AI Forgets You in 90 Days: Content Freshness as an AEO Signal

The new front page forgets fast

For two decades, an authoritative page could rank for years on accumulated link equity and age. "Old but good" was a viable strategy. Answer engines broke it. When an AI assembles a response, it strongly favors recent sources, it is optimizing to give the user a current answer, and a page last touched in 2023 reads as a risk of being out of date. The result is a recency bias far stronger than classic SEO's mild freshness signals.

The numbers make it concrete. Per Salespeak's citation-freshness study, about 50% of AI-cited content is less than 13 weeks old, and roughly half of Perplexity's citations come from the current year alone. AI-cited content is on average 25.7% fresher than the content ranking in traditional organic Google results, a quantified gap between what wins in AEO and what wins in SEO. Freshness is not a tiebreaker in answer engines; it is closer to a gate.

The decay timeline

Citation decay follows a measurable curve, and knowing it tells you when to act. Drawing on the timelines in AuthorityTech and Ten Speed's AEO freshness guide:

• 0-90 days: peak citation eligibility. Newly published or substantively updated content is most likely to be pulled into answers.
• 90-180 days: decay begins. Without updates, the page starts losing retrieval priority to newer competitors covering the same queries.
• 180-365 days: significant decay. Only high-authority domains with strong entity signals reliably maintain citation rates; everyone else fades.
• 365+ days: for most pages, effectively out of the consideration set on time-sensitive queries unless refreshed.

This means a publish-and-forget content library, the thing most companies have, is a depreciating asset in AEO terms. The blog post that drove AI citations and high-intent referrals in Q1 is quietly aging out by Q3. And because the loss shows up nowhere in your keyword rankings (the page may still rank in classic search), it is invisible unless you specifically monitor it.

Freshness means substance, not a touched timestamp

Here is the critical nuance that separates real freshness work from gaming: answer engines evaluate what you changed, not just that you changed something. Bumping a dateModified field or swapping one word does not fool a model that reads the content. Per the same Salespeak analysis, AEO freshness is about accuracy, recency, and substantive updates, real improvements to the content, not cosmetic timestamp churn.

That reframes the work. The goal is not to touch every page constantly; it is to keep your high-value pages genuinely current, updated statistics, new developments, corrected facts, expanded sections, refreshed examples. A page that says "as of 2026" and actually reflects 2026 reality is what clears the bar. A page with a 2026 timestamp and 2023 content does not, and risks looking deceptive.

Find what is decaying before it disappears

You cannot refresh what you cannot see. The first step is an inventory of your content by age and value, so you can prioritize the high-traffic, high-intent pages crossing the 90-day decay line. A scan of your content graph that joins last-meaningful-update dates with traffic and citation value turns "we have a thousand posts" into a ranked refresh queue.

# Build a refresh queue: pages crossing the decay line, by value
from datetime import datetime, timezone

def days_since(iso_date):
    d = datetime.fromisoformat(iso_date).replace(tzinfo=timezone.utc)
    return (datetime.now(timezone.utc) - d).days

def refresh_score(page):
    age = days_since(page['last_substantive_update'])
    # decay multiplier ramps after 90 days, steep after 180
    if   age < 90:  decay = 0.0
    elif age < 180: decay = (age - 90) / 90        # 0 -> 1
    else:           decay = 1.0 + (age - 180) / 180  # > 1, accelerating
    # weight by what the page is worth to you
    return decay * (page['monthly_traffic'] + 5 * page['ai_citations'])

pages = load_content_inventory()
queue = sorted(pages, key=refresh_score, reverse=True)
for p in queue[:20]:
    print(f"{days_since(p['last_substantive_update']):4}d  "
          f"score={refresh_score(p):8.0f}  {p['url']}")
# Work the top of this list every sprint. Substantive updates only.

The distinction in the data model matters: track last_substantive_update, not dateModified. A timestamp that moves on every trivial republish is useless for prioritization. You want to know when the content last actually changed in a meaningful way, because that is what the answer engines are effectively measuring.

Operationalize freshness as a recurring program

Freshness is not a project; it is a rhythm. The teams that win AEO treat their top content like a maintained product, not a published artifact. A workable cadence:

• Tier your library. Identify the 10-20% of pages that drive most of your AI citations and high-intent traffic. These get active maintenance.
• Set review SLAs. Tier-1 pages reviewed every ~90 days, before the decay line; lower tiers on longer cycles or as topics move.
• Update substantively. Each refresh adds real value, new data, new developments, corrected or expanded content, then honestly updates the date.
• Monitor citations, not just rankings. Watch whether your refreshed pages re-enter AI answers; that is the metric this work moves, and it will not show in a rank tracker.

The bottom line

Answer engines forget you fast. With half of AI citations going to content under 13 weeks old and decay beginning at 90 days, a publish-and-forget library is a depreciating asset, and the loss is invisible because it does not show in keyword rankings. The work is not timestamp games, which backfire, but genuine substantive updates to your highest-value pages on a recurring cadence. Inventory your content by real last-update date and value, prioritize the pages crossing the decay line, refresh them with actual new substance, and measure citations rather than rankings. In AEO, currency is a gating signal, and the page you wrote last year is already aging out of the answers your buyers are reading today.