Your AI citations have a shelf life

Citation decay is the rate at which pages an AI engine cites today stop being cited later, and the published data now says it is fast enough that a citation should be planned as a lease, not an asset.

Most AI visibility reporting still treats a citation as something you win and then hold. Several studies published this year followed the same prompts over weeks or months instead of taking one snapshot. They disagree on the exact rate, partly because they measure different things. They agree on the direction: the set of cited sources turns over, and the turnover is uneven across engines, content types and URL patterns.

#What eleven months did to one vertical

The clearest long-horizon example comes from Green Flag Digital's ChatGPT citation study, which followed the personal finance pages ChatGPT cited most in October 2025. Of the 215 pages analysed, 107 got no citations at all in September 2026, and the group as a whole kept 14% of its citations.

The worst performers had a date in the URL. In the same data, twenty of the 215 pages had 2024 or 2025 in the address. Of those, 80% went to zero, and together they kept 0.3% of their citations against 16% for the rest. One Bankrate page titled with a specific May 2025 date went from 237 citations to zero.

The source mix changed too. In the same study, Wikipedia and Reddit fell from 18% of top-500 citations to none, while bank, card issuer and government pages rose from 2% to 28%. That is one vertical on one engine, and the study does not claim to explain the cause. It does show that a year is long enough for most of a winning set to be replaced.

#What one month does

Shorter windows show the same pattern at smaller scale. Alex Birkett compared the top 500 cited URLs across two snapshots one month apart, from roughly 100 prompts written for an AI recruiting platform. Only 245 of the 500 URLs, or 49%, appeared in both. At the domain level the picture was steadier: 133 of 204 domains, or 65%, remained.

That gap is the useful finding. Domains hold better than URLs. An engine may keep trusting your site while swapping which of your pages it cites, or it may keep citing the same kind of source while swapping which site supplies it.

Birkett's write-up also summarises an earlier Profound study that measured month-to-month change in cited domains for identical prompts: 59.3% for Google AI Overviews, 54.1% for ChatGPT, 53.4% for Microsoft Copilot and 40.5% for Perplexity, rising to 70% to 90% over six months.

#Engines decay at very different rates

The largest dataset so far is SISTRIX's citation drift study: 82,619 prompts and 1,548,213 snapshots across six countries, resampled weekly for 17 weeks between December 17, 2025 and April 8, 2026.

Weekly domain replacement differed by more than an order of magnitude. SISTRIX measured 5% for Google AI Overviews, 56% for Google AI Mode and 74% for ChatGPT Search.

SISTRIX also found structure inside the churn. For 86% of prompts there was a stable core of cited domains with a rotating periphery, and in AI Mode that periphery rotated at 89% per week. ChatGPT's median prompt had zero stable core domains, against about two in AI Mode.

Content type predicted who made the core. In AI Mode, video and YouTube sources held core positions at 24%, Wikipedia at 12%, forums and user-generated content at 3%, and news and media at 1.4%.

The AI Overviews figures from SISTRIX and Profound look contradictory, at 5% weekly against 59.3% monthly. They are not measuring the same thing. The prompt sets, periods and change definitions differ, so the two numbers should not be placed on one chart. Each is internally consistent, and each says the engines behave differently from one another.

#Read decay studies before quoting them

Decay is easy to overstate. Digital Applied took apart one widely shared study of Australian insurance citations, whose headline was that 57.2% of domains were cited in exactly one of seven sampled months. The seven months were not consecutive, with April and June missing, so a domain cited only in those months could not be observed at all. The survival table summed to 10,446 domains, not the 28,725 in the headline, which leaves the denominator unclear. The data covered one country and one vertical.

The general lesson applies to any decay figure, including your own:

• Absence of observation is not absence of citation. Gaps in sampling look like losses.
• Check the unit. URL churn, domain churn and citation-count loss are three different measures.
• Check the prompt set. A study of news-heavy prompts will decay faster than one of product or definitional prompts, because news sources barely hold core positions.

#What decay rewards

Keep the year out of the URL. The Green Flag result is the most actionable number here. A page whose address names a year announces its own expiry. Put the year in the title and the visible text, keep the URL stable, and update the page in place with an accurate dateModified rather than publishing a new dated URL each year.

Compete for the core, not the periphery. If most prompts have a small stable core and a fast-rotating edge, one appearance at the edge says little about next month. The content types that most often held core positions in SISTRIX's data were video, big tech platforms and Wikipedia, while news and forums rarely did. Our reading, not a measured result: a page that is the primary record for a fact has a better chance of staying than one of many summaries of it.

Refresh on a schedule. Birkett also cites AirOps data on 4,000+ ChatGPT-cited pages: 35% had been updated within three months and 53% within six. That is correlation, not proof that updating causes citation, but it describes the competition. A page left untouched for a year is up against pages that were not.

Measure over windows, per engine. A single audit is a snapshot of a moving set. With ChatGPT turning over most of its cited domains in a week, one run tells you little. Report citation share across a rolling window, keep each engine as its own series, and read one week's gain or loss as noise until it persists. We covered why visibility is a distribution and how many reruns are worth paying for in earlier posts.

Watch URLs and domains separately. When a cited URL drops out, check whether another page on your domain replaced it before treating it as a loss. Domain-level presence holding while URLs rotate is a different problem from the domain disappearing.

#The short version

AI citations turn over. Google AI Overviews changes slowly, AI Mode and ChatGPT change most of their cited sources every week, and over months even top-cited pages can fall to zero. The pages that last tend to be primary sources with stable URLs that are kept current. Plan content and reporting as if every citation has to be re-earned, because the data says it does.