Content Freshness and AI Citations: What the Evidence Leaves Unresolved

Nobody has shown that updating a page makes AI engines more likely to cite it. That includes the vendors themselves. The evidence splits into three parts, and none of them closes the question.

• Exposure. Some engines pass a page's last-updated date to the model. Exposure is not weighting.
• Observation. One widely shared study found most cited pages carry recent dates. It never checked uncited pages.
• Experiment. One lab study found a recent timestamp helped. It ran on synthetic pages already placed in the model's context.

What this means for you: freshness is not a lever you can pull with confidence. It is also not proven useless. The honest position is that nobody knows yet. If you want the definition and grading in brief, see content freshness: definition and evidence.

A date only matters if the engine sees it. That depends on which stage of the answer you are talking about.

Parametric answers (no web access)

When a model answers from memory, it knows nothing published after its training cutoff. OpenAI's model docs list a February 16, 2026 cutoff for the GPT-5.6 family and August 31, 2025 for the Instant path ChatGPT uses. Updating a page today cannot reach those answers. No publisher controls this stage.

Retrieval answers (live search)

When an engine searches, it may see dates. Anthropic's web search tool documentation lists a page_age field described as "When the site was last updated". That proves Claude's API surfaces the date. It does not say the model prefers newer pages. The Claude visibility factors page grades this signal alongside the others.

The engine also writes its own queries. OpenAI says ChatGPT search rewrites prompts into targeted queries. Google may fan out across subtopics. So you cannot tell from outside whether a query even asked for recent material.

Where dates come from

Engines and studies read dates from four places, all of which you write yourself:

• the sitemap lastmod value
• the schema dateModified property
• the HTTP Last-Modified header
• the date shown on the page

Every one of these is self-declared. That fact drives most of what follows.

The study most often cited for "update your content for AI" is Seer's, published July 24, 2026. Seer reported that 75% of cited pages had been updated within a year, with Gemini at 78%, ChatGPT at 73% and Perplexity at 65%. The sample covered 7,683 pages across four brand categories.

The problem is the missing denominator. Seer looked only at pages that were already cited. It had no group of uncited pages to compare against, and no base rate for the web as a whole.

Say most crawlable pages already carry a recent date because a CMS stamps lastmod automatically. Then 75% of cited pages being recent is exactly what you would expect by chance. The finding would carry zero signal. The study cannot rule that out.

The dates were also read from publisher-declared fields. Those are the values an optimiser can change without touching the content. So the study measured what sites claim, not what changed.

Seer's more solid finding describes the web, not the engines: a 72% versus 42% gap between a page's apparent update date and its original publication date. Many pages look newer than they are. That is a reason for caution, not a tactic.

There is one experimental result in favour of freshness. The 252,000-trial factorial study, What Gets Cited, reported that "including explicit price information and a recent timestamp also helps consistently."

Read the design before you act on it:

• The corpus was synthetic: 100 e-commerce blog posts.
• Each trial used two candidate documents.
• No search engine was queried. The documents were placed straight into the model's context.
• The paper comes from an industry team that ships a GEO checklist.

So it shows that a visible date can change which of two pre-loaded documents gets cited. It does not show that a date helps a page get retrieved in the first place. A critical survey of 45 GEO studies found this is the pattern across the field: already-retrieved content can change its own citation, but no technique shows a stable, cross-platform effect on being found.

The same study found the strongest drivers of citation were topical relevance and list position. Freshness was a side effect, not the headline. Our review of what citation optimization studies show covers the wider set.

Most of the freshness question sits outside a publisher's reach. Here is the split.

One more boundary: change notification. Bing's webmaster team has said IndexNow helps AI systems reference current versions of pages. That is an assertion with no study behind it. The IndexNow documentation never mentions AI, and a successful ping only confirms receipt. Google does not consume IndexNow.

Google's own AI optimization guide says there are no special optimizations needed for its generative AI features. Freshness tricks are not on its list of requirements.

Work through these in order. Each step is either verifiable or fixes something real.

1. Confirm crawlers can fetch the page. A date means nothing if OAI-SearchBot, PerplexityBot or Claude-SearchBot get a 403 from your firewall, or if the content only appears after JavaScript runs. Access comes first.
2. Find facts that are actually out of date. Prices, plan names, feature lists, team size, locations. A wrong price in an AI answer is a real problem with a real fix.
3. Update the content, then the date. Change the date only when the substance changed. A typo fix is not an update.
4. Make your four date sources agree. Sitemap lastmod, schema dateModified, the Last-Modified header and the visible date should tell the same story. The date consistency check walks through each one.
5. Show the original publication date too. Readers and engines can then see both when it was written and when it last changed.
6. Re-check answers over time. AI answers shift between checks for reasons that have nothing to do with your page. Read what changes between AI visibility checks before attributing any movement to an update.

If you want a baseline before you start, the free checker on this site shows whether AI engines name you for up to three buyer questions. It is a snapshot, so treat one result as a starting point, not a trend.

These are the freshness habits most likely to waste effort or backfire.

• Bumping dates without changing content. Setting dateModified to today on an unchanged page games your own measurement. If engines ever do weight dates, it is the obvious pattern to discount. There is no evidence it helps. The specific question is covered in does changing the last updated date help AI cite you.
• Letting the CMS stamp every page daily. Automatic lastmod values on every build make all your pages look equally new. That removes any information the date could carry.
• Treating Seer's 75% as a target. It describes cited pages, with no comparison group. It is not a threshold to hit.
• Quoting the lab result as proof. A timestamp helping in a two-document, pre-loaded test does not show it helps you get retrieved.
• Stacking freshness on other myths. Adding schema and rewriting in an "AI-friendly" style alongside new dates is not supported either. See does schema markup help AI citations.
• Reading one engine's behaviour as all engines. Claude's API exposes page_age. That says nothing about ChatGPT, Perplexity or Google AI Overviews.

The gap is a missing number. Someone needs to measure what share of all crawlable pages carry a recent lastmod or dateModified. If that share is also near 75%, Seer's finding is noise. If it is much lower, freshness becomes a real candidate signal.

A stronger test would compare pages that genuinely updated content against matched pages that did not, across several engines, over months. Nobody has published that.

Until then, the defensible approach is simple. Keep facts current because readers and engines both quote them. Keep dates honest and consistent because they are cheap to get right. Don't sell, buy or perform date changes as an AI tactic.