Content Freshness & AI Search: Why Recency Is a Citation Signal

Why Freshness Matters More in AI Search Than Traditional SEO

In short

In traditional SEO, freshness is one ranking factor among hundreds. In AI search, freshness is closer to a gating factor: LLMs increasingly retrieve in real time, and a model deciding which two or three sources to cite for a time-sensitive query will heavily prefer recent ones. Stale content is filtered out before the citation slot is awarded.

Google has treated freshness as a ranking signal since at least the 2011 "Freshness Algorithm Update". Newer is not always better, but for time-sensitive queries it usually wins.

AI search behaves the same way — only more so. ChatGPT Search, Perplexity, and Google AI Overviews all retrieve sources at query time, and only the top few make it into the visible answer.

That citation slot is small. A query like "current EU AI Act timeline" will surface two or three sources, not twenty. Recency is one of the cheapest filters the model can apply first.

Three reasons freshness weighs heavier in AI search.

• Cited answers are sparse. Generative engines pick a handful of sources per answer. Stale pages are pruned early.
• Models are explicit about dates. Many LLM answers now include "as of [date]" framing, which makes outdated sources visibly disqualifying.
• Trust compounds. A page that is regularly updated signals an active maintainer, which is itself an E-E-A-T signal that traditional and AI search both reward.

The practical consequence: in AI search, freshness is not a tie-breaker. It is part of the qualifying round.

Schema.org Freshness Signals: datePublished, dateModified, lastReviewed

In short

Schema.org Article defines datePublished (first publication date) and dateModified (most recent meaningful update). Both are ISO 8601 strings inside JSON-LD and are read by both Google and LLMs. lastReviewed is a separate property — primarily defined on MedicalEntity — but the convention of surfacing 'last reviewed' in JSON-LD or in visible UX is now widespread.

Schema.org Article is the canonical schema for editorial content. Two of its date properties matter for freshness: datePublished and dateModified.

datePublished. The date the article was first published. Set once. Do not change it when you make minor updates — that destroys the historical signal.

dateModified. The date of the most recent meaningful update. This is the field freshness depends on. Update it when content changes substantively, not for typo fixes.

Minimum example.

{
  "@context": "https://schema.org",
  "@type": "Article",
  "headline": "Content Freshness & AI Search",
  "datePublished": "2026-05-06",
  "dateModified": "2026-05-06",
  "author": { "@type": "Person", "name": "Linus Ingemarsson" }
}

lastReviewed. Schema.org defines this property on MedicalEntity, but the underlying idea — "we looked at this on date X and confirmed it is still correct" — is broadly useful. Many sites now surface a "Last reviewed" line in the rendered HTML even when JSON-LD does not include it.

What LLMs actually read. Generative engines parse both the JSON-LD block and the visible HTML. If your dateModified in JSON-LD says May 2026 but the rendered byline says "Posted 2022", the model will hedge — and may pick the older signal. Reconcile both.

Three rules to follow.

1. Use ISO 8601 (YYYY-MM-DD) in JSON-LD — not human strings like "May 6, 2026".
2. Update dateModified only on meaningful edits. Do not bump it daily for a stale page; LLMs and Google detect the pattern.
3. Make the visible date in the rendered HTML match the JSON-LD dateModified. Mismatch destroys trust.

Visible Freshness UX Patterns: 'Updated April 2026' Badges

In short

Visible freshness signals are the user-facing counterpart to Schema.org dates. The most common pattern is an 'Updated [Month Year]' badge near the title, sometimes paired with 'Originally published' and 'Last reviewed' lines. They serve two audiences: humans (trust) and LLMs (which parse rendered HTML alongside structured data).

Schema is invisible. Most readers — and many AI crawlers — also evaluate the page they actually see. Visible freshness signals close that gap.

Common patterns.

• Updated badge. "Updated April 2026" near the title or under the byline. Short, scannable, persistent.
• Three-line metadata. "Published [date] · Last reviewed [date] · Next review [date]". Used on regulatory, medical, and high-stakes editorial content.
• Inline 'as of' framing. "As of May 2026, the EU AI Act schedule is …" — this is the same pattern LLMs use, and mirroring it improves model alignment.
• Changelog block. A collapsible "What changed in this update" list. Especially useful for guides and playbooks.

Where to put the signal. Above the fold, near the byline, is the most extractable location. Putting it only in the footer means the model may not read far enough to find it.

Match the schema. If JSON-LD dateModified says 2026-05-06 and the visible badge says "Updated April 2026", you have a 30-day mismatch. Pick one source of truth and render from it.

The Alice Labs convention. Every article carries a three-date metadata block — Published, Last reviewed, Next review due — visible above the fold and mirrored in JSON-LD. The "next review" date is a public commitment that the page is on a cadence, not abandoned.

Review Cycle Methodology: 90 Days Default, Faster Where Needed

In short

A review cycle is a documented schedule for re-examining published content. A 90-day cycle is a sensible default for evergreen guides. News, pricing, and regulatory content need shorter cycles (7-30 days). Glossary and definitional content can stretch to 180 days. The cycle is the discipline; the dateModified is the proof.

Reviewing content every 90 days is common practice across SEO and editorial teams. It is short enough to catch drift, long enough to be sustainable.

But 90 days is a default, not a universal rule. Different content types decay at different rates.

Choosing the cycle by content type.

• News / current-events posts. 7-14 days. After that, re-frame as "[Date] retrospective" or unpublish.
• Pricing / commercial pages. 30 days. These pages are often consulted by AI agents for comparisons; outdated prices are a trust break.
• Regulatory / compliance content. 30-60 days, plus an event-trigger (review immediately on legal change).
• Evergreen guides and deepdives. 90 days. The Alice Labs default for the LLMO content stack.
• Glossary / definitional pages. 180 days. Definitions move slowly; reviewing too often is wasted effort.
• Statistics roundups. 90 days, plus an immediate refresh whenever a referenced source publishes new data.

What "review" actually means. A review is not a re-read. It is a structured check.

1. Are all cited statistics still current? Replace where superseded.
2. Are all external links still resolving? Fix or remove.
3. Has the underlying landscape changed? Edit the relevant section.
4. Update dateModified and the visible "Updated" badge — but only if step 1-3 produced a real change.

Source: Alice Labs methodology, applied across 100+ Nordic enterprise implementations (2024-2026)

LLM Citation Behaviour on Stale vs Fresh Content (Qualitative)

In short

Across audits of ChatGPT Search, Perplexity, and Google AI Overviews, fresh sources are consistently preferred for time-sensitive queries — and the freshness threshold appears tighter than in traditional SERPs. This is a publicly observed pattern; specific lift percentages vary by query and engine, so we treat the relationship as qualitative, not quantitative.

We do not claim a specific percentage uplift from freshness alone — published research isolates that variable poorly. What we do observe consistently across LLMO audits is a behavioural pattern.

Patterns observed in citation audits.

• Time-sensitive queries. Sources older than 12 months are routinely passed over for "current state of X" queries in ChatGPT Search and Perplexity, even when older pages still rank well in Google's blue links.
• Date-conditioned answers. When a model produces "as of [date]" framing, it tends to choose sources whose visible date is close to that frame — not earlier than it.
• Statistical claims. Models prefer sources published within ~24 months of the cited statistic. Older numbers get hedged with "older estimates suggest …".
• Definitional content. Freshness matters less for definitions and glossary terms — these can be older without being filtered out.

The compounding effect with citations. Aggarwal et al. (2024) showed citation-rich content can lift generative engine visibility by up to 40%. In our audits, the citation effect and the freshness effect compound: a fresh, citation-rich page consistently out-cites a fresh, citation-light page.

SparkToro's 2024 zero-click study reported roughly 60% of searches end without a click. The implication is that being cited is the new visibility — and being cited depends on being crawled fresh.

We deliberately avoid quoting a specific freshness lift. The honest framing: freshness is a qualifier, citations are an amplifier, and the two together drive most LLMO outcomes.

The Alice Labs Review Cadence: A Proprietary Methodology

In short

Alice Labs runs a documented 90-day default cycle across the LLMO content stack, with shorter cycles for high-decay content types and event-driven triggers for regulatory and pricing pages. Every article carries a visible Published / Last reviewed / Next review block, mirrored in Schema.org JSON-LD, and a quarterly LLMO Citation Benchmark verifies that citations move with the cadence.

The Alice Labs review cadence is the operational layer that turns "freshness matters" from a slogan into a process. It runs across 100+ Nordic enterprise implementations.

Five components.

1. Default 90-day cycle. Every evergreen article has a nextReviewDue set to publishDate + 90 days. The date is public and visible above the fold.
2. Tier overrides. Pricing, regulatory, and statistics roundups are tagged with tighter cycles (30, 30-60, 90 days plus source-trigger).
3. Event triggers. Major rule changes, competitor launches, and referenced-source updates pull a review forward regardless of the schedule.
4. Visible + structured dates. Published, Last reviewed, Next review due — rendered in HTML and mirrored in Schema.org datePublished / dateModified.
5. LLMO Citation Benchmark. Quarterly audits across ChatGPT, Perplexity, Claude, and Google AI Overviews to confirm that citation share moves with the cadence — not just the rank.

What we measure each cycle.

• Citation share per target query in each generative engine.
• Crawl frequency by AI user-agent (GPTBot, ClaudeBot, PerplexityBot, OAI-SearchBot).
• Visible and structured-data dates — are they reconciled?
• External and internal link health.
• Whether referenced statistics or sources have been superseded.

What we deliberately do not do. We do not bump dateModified for typo fixes. We do not auto-rewrite content on a timer. We do not promise a freshness uplift in percent terms. Cadence is a discipline; outcomes follow when content quality is already strong.