 349 had never been saved by the Wayback Machine, 123 turned out to be new, and 88 had no copy from after their update. That left 123 pages I could compare by hand.

What changedPagesCited, Google top 10Cited, pages ChatGPT read
Rewrite (20% or more of the text)3879%33%
Fact update (a number, price, rule or named item)3386%44%
Light edit (wording only)154 of 464%
Date only3769%32%

Date-only pages were cited a little less often than pages with real updates. The formal test was inconclusive, with an odds ratio of 1.49 and a 95% confidence interval from 0.23 to 9.5 across only 13 questions. Many of the date-only pages were Google's own Search Central docs, which update their "Last updated" line automatically.

Do ChatGPT and Claude look for fresh pages?

Default settingTime-sensitive questions searchedEvergreen questions searched
ChatGPT30 of 300 of 30
Claude30 of 300 of 30

The split was the same on both runs of every question (Fisher's exact test, p < 0.0001). For evergreen questions, neither tool looked at the web at all, so a page's freshness couldn't come into it. For time-sensitive ones, both went looking every time.

Lookalike sites

Some of the pages the AI tools surfaced were copies of real sites on other domains, such as a page made to look like Google's developer docs at an unrelated address and proxies copying Shopify's help pages. I left them out of the analysis, but they're worth knowing about if you rely on AI answers for technical documentation.

What this means

  • An old visible date is the clearest risk. If a page shows a date from years ago, update it properly or reconsider showing the date. In this data, no date did as well as a recent one.
  • Don't count on a date bump alone. Pages that changed only their date were cited no more often than pages with real updates, and slightly less in this sample. Update the facts that change, like prices, limits and rates, and the date will follow.
  • For evergreen topics, freshness may not come into it. On default settings, ChatGPT and Claude answered evergreen questions from what they already knew, so being well known and widely cited matters more there than a recent timestamp.
  • Time-sensitive pages do get looked at. Every time-sensitive question sent both tools to the web, so pages about prices, limits and rules are where keeping facts current has a chance to pay off.

Limitations

  • Observational. The study shows which pages were cited, not why. Sites that keep their dates current may also be better maintained in other ways.
  • Archive gaps. Half the pages claiming a recent update had never been saved by the Wayback Machine, and smaller sites are missed most. I also tried Common Crawl to fill the gaps, but its index was unreliable and most recently updated pages hadn't been crawled since their update, so the date-only comparison rests on 123 pages, mostly from larger sites.
  • A broad definition of cited. Perplexity's full sources list and Google AI Mode's sources panel count as cited, which makes most top-10 pages count as cited.
  • Dates read once. Pages were visited a day or two after the answers were collected, so a few may have changed in between.
  • One location. Everything ran from Victoria, BC, so answers and sources lean Canadian.
  • A snapshot. The answers were collected on November 10, 2026, and AI tools often change how they find sources.

What's next

In about six weeks I'll visit the same pages again, record which ones changed their date and what else changed, and ask the same 60 questions again. That avoids relying on the archive and should give a clearer answer on whether changing the date works. I'll add the results to this page.

Study data

Download the pages (CSV), with one row per question and page: the question and its type, the page URL, its Google rank, whether ChatGPT read it, which tools cited it, its claimed update and publish dates, its date group, its word count and, where available, what the archive comparison found.

Methodology

Conrad Studio original research. Answers were collected on November 10, 2026, page dates on November 11 and 12, and archive copies on November 12, 2026.

Questions

60 questions in three areas: home services, B2B SaaS and ecommerce marketing, and personal finance and travel. In each area, 10 were time-sensitive, where the right answer changes from year to year ("RRSP contribution limit", "how much does Shopify cost per month"), and 10 were evergreen ("how does a heat pump work", "what is customer lifetime value"). Questions didn't name a year. A second AI scorer labelled each question without seeing my labels, and any question we disagreed on was replaced. None of the chunking study's questions were reused, and "best company" questions were left out. The four hypotheses and the result that would count as a confirmation were written down before any answers were collected.

Tools

Each question was asked twice in each of six setups, for 720 answers in all.

ToolSetup
ChatGPTTemporary chat set to Unpersonalized, default settings, so it searched the web only when it decided to
ChatGPT, search onThe same, with web search switched on
ClaudeIncognito chat, default settings
Claude, search onIncognito chat, with "Search the web:" at the start of the question
Perplexity ProDefault search, with memory switched off
Google AI ModeSigned in, with Personalize Search switched off

All the tools were used from the same Chrome browser in Victoria, BC. Every answer was saved word for word with a checksum, and a screenshot was taken at the time. Perplexity's full sources list and Google AI Mode's sources panel count as cited.

Pages

For each question I kept Google's organic top 10 and every page ChatGPT read while answering, cited or not. That gave two fair comparisons: top-10 pages any tool cited against top-10 pages none cited, and pages ChatGPT read and cited against pages it read and passed over. A third pool, every page any tool cited, was used only to describe how old cited pages are. Reddit, forums, social media, YouTube and PDFs were left out, along with lookalike sites that copy real pages on other domains.

Dates

A script visited each page and recorded its claimed update date (the latest of a visible "updated" date, schema dateModified and the modified-time meta tags) and its claimed publish date. Visible dates were only used when the page had none in its code, because sidebar lists of other posts can make a page look newer than it is. Dates after November 10 were ignored, since the tools couldn't have seen them. A page counts as recently dated if its newest date is within 12 months of November 10, 2026. 2,546 pages had enough text to use (40 words or more), including 1,462 of the 1,547 pages in the two comparison pools.

A reviewer checked 50 random pages by hand without seeing the script's dates. After two rounds of fixes to the script, they agreed on whether a page was recently dated for 49 of the 50.

Real changes

For the 708 pool pages claiming an update in the last 12 months, I used the Wayback Machine to find when each page was first saved, a copy from at least 30 days before the claimed update, and the latest copy after it. Pages first saved within the last 12 months count as new. For the 123 older pages with both copies, a reviewer who couldn't see the script's results compared the main text and labelled each page: rewrite (20% or more of the text changed), fact update (any number, price, limit, rate, rule or named item changed), light edit (wording only) or date only (nothing but dates or years). The script's own labels agreed with the reviewer on only 22 of 40 pages checked, so the published results use the reviewer's labels.

Analysis

Each page was compared only with the other pages for the same question (conditional logistic regression grouped by question), with Google rank and page length as controls. Results are odds ratios. The search question was tested with Fisher's exact test at the question level, because the split between question types was complete. Checks included each tool on its own, standard errors grouped by site, leaving out canada.ca (more than 5% of pages), and counting only links inside the answer. An independent checker recomputed every number from the raw files.

let’s Work together