Content decay in the AI-search era: when and what to refresh

The one-line version first: content decay is the slow, quiet slide of a page that used to perform, and in the AI-search era it costs more than it used to, because a model will happily quote your old numbers as fact in an answer that never links back.

I catch myself doing it too. A post sits at a comfortable position for a year, nobody on the team opens it, and one day I read it the way a stranger would — and there it is, a figure that stopped being true a while ago, sitting in the second paragraph with no date next to it. That’s decay: rankings drift down, traffic thins out, the facts go stale, and AI answer engines stop citing you. It is not a penalty and rarely a single event; it is entropy. So what is worth watching, when is it time to act, and how deep should the fix go — refresh, rewrite, or consolidate?

The four decay signals

Decay shows up as a sequence, not a single alarm, so I read the signals together rather than reacting to whichever one fired first.

Signal	Where you see it	What it means
Rank slip	Search Console average position	A competitor updated, or the query moved on
Traffic decline	Clicks/impressions month over month	Ranking loss is now costing you sessions
Stale facts	The page itself	Prices, dates, versions, “as of 2024” no longer true
Lost citations	AI-answer visibility checks	Models cite someone fresher — or quote your old figure

The first two are the classic SEO signals every analytics tool surfaces. The last two are the ones teams miss, and they are exactly the ones that matter most for generative search — probably because nobody’s dashboard has a column for them. A page can hold its blue-link ranking while quietly disappearing from AI answers — or, worse, keep appearing with a number you corrected a year ago. Which of the two is worse? Honestly, the second: the first is invisible, the second is actively misinforming people under your name.

Why stale numbers hurt GEO more than SEO

In classic search, a stale statistic costs you a sliver of trust on one page that one reader is actively looking at, and that reader can see the “updated 2023” stamp and discount it themselves.

Generative engines strip that safety net. A model extracts the figure, drops the context, and repeats it as a flat assertion inside an answer that often shows no source at all. Your outdated “average cost is $40” becomes the AI’s confident answer to thousands of people who never see the URL, never see a date, and have no way to sanity-check it. One stale number doesn’t decay one page — it propagates into every answer that ingested you! (And there is no button anywhere to tell the model it was corrected — at least none I’ve found.)

The refresh trigger

Not every dip deserves a rewrite, and I’ve wasted enough afternoons on pages that didn’t. The trigger I use is simple: act when two or more signals fire on a page that still matters. Here is the checklist I run before touching anything.

• Average position dropped 3+ places over the last 90 days.
• Clicks are down month over month for two consecutive months.
• The page contains a number, date, price, or “current as of” claim older than your niche’s shelf life.
• The page has lost — or never earned — AI-answer citations for its target query.
• The topic still has demand (impressions haven’t gone to zero; the query still exists).
• The URL is worth saving (it has backlinks, brand equity, or internal-link authority).

If the last two are false, decay is a signal to retire or redirect, not refresh. Is anyone still asking this question at all? If not, don’t spend effort re-inflating a page for a query nobody asks anymore — that’s not maintenance, that’s taxidermy.

Refresh vs rewrite vs consolidate

Once a page trips the trigger, one decision remains: how deep to go.

Action	When	Effort	Keeps the URL?
Refresh	Structure is sound, details aged	Low	Yes
Rewrite	Angle/format is broken, thin, or off-intent	High	Yes
Consolidate	Several thin pages split the same query	Medium	One survives, others redirect

• Refresh is the default and the cheapest win. Update the numbers, re-verify every claim, add the section the query grew to expect, bump the visible date. The page keeps its history and its links.
• Rewrite when editing costs more than starting over — the angle is wrong, the structure fights the query, the writing shows its age. Same URL, new body.
• Consolidate when the site’s own pages are the competition. Three thin posts each ranking on page two for one query will keep decaying against a single competitor’s deep guide. Merge them into one canonical page, redirect the rest, and inherit their combined equity — a move worth auditing the content-decay and refresh queue for before anything new gets written.

What a refresh actually touches

A real refresh is boring and specific, and I go through a page in this order:

• Every statistic re-sourced to the current year, with the source visible.
• Dead links fixed, screenshots and versions updated.
• One new section answering the sub-question the query grew to include.
• Structured, extractable phrasing so answer engines can lift a clean passage.
• A genuine, honest updated date — because you genuinely updated it.

That last point is the whole game for GEO: a fresh date on genuinely fresh facts is what pulls a page back into the answer. Could a date bump alone do it? Perhaps briefly — and then the engine reads the same stale figure under a new stamp, which is exactly the pattern the warning above is about.

Stale facts / numbers 55%

Fresher competitor 30%

Structure not extractable 15%

source: internal audits, 2026

The split is illustrative, but the lesson holds across audits: most lost AI citations trace back to a fact that went stale, not to a page that was never good. Decay is usually fixable, which is oddly reassuring — the page was fine, it just kept living in the year it was written!

In short

• Content decay is entropy, not a penalty — rankings, traffic, facts, and citations all slide.
• Watch four signals; act when two or more fire on a page that still has demand and equity.
• Stale numbers hurt GEO more than SEO because a model repeats them as sourceless fact.
• Refresh a sound page, rewrite a broken one, consolidate pages that compete with each other.
• A real refresh updates the facts and the structure — then the honest date does the pulling.

And the post that made me write this? It’s on the refresh queue — second paragraph, no date next to it.

Sources

• Aggarwal et al., “GEO: Generative Engine Optimization,” KDD ‘24 — arXiv:2311.09735: citation likelihood depends on extractable, credible passages — which stale facts undermine.
• Google Search Central — “A guide to Google Search ranking systems” — documents the freshness systems that reward recency for queries that warrant it.
• Google Search Central — “Creating helpful, reliable, people-first content” — the self-assessment questions that flag out-of-date, unmaintained pages.