Does Updating Old Content Help, and Which Freshness Signals Count?

Content freshness

How recently a page was published or genuinely revised, expressed both as a stated date and as whether the facts on the page are still current.

Also known as: Recency, Content recency

Three independent lines of evidence agree, which is unusual in this field. The SIGIR 2026 test by Vishwakarma, Kumar and Jamidar isolated recency as one of four unanimous gatekeepers across 252,000 trials. The GEO-16 analysis by Kumar and Palkhouski, covering 1,702 citations from 1,100 URLs across 16 B2B SaaS verticals, found "Metadata and Freshness" the strongest single correlate of citation at r=0.68. And Anthropic documents that its web search tool returns a page age field with every result handed to the model.

That third point is the one that changes how to think about it. Freshness is not inferred from ranking or from crawl frequency on that engine, it is a literal field in the payload the model reasons over. The model knows how old your page is before it decides which passage to quote.

Seer Interactive, across more than 5,000 URLs with extractable publish dates, found 31% of ChatGPT citations came from 2025 content, 50% for Perplexity and 44% for AI Overviews, and that 65% of AI bot log hits landed on content published within the previous year. The skew varies by engine, and it points the same direction on all of them.

The practical instruction is short. Show one visible last-updated date near the top, emit a dateModified that matches it exactly, keep sitemap lastmod accurate, and make the three agree by generating all of them from the same field.

Start with the mechanical reason it fails. Recency matters because the model is deciding which of two sources describes the current state of something. If the text still references a feature that was renamed, a rich result that was withdrawn or a price that changed, the date is not evidence of currency, it is evidence of neglect. The same trials that found recency a gatekeeper also found internal consistency a differentiator.

Then the policy reason. Google's spam policies define scaled content abuse by intent and value: pages generated primarily to manipulate search rankings rather than to help users. Republishing a library of unchanged pages under new dates on a schedule is exactly that shape, whether a person or a model did the republishing.

The honest version costs one line per revision. Keep a visible changelog entry naming what changed, and only move the date when there is an entry to add. That gives a reader and a model a reason to believe the date, and it makes the decision about whether to update explicit rather than automatic.

Competitor-triggered updates are the ones most teams never run, because nobody is watching. Oppira tracks changes to competitor landing pages, ads, posts and reviews and raises them as alerts, which turns "something in our market moved" into a dated entry you can act on rather than something you notice a quarter late.

A calendar-based refresh cycle is not useless, but it should be a backstop rather than the plan. Once a year, list the pages with no changelog entry in the last twelve months and check whether each one is still true. Most will be. The ones that are not are the whole return on the exercise.

What deliberately is not on this list: expanding the page, splitting it into shorter chunks, and adding markup. Ahrefs measured Spearman 0.04 between word count and citation position across 174,048 cited pages, and the SIGIR 2026 test found formatting changes negligible. Neither is what a refresh should spend its time on.

Schulte, Bleeker and Kaufmann, in "Don't Measure Once: Measuring Visibility in AI Search", show that single-point measurements of AI visibility are unreliable and recommend roughly 10 to 30 independent samples per prompt with reported confidence intervals. Citera separately noted that AI citation sources shift 40 to 60% month over month, which sets the size of the noise you are trying to see through.

So run it as a cohort comparison rather than a before-and-after. Pick the ten pages you are updating and ten comparable pages you are not, sample the same questions across both sets for four to eight weeks, and read the difference between the groups. That is a crude design, and it is far more evidence than a single-page anecdote.

Pair that with the two reports that give real data for free. Bing Webmaster Tools AI Performance shows total citations, page-level citation activity and the grounding queries an engine actually retrieved on, which no other free tool provides. Search Console reports impressions of your URLs in Google generative AI features, where the site has that report available.

Then close the loop with the changelog. Because every real update leaves a dated line, you can join citation and impression movements back to the specific edit that preceded them, which is the only way a refresh programme accumulates knowledge rather than opinions.