Why AI Citations Decay — And What to Do Before You Lose Them

On August 13, 2026, Profound shipped a feature called "Citation Decay." Its job: show, page by page, exactly when a citation peaked and how fast it has been fading since. That's not a minor product tweak. It's one of the most widely used AI-visibility measurement tools admitting something most marketing teams still aren't tracking: an AI citation is never permanent. It has an expiry date, and it arrives faster than most people assume.

A lot of marketing teams still treat AI visibility like a counter that only goes up: once you earn a mention, it's banked, you check the box, and move to the next topic. The data published this summer tells a different story. A citation behaves more like a subscription than a trophy — it renews, or it lapses.

Here's what the recent studies say about how fast that decay happens, why it varies so much between engines, and how to build a refresh routine that reacts to the right signal instead of an arbitrary schedule.

What is citation decay?

Citation decay is the gradual drop in a page's citations after it reaches peak visibility in AI answers. A page moves through a cycle: it first shows up in an answer, gains citation frequency, peaks, then loses ground if nothing on the page changes. Without a refresh, it can eventually all but disappear from AI answers.

Profound's own announcement describes the problem it's solving: most teams still decide what to update "on an arbitrary cadence," so "the page that actually needed updating sits untouched for another quarter" (tryprofound.com, August 13, 2026). The tool calculates, for every URL, its first-cited date, its rise period, its peak, its half-life (the number of days from peak until weekly citations fall to 50% of that peak), and its most recent known citation date.

The idea isn't unique to Profound. It echoes a pattern that several independent studies have confirmed since mid-2026, with hard numbers on exactly how fast that decay moves.

How long does an AI citation actually survive?

A study by Scrunch, run with Stacker across 3.5 million citation events between September 2025 and March 2026, puts the average half-life at 4 to 5 weeks — meaning half of the citations a page earns are gone within about a month. That average hides sharp differences between engines.

The researchers tracked cohorts of sources that first appeared in a given week, then measured how many were still being cited in the weeks that followed (scrunch.com, "The half-life of AI citations"). Broken down by engine:

The same study also found industry-level differences: roughly 5 weeks for insurance, 4.8 for financial services, 4.1 for healthcare, 4.3 for retail and ecommerce. One notable outlier: domains inside Stacker's editorial partner network held on roughly twice as long as average, up to 12.3 weeks on Perplexity.

Why does Perplexity hold on to sources longer than ChatGPT?

The Scrunch and Stacker study does not prove a single cause for Perplexity's longer half-life. Still, the measured gap — 5.8 weeks on Perplexity versus 3.4 on ChatGPT — is large enough to justify different monitoring cadences, because each engine discovers, revisits, and selects sources differently.

Perplexity, built from the ground up around citing sources, appears to hold on to its reference sources longer before swapping them out. ChatGPT, which recomposes its answer more heavily on each query, cycles through sources faster.

What that means in practice: your monitoring and refresh cadence can't be identical if your AI traffic comes mostly from ChatGPT versus mostly from Perplexity. Content that "holds" for five weeks on Perplexity may already have lost half its ChatGPT citations.

It also means a single one-time audit, run once and then filed away, tells you nothing about a page's real trajectory. If AI visibility matters to your acquisition, what matters isn't just whether you're cited today, but for how long, and on what slope.

Is fresh content really cited more, or just updated more often?

A July 2026 Seer Interactive study covering 7,683 pages and 47,097 citations across ChatGPT, Gemini, and Perplexity found that 75% of AI-cited pages had been updated within the past year. But most of that freshness comes from maintenance, not new publishing.

The detail is telling: among 4,124 pages where both a publish date and an update date were readable, 72% looked "fresh" by their last update, versus only 42% if you go by their original publish date. Over a quarter of those "fresh" pages had actually been first published more than two years earlier (Seer Interactive, "Content Recency's Impact on AI Visibility," July 2026). In other words: an older page that's genuinely maintained can stay cited just as well as a brand-new one.

The study also breaks down the share of cited content updated within the past year, by engine:

One more useful data point, for telling a durable citation apart from a one-off spike: pages cited across all four months of the study were only 68% "fresh," with a median update age of 5.6 months. Pages cited in just one month were 86% fresh, with a median age of 1.9 months. An isolated citation spike usually comes from very recent content; a stable, ongoing presence can tolerate a somewhat older page, as long as it's kept up.

Does source rotation mean my content is failing?

No — a share of that rotation is simply how these engines normally operate, independent of your content's quality. A Profound study covering roughly 80,000 prompts per platform found that 40 to 60% of the domains cited for the same query changed within a month, and that 70 to 90% changed between January and July 2025 (tryprofound.com, "AI Search Volatility").

That number puts things in perspective. If a page is losing citations, two very different stories can produce the same downward curve:

That's exactly the distinction Profound's feature is trying to automate: sorting pages by half-life instead of age, to prioritize "pages worth maintaining" instead of whichever happens to be oldest.

How do you spot a page that has passed its peak?

Three signs together point to real decay rather than measurement noise: weekly citation counts drop for several consecutive weeks, no competing page seems to have taken your place in the same answers, and the content hasn't changed since the last uptick. Any one of these signs alone can just be normal variation. Together, they describe a page that needs attention.

Without dedicated citation-decay tooling, a simple method works to start: pick a stable panel of 15 to 20 queries tied to your most important pages, run it manually across the main engines once a week or every two weeks, and note whether each page still shows up, has disappeared, or was replaced by a new competitor. A fixed panel avoids the most common mistake — judging your visibility off a single query tested once, when the natural variability of AI answers makes any one-off check unreliable.

How do you build a refresh routine around decay instead of a calendar?

Replace the arbitrary calendar ("republish everything every quarter") with a trigger based on each page's actual behavior. Four steps are enough to get started, even without dedicated citation-decay tooling.

1. Track citations by page, not just overall brand visibility. A brand-level average can stay flat while one specific page collapses entirely. Citation tracking by page and by engine lets you spot exactly which URL has passed its peak, instead of guessing.

2. Tell decay apart from rotation before you act. Use the table above: a one-week dip isn't a signal. Several consecutive weeks of decline with no identifiable competitor taking your place is.

3. Compare decay speed by page type instead of treating the whole site the same way. A news piece or an event-tied analysis tends to decay faster than a reference guide or a definition page. Prioritizing refreshes by content type, not just by age, keeps you from spending time on pages that decay slowly by nature.

4. Refresh the content, not the date — then verify. An update that counts changes facts, numbers, examples, or an entire section — not just the updatedAt field. Once the change is live, a new audit on the same page confirms whether it actually reopened a citation window. That's the principle behind our Track, Fix, Prove method: track, fix, then prove the effect instead of assuming it.

This routine complements, rather than replaces, the case for publishing on a regular cadence: new content still matters for covering new questions, but maintaining pages that are already cited needs its own discipline, with its own trigger.

For a small marketing team, none of this requires expensive tooling on day one. A simple spreadsheet with your priority pages, a fixed query panel, and a "last citation uptick observed" column is enough to get moving. Dedicated tooling becomes worth it once the number of pages and queries you're tracking outgrows what a weekly manual review can handle.

Citation decay isn't inevitable — it's a signal worth measuring

The GEO market has long treated an AI citation as a one-time goal: earn the mention, then move on to the next topic. The 2026 data shows that's a scheduling mistake. A citation earned in week 3 can have lost half its frequency by week 8, with no alarm going off unless someone is actually watching.

The good news: this pattern is measurable, broadly predictable, and manageable with a simple routine. You don't need to republish your entire site every quarter. You need to know which pages have passed their peak, and act on them before a competitor takes their place.

Run a free AI visibility audit to see where your most-cited pages stand today, and which ones are approaching their decay point.