What does the 11-day AI citation half-life actually mean?

It means the median page fell to half its peak citation share 11 days after peaking, not that a citation ends after 11 days. Profound measured this across 883,000 pages and seven AI engines, using page-level citation share smoothed over a trailing 14 days . That smoothing caps a single burst at a 14-day half-life. The finding that 78% of pages halved within two weeks is partly built into the method. Model updates were not controlled. The same study found pages cited by five or more engines 5.5 times as likely to hold half their peak at day 56. The figure describes decay from a peak, and it should always be quoted with that unit.

Is it true that only 10.6% of AI citations survive a month?

The figure is real, but its denominator makes survival look about half as large as it is. Digital Authority Partners ran 30 queries on five engines in three waves 14 days apart and found 119 of 1,127 unique URLs in all three . The 1,127 counts every URL seen in any wave. A URL first cited in the second or third wave could never appear in all three. Against the 530 URLs cited in wave one, survival is 22.5%, and the study's own 28-day retention rate averages 33%. It is one agency study with 30 queries and one run per wave, so the result is provisional.

Why do AI citations change when the AI model changes?

A new model can choose different sources for the same query, so a model switch reshuffles citations independently of time. When Gemini 3 became the default AI Overviews model on 27 January 2026, 42.4% of previously cited domains disappeared across 100,000 US keywords . Almost all of the loss came from the long tail; among the 500 most-cited domains, only one dropped out. Sources per AI Overview rose from 11.55 to 15.22 at the same time. Three of the six decay datasets span that switch without controlling for it. Part of what they report as citation decay is a one-off model change.

Go back

Citations & Retrieval

How long do AI citations last? Most rotate within days, a small core stays for months

Q: How long do AI citations last?

It depends on what is counted, and published answers range from under one day to longer than 17 weeks. Rerunning identical prompts daily, GetMentions found 69% of an answer's cited domains changed overnight . Tracking source cohorts across a live prompt mix, Scrunch and Stacker put the average half-life at about 4.5 weeks . Sampling fixed prompts weekly, SISTRIX found the brand's own domain cited in all 17 weeks for 43% of brand queries in Google AI Mode . The common pattern is a fast-rotating periphery around a small core of sources that keep returning. A single half-life blends the two and describes neither.

Q: Do you need to refresh content every month to stay cited by AI?

No study has tested a monthly refresh cycle. The turnover datasets measure churn, not the effect of updating a page. AI-cited content averages 1,064 days old, about a quarter fresher than Google's organic results . One industry report found pages not updated for a quarter over three times more likely to lose AI citations, a correlational Tier C finding . Model weights move slower still: the gap between knowledge cutoff and release runs 2.4 to 4.6 months for recent models . The evidence supports keeping evergreen pages current on a quarterly rhythm, not rebuilding visibility every four weeks.

Six datasets put the AI citation half-life anywhere from under one day to longer than 17 weeks, depending on what is counted. Turnover is real, but survival flattens around a recurring core: in Google AI Mode the brand's own domain stayed cited all 17 weeks for 43% of brand queries.

A reading circulating among GEO practitioners holds that AI citations have a half-life of about four weeks, so AI visibility must be re-earned every month. The evidence rates that claim Overstated. Turnover is real and replicated, but the reported half-life depends on what is counted. On fixed prompts rerun daily, 69% of an answer's cited domains changed overnight, a half-life under one day Agarwal · GetMentions AI · 2026. On weekly samples, the brand's own domain stayed cited in all 17 weeks for 43% of brand queries Beus · SISTRIX · 2026. Survival flattens around a core of recurring sources, and no study has tested a monthly refresh. The sections below take the six datasets in turn, then the model updates that sit inside every decay figure.

AI citation turnover is real and replicated across engines and years

The direction of the claim is among the best-replicated findings in generative engine optimization (GEO). Across roughly 80,000 prompts per platform, Profound measured domain drift from June to July 2025 of 59.3% on Google AI Overviews and 40.5% on Perplexity Blyskal · Profound · 2025. Its January-to-July comparison reached 70–90%.

Weekly data agrees. SISTRIX resampled 82,619 prompts weekly for 17 weeks across six countries Beus · SISTRIX · 2026. Google AI Mode replaced 56% of cited domains each week and ChatGPT Search 74%.

Daily data agrees too. GetMentions reran 2,398 identical queries once a day for seven days in June 2026, across four engines Agarwal · GetMentions AI · 2026. Overnight, 69% of a typical answer's cited domains changed, from 44.4% on Perplexity to 88.3% on Gemini.

The reported AI citation half-life runs from under a day to over 17 weeks

The four-week figure traces to a joint Scrunch and Stacker study of 3.5 million citation events from September 2025 to March 2026 Scrunch; Stacker · 2026. It put the average source half-life at about 4.5 weeks, from 3.4 on ChatGPT to 5.8 on Perplexity. The study follows cohorts of source domains across a live prompt mix, not reruns of one prompt. A source can decay there because its prompts stopped being asked.

Its secondary finding needs a label. Citations on Stacker partner domains lasted roughly twice as long Scrunch; Stacker · 2026. Stacker sells that distribution, so the durability result is directional.

Fixed-prompt designs give other numbers. GetMentions holds the prompt, location and session constant, and its overnight churn implies a half-life under one day Agarwal · GetMentions AI · 2026. SISTRIX samples fixed prompts weekly and found AI Overviews replaced only 5% of cited domains a week Beus · SISTRIX · 2026. A figure that moves this far with the unit describes the ruler more than the engine.

Profound's 11-day AI citation half-life measures decay from peak, not lifespan

The sixth dataset has already mutated in circulation into "an AI citation lasts 11 days". Profound tracked 883,000 pages through 1.19 million page-engine lifecycles on seven engines from September 2025 to September 2026 McCain · Profound · 2026. The median page fell to half its peak citation share 11 days after peaking.

That is peak to half-peak, not the time a page stays cited. Profound smooths citation share over a trailing 14 days, which caps a single burst at a 14-day half-life McCain · Profound · 2026. The headline that 78% of pages halved within two weeks is therefore partly definitional.

The misreading already shapes practice: practitioner threads report moving from biweekly to weekly tracking on the premise that citations end after 11 days. The same study points the other way on persistence. Pages cited by five or more engines were 5.5 times as likely to hold half their peak at day 56, 6.8% against 1.2% McCain · Profound · 2026.

AI citation survival flattens around a core, often the brand's own domain

The persistent layer appears in weekly data as a core. In Google AI Mode, 86.5% of prompts held a stable core of one to five domains, while the remaining domains rotated at 89% a week Beus · SISTRIX · 2026. On brand queries the brand's own domain is often that core, cited in all 17 weeks for 43% of them.

A model switch shows the same split. When Gemini 3 took over Google AI Overviews, nearly all the domains that vanished came from the long tail Deda · SE Ranking · 2026. Among the 500 most-cited domains, one dropped out.

The core is a recurring pool rather than a fixed slot. Over seven daily runs, only 0.4% of sources on Gemini and 11.1% on Perplexity were cited every day Agarwal · GetMentions AI · 2026. ChatGPT turned over 79.2% of domains overnight, which makes it the engine where the re-earning reading comes closest to accurate.

Model updates and rerun noise sit inside every AI citation decay figure

None of the half-life studies separates time decay from model changes. Gemini 3 became the default AI Overviews model on 27 January 2026 and removed 42.4% of previously cited domains across 100,000 US keywords Deda · SE Ranking · 2026. The Scrunch, Profound and SISTRIX windows all span that date, and Profound states that model updates were not controlled McCain · Profound · 2026.

That qualifies the most-quoted stability figure. SISTRIX measured its 5% weekly AI Overviews drift from December 2025 to April 2026, across the switch Beus · SISTRIX · 2026. A 17-week average can absorb a one-off shock, so the 5% describes a stable head of sources, not immunity to model swaps.

Model weights run on a slower clock. Graphite puts the gap between knowledge cutoff and release at 2.4 months for Anthropic and 4.6 for OpenAI Druck et al. · Graphite · 2026. It treats that gap as the minimum before new content reaches the weights. All six datasets measure the fast clock of search-fed citations, which moves in days.

Part of what reads as decay is rerun noise. Typical 95% confidence intervals on citation share span 3 to 6 points, and week-over-week movement often sits inside that band Sielinski · IQRush · 2026. The library article "You can't measure AI search with one number" sets out why one run cannot separate the two.

No study has tested a monthly AI visibility refresh

The operational half of the claim has no direct evidence. None of the six datasets refreshed a page and then measured whether its citations returned. The freshness evidence points to a slower rhythm: AI-cited content averages 1,064 days old, against 1,432 for organic Google results Law · Ahrefs · 2025.

The strongest refresh signal in the library runs on a quarter. Pages not updated for a quarter were over three times more likely to lose AI citations AirOps · 2026. That is a Tier C industry report and correlational. It supports keeping core pages current, not a four-week cycle, and a matched refresh experiment would settle the question.

FAQ
Frequently Asked Questions

Industry report from AirOps with Kevin Indig, drawing on millions of citation datapoints across ChatGPT, Google AI Overviews, AI Mode, Gemini, and Perplexity. Findings are organised around freshness, on-page structure, schema use, user-generated content, off-site mentions, and visibility stability, with specific percentage gaps tied to each signal.

airops.com·Accessed 20.05.2026

About the author Max Ackermann

Max Ackermann is founder and Managing Director of info.link, the product data platform that makes brands visible in AI search and connects every physical product to the web through GS1 Digital Link. He writes about AI search and generative engine optimization (GEO), AI-powered commerce, and how brands can structure product data for ChatGPT, Gemini, Perplexity, and retailer AI assistants like Amazon Rufus. For the past two years he has built the pipelines that put structured product data into AI answers, and run the experiments that test what actually moves AI citations.

Max has 20+ years of experience building digital products and businesses. He previously led McKinsey's Corporate Venture and Design teams across Europe, and as Managing Director of a leading US digital agency he built platforms with Nike, Google, Meta, and Airbnb. He founded the UX Design program at Central Saint Martins College, University of the Arts London, and is a Fellow of the UK's Higher Education Academy. Based in Hamburg, he works closely with GS1 on Digital Link adoption; info.link is headquartered in Hamburg and Berlin and counts GS1 Germany among its investors.

Follow Max on LinkedIn.