The cheapest citation lever we know is a date. Not writing new content, not link building: keeping the updatedAt on your existing pages current, consistent, and honest. In our May cohort, refreshed pages saw AI citation mentions rise roughly 60% within four weeks, at a median cost of 20 minutes per page.
We have written before about visible freshness stamps. This note is about the plumbing underneath them: the four places the date lives, the rule for when it may change, and the monthly pass that keeps a whole site current.
It is unglamorous work. It is also the best effort-to-citation ratio on the menu, and the results show up within a month.
What counts as a freshness signal to an AI engine
A freshness signal is any machine-readable claim about when your content last changed. There are four that matter: the visible updated date rendered on the page, the dateModified field in your Article schema, the lastmod value in your XML sitemap, and the Last-Modified HTTP header. Answer engines read the first two directly, and crawlers use the second two to decide what to refetch and how often.
Most sites maintain one of the four, usually the visible stamp, and let the other three drift or contradict it.
That drift is invisible to humans and loud to machines. A crawler comparing surfaces sees a page that cannot agree with itself about its own age.
In a CMS-driven stack, one updatedAt field should feed all four automatically. That single pipe is the whole technical fix.
Why answer engines weight recency so hard
Engines weight recency because their users punish staleness. An assistant that quotes 2024 pricing in 2026 loses trust instantly, so retrieval layers boost recently updated sources for any query where facts can drift: prices, tools, statistics, recommendations, anything with a year in it. In our audit data, a visible and current updated date roughly doubles citation rate against identical content with no date, and the gap widens on query topics that change fast.
Recency also drives crawl behavior. PerplexityBot in particular refetches pages within about 48 hours of a sitemap lastmod change in our logs, which means honest updates propagate into answers fast.
Stale pages fall out of answers quietly. Nothing alerts you. The engine just starts preferring a competitor whose numbers claim to be from this quarter.
Freshness is not a bonus signal. For fact-heavy queries it is closer to an entry requirement.
The four date surfaces must agree
The trust value of a date comes from consistency. When the visible stamp says June, the schema says June, the sitemap says June, and the header agrees, the claim is corroborated four times. When the page says June but the sitemap says 2024, the claim is contradicted, and a contradicted signal is worth less than no signal. We fixed exactly this mismatch on one client site and their cited-page count rose 40% inside a month, with no content changes at all.
The failure is almost always plumbing. A rebuild regenerates every sitemap lastmod, or the schema hardcodes the publish date, or a CDN mangles Last-Modified.
The audit takes ten minutes: pick three pages, check all four surfaces on each, and note disagreements. Most sites fail on the first page.
Wire everything to the CMS updatedAt field once, and the surfaces can never argue again.
What earns a new updatedAt (and what does not)
Our rule: the date changes when a returning reader would notice the difference. Refreshed statistics, a rewritten or added section, corrected facts, a replaced example, an updated recommendation, all qualify. Typo fixes, comma repairs, image swaps, and styling changes do not. The working threshold is that at least 10% of the page changed or a load-bearing fact was corrected. Below that, the old date stays, no matter how tempting the bump.
The rule exists because the date is a promise. It tells the engine and the reader that the substance was re-verified on that day.
Keep a one-line change note per update in the CMS. Ours read like commit messages: refreshed 2026 pricing, replaced retired tool, added crawler section.
That note keeps you honest when you cannot remember whether March's edit was substance or punctuation.
The monthly freshness pass we run
On the first Tuesday of each month, the engine flags every page that is either older than 90 days by updatedAt or contains a fact its audits found drifting, a price, a version, a statistic. June's pass flagged 25 pages. Eleven earned a genuine refresh at about 20 minutes each, updated numbers, one rewritten section, a sharper lead answer. Fourteen were still accurate and kept their dates. Total human time: under four hours.
The 90-day threshold fits our topics, which drift fast. Slow-moving fields can run 180 days without penalty.
Prioritize by stakes: pages that already earn citations get refreshed first, because those are the rankings you can lose.
Four hours a month keeps a 90-page site perpetually current. That is the entire program.
What honest refreshes did to citations
The May cohort is our cleanest evidence. Eleven pages got genuine refreshes, and across the following four weeks their combined AI citation mentions rose about 60%, from 41 tracked mentions to 66, while a matched set of untouched pages stayed flat. Three of the refreshed pages entered AI Overviews for queries they had never surfaced on. Median time from update to first new citation: 16 days.
The winners shared a pattern: the refresh sharpened the extractable answer as well as the facts. Freshness got the recrawl, structure won the citation.
Treat the refresh as a package: current numbers, current date, and a tighter 40-90 word lead under the main heading.
That pairing is why this lever is cheap. The recrawl is free attention, and you decide what it finds.
Do not fake the stamp
We tested the dishonest version so you do not have to: six pages, dates bumped, zero content changed. Over eight weeks, citation lift was exactly zero. Engines appear to compare the claimed date against the crawled diff, and a claim with no diff behind it earns nothing. The downside is worse than nothing, because a domain that repeatedly claims freshness it cannot show is teaching engines to discount every date it publishes.
Users run the same check with more contempt. A 2026 stamp on visibly 2023 advice reads as deceit, not diligence.
Honesty here is not just ethics, it is mechanism design. The signal works because it is expensive to fake well and cheap to earn honestly.
Twenty minutes of real revision, one field updated, four surfaces in agreement. The cheapest lever only works because it is true.