Content freshness has two sides. The first is the content itself: Are the prices, services, opening hours, contact people, and figures on a page still correct? The second is how a page is dated: the dates that tell people and software when it was published and when it last changed. A guide marked “Last updated September 14, 2026” shows at a glance how current it is; without a date, readers can’t tell.

Content freshness is different from the knowledge cutoff, the date up to which a language model’s training data reliably covers information. Content freshness mainly concerns the content an AI system retrieves from the web when answering a question and uses as the basis for its answer, an approach called retrieval-augmented generation. What counts is the version of a page the system has at hand: often the one its web index captured most recently.

How publication and update dates are signaled

A date can be stated on several levels that complement each other:

  • Visible date: a clearly labeled date on the page, such as “Published March 2, 2026” or “Last updated September 14, 2026.” Google recommends featuring the date prominently and labeling it this way. On articles, it often sits in the byline, which also names the author and so makes the authorship visible.
  • HTML time element: in the page’s code, its HTML, the visible date can be wrapped in a time element, a building block the HTML standard provides for dates and times that also makes them machine-readable. Its datetime attribute, an added detail inside the element, holds the date in an unambiguous form: “September 14, 2026” becomes 2026-09-14.
  • Structured data: in structured data based on schema.org, the properties datePublished and dateModified state the date of first publication and the date of the last change. Google recommends them in types such as Article for articles and BlogPosting for blog posts.
  • XML sitemap: the lastmod value in an XML sitemap tells crawlers—programs that fetch websites automatically—when a page last changed.

To determine a page’s date, Google says it doesn’t rely on a single factor, because any factor can be prone to errors; it looks at several instead. The date should describe when the page was published or updated, not, for example, the date of an event the page reports on, and it shouldn’t be a future date. If the wrong date is still picked up despite these rules, Google suggests removing some or all other dates on the page.

What AI systems learn about a page’s age

AI providers disclose only parts of how their systems use dates. For the web search tool of the Claude API, which developers use to let Claude search the web, Anthropic documents that the tool can return a page_age field for each page it finds, next to the page’s address and title: when the page was last updated. The field can also be empty. Perplexity’s Search API for developers returns a publication date and a last-updated date for each result where available and lets developers filter results by both. Neither provider describes how much a date influences which sources an AI answer draws on.

Microsoft stays more general. Bing’s Webmaster Guidelines also cover Copilot and web content Bing supplies to AI applications. They say that content should be updated when facts or guidance change and that freshness signals should be used appropriately.

A lab study suggests that language models can favor content that is dated more recently. Researchers from Waseda University in Tokyo and The Hong Kong Polytechnic University presented it at the SIGIR-AP 2025 conference. They had seven language models in versions from 2023 and 2024, including OpenAI’s GPT-4o, sort text passages by how well they matched a question. When the researchers put a made-up publication date in front of each passage, the “newer” passages were consistently moved up; larger models weakened this effect but did not eliminate it. The passages themselves came from a standard research test collection, and only the dates were artificial; the study did not measure AI systems such as ChatGPT or Perplexity.

Real changes, not just new dates

An update date is only worth as much as the change behind it. Among the warning signs in Google’s guidance on helpful content is changing the date of pages to make them seem fresh when the content has not substantially changed. A new date therefore belongs with a real revision, such as new prices, an added section, or corrected facts. A date that changes with every minor tweak, or automatically every day, no longer tells people or machines anything.

The most important freshness signal is still content that is correct. When an AI system retrieves a page, it bases its answer on what the page says: if a pricing page still lists last year’s prices, the answer may repeat them, even if the page was retrieved moments earlier. When prices, services, locations, or contact people change, it pays to revise the affected pages promptly and update the visible date (where there is one), dateModified, and lastmod together. If the content management system used to maintain the website sets these values automatically, it’s worth checking that they reflect real changes.