Median AI citation half-life is 4.5 weeks. Only 33% of cited URLs survive 28 days. The 2026 churn data.
The median AI citation half-life is 4.5 weeks. Scrunch and Stacker tracked 3.5 million citation events across six AI platforms between September 2025 and March 2026, applied survival-curve analysis with 200 bootstrap resamples, and found that half of any given week's cited sources fall out of AI answers within 32 days. On ChatGPT the figure drops to 3.4 weeks. Your team publishes a page, wins a citation, screenshots it for the monthly report, and by the next report that citation is gone.
A second study puts a harder number on it. Digital Authority Partners tracked 1,127 unique URLs that ChatGPT, Perplexity, Gemini, Copilot, and Google AI Overviews cited across 30 queries in six industries. Six weeks and three collection waves later, the engines still cited 119 of them, or 10.6%.
Here is what the churn data says, which platforms rotate fastest, and what refresh cadence keeps you in the answer.
AI citation half-life measures how many weeks pass before 50% of a cohort of cited sources stops appearing in AI answers. Scrunch and Stacker measured a 4.5-week median across 120,000+ domains, eight industries, and six platforms in their March 2026 analysis. The metric borrows from cohort survival analysis: take every source cited in week one, then count how many survive into week two, week three, and so on.
The metric matters because most B2B teams still report AI visibility as a state rather than a flow. A dashboard screenshot showing 14 citations tells you nothing about whether those 14 will exist next month. Semrush's 2026 AI Visibility Index, built on 126 million U.S. AI search prompts from January through April 2026 across 22 industries, found that 45% of marketing leaders cannot measure their brand's presence in AI answers at all. Among those who can, few measure it twice.
Quarterly reporting worked for rankings because rankings held for months; citations turn over inside a single reporting cycle. Track this metric for your brand across all five engines → nobori.ai
AI engines replace two-thirds of their cited URLs within 28 days. Digital Authority Partners measured an average Citation Retention Rate of 33% across five AI platforms over four weeks, drawing on 630 data points from 30 queries spanning six industries. The researchers also pulled Google and Bing organic top-20 results as a benchmark.
The distribution is worse than the average suggests. Of every query-and-platform pair the researchers tracked, 87% fell below a 70% stability threshold. Nearly a quarter, 24%, showed complete turnover, with zero overlap between the sources cited in wave one and those cited in wave three.
SISTRIX reached a compatible conclusion from a much larger sample. Its AI Research Index analyzed 82,619 qualified prompts and 1,548,213 snapshots across six countries and 17 weeks, from 17 December 2025 to 8 April 2026. Google AI Mode swapped 56% of cited domains every week. ChatGPT Search swapped 74%. At the URL level rather than the domain level, ChatGPT's weekly churn climbed to 85%.
ChatGPT rotates sources faster than any other engine, and Perplexity holds them longest. The gap between the two runs close to 70%.
| Platform | Citation half-life (Scrunch/Stacker, Mar 2026) | 28-day retention rate (Digital Authority Partners, Jun 2026) |
|---|---|---|
| ChatGPT | 3.4 weeks | 31% |
| Google AI Mode | 4.3 weeks | Not measured |
| Google AI Overviews | 4.7 weeks | 27% |
| Google Gemini | 4.8 weeks | 11% |
| Microsoft Copilot | Not measured | 34% |
| Perplexity | 5.8 weeks | 44% |
Gemini posted the weakest retention of any engine at 11%. Digital Authority Partners recorded Gemini and AI Overviews as 100% volatile in their sample: not one query held its sources steady across the full 28 days.
Perplexity's durability changes the ROI math on expensive content. A citation that survives 5.8 weeks accumulates roughly 70% more buyer impressions than one that survives 3.4 weeks, which makes Perplexity the better home for research reports, original datasets, and anything with real production cost behind it. See which engines hold your citations and which drop them → nobori.ai
A 33% retention rate means you need to win three citations to keep one. Run the arithmetic on a mid-sized B2B content program: you publish 12 pages a quarter, earn citations on eight of them, and report eight wins to your leadership team. Four weeks later, at 33% retention, between two and three of those citations still exist. The other five or six answer the same buyer question with a competitor's page.
Referral volume compounds the exposure. Previsible's 2026 State of AI Discovery Report analyzed 6.77 million LLM-driven sessions across 166 GA4 properties from November 2024 through May 2026 and recorded 9.9x growth in 19 months, reaching 644,478 monthly LLM sessions by May 2026. ChatGPT carried 92.4% of trackable standalone AI referral traffic, up from roughly 84% in December 2025.
Read those two findings together. The channel sending you the most AI traffic is also the channel that discards your sources fastest. A brand concentrating its AI visibility on ChatGPT is building on the least stable surface available.
AI Overviews split into two populations, and the average hides both. SISTRIX measured 5% weekly domain churn in AI Overviews, the lowest of the three platforms it tracked. Break the sample apart and the picture sharpens: for 53% of prompts, not a single cited source changed across all 17 weeks. For 28%, sources changed occasionally. For the remaining 19%, sources drifted as hard as AI Mode, at 46% weekly churn-in.
AI Overviews cite 11 domains on average, and eight of those stay put. SISTRIX verified this is not a measurement artifact: for 87% of the stable prompts, the generated answer text changed week to week while the cited domains stayed identical. Google rewrote the sentences and kept the sources.
ChatGPT works the opposite way. It cites 3 to 4 domains per response, and the median prompt contains no domain that appears in all 17 weeks. Stable core structures exist there, but they are exceptions. AI Overviews concentrate their stability in a majority of prompts where the source set never moves, so a domain that gets in tends to stay in. ChatGPT draws from a pool that reshuffles weekly, so few domains hold position on any prompt.
Your own domain survives more often than anything else, and still not reliably. SISTRIX found that for branded queries in Google AI Mode, the brand's own domain appeared in all 17 weeks just 43% of the time. It held for at least 80% of the weeks in 66% of cases. Branded queries drift about 20% less than average, at 45% weekly churn-in versus 56%, which makes them steadier without making them safe.
The interesting movement happens next to you. Of the 12 to 15 other domains cited alongside a brand in AI Mode, 70% change every week. When a buyer asks about your product by name, the answer names you and then reassigns the surrounding real estate weekly. This week a review site sits beside you. Next week a competitor comparison page does.
SISTRIX's example: for "bbc newsreaders female," bbc.co.uk appeared in 13 of 17 weeks. Its two co-cited domains in week one were both gone by week 17, replaced by different sites. Watch who gets cited beside you week over week → nobori.ai
Video survives citation drift better than any other content type, and news survives worst. SISTRIX classified cited domains by type and measured how often each type entered the stable core of an AI Mode answer.
| Domain type | Median appearance rate | Core rate (stays cited) |
|---|---|---|
| Video (YouTube) | 53% | 24% |
| Big Tech (Google, Apple, Microsoft) | 41% | 16% |
| Wikipedia | 24% | 12% |
| Marketplace (Amazon and equivalents) | 18% | 10% |
| Forums and UGC (Reddit and equivalents) | 12% | 3% |
| News and media | 12% | 1.4% |
A 1.4% core rate for news means editorial coverage functions as a one-way ticket. AI engines cite the article, then drop it the following week. Publishers who build an AI strategy on news volume are paying for placements that expire within a week or two.
The evergreen signal runs deep in the same data. In AI Mode, 85% of core URLs qualified as evergreen against 77% of peripheral ones, and product and shop pages dominated the core while guides and news articles rotated through. That reinforces what Nobori's analysis of why AI engines cite product pages over blogs found on citation share: structured page types win more citations, and the drift data shows they also keep them longer.
Earned editorial distribution roughly doubles citation half-life. Content syndicated through the Stacker editorial network held citations for close to 10 weeks against 4.5 weeks for non-network domains, a 2.1x durability advantage that held across all six platforms and all eight industries in the study. The strongest pairing in the dataset was Perplexity plus editorial distribution at 12.3 weeks, more than 3x an average non-network ChatGPT citation.
The mechanism is structural rather than reputational. When your claim lives on one domain, the model has one retrieval path. If that path drops below a relevance threshold during a refresh cycle, the citation vanishes. When the same claim appears across dozens of trusted editorial domains, individual sources rotate out while the underlying information stays above the citation threshold because it keeps showing up somewhere credible.
Industry results narrow the decision for B2B teams:
| Industry | Half-life, non-network | Half-life with editorial distribution | Weeks gained |
|---|---|---|---|
| Real estate | 4.2 weeks | 10.6 weeks | +6.4 |
| Financial services | 4.6 weeks | 10.4 weeks | +5.8 |
| Retail and ecommerce | 4.1 weeks | 9.1 weeks | +5.0 |
| Marketing and advertising | 4.5 weeks | 8.9 weeks | +4.4 |
| Insurance | 4.8 weeks | 8.8 weeks | +4.0 |
| Healthcare | 4.0 weeks | 8.6 weeks | +4.6 |
| Tech and SaaS | 4.4 weeks | 7.0 weeks | +2.6 |
Tech and SaaS gains the least from editorial distribution, at +2.6 weeks, and still improves durability by 59%. Nobori's guide to off-page AEO and third-party mentions covers where to place those placements.
Recency governs which pages enter the retrieval pool in the first place. Seer Interactive published updated research in July 2026 that layered citation data onto its earlier freshness work, examining pages cited in non-branded LLM answers for four brands: a national pet retailer, a vacation rental marketplace, a retail energy provider, and a commercial bank. Close to 65% of log hits went to content published within the previous year, and 89% of the content LLMs surfaced was published between 2023 and 2025.
Seer made one methodological change worth copying. The 2026 study compared each page's last-updated date against its original publish date, treating the two as different signals. Teams that update a 2023 page and leave the original publish date visible are sending a stale signal on a fresh page.
The effect varies by category. Finance and travel queries showed the sharpest recency demand. Energy and DIY topics still surfaced older, well-built pages. B2B software sits closer to the finance end of that range, which means your refresh cadence needs to be tighter than a general content calendar would suggest.
No. AI citations and organic rankings now operate as separate systems with separate decay curves. Digital Authority Partners captured Google and Bing organic top-20 results alongside their AI citation waves and found that 60% of AI-cited URLs did not rank in the top 20 for the same query on either engine. Moz analyzed 40,000 queries in 2026 and found 88% of Google AI Mode citations sitting outside the organic top 10.
Cross-platform overlap is thin even inside Google's own products. SISTRIX calculated a Jaccard index of 0.17 between AI Overviews and AI Mode citations for identical prompts, meaning the two surfaces cite different domains 83% of the time. Between AI Mode and ChatGPT the overlap drops to 0.125.
The practical consequence: a stable Google ranking gives you no read on citation durability, and a citation on one AI surface predicts almost nothing about the others. Nobori's breakdown of the B2B AI visibility gap covers the ranking-to-citation disconnect in detail.
Google AI Mode churns at nearly the same rate everywhere. SISTRIX measured weekly churn-in between 54% and 59% across all six countries it tracked, with no trend toward stabilization over 17 weeks.
| Country | AI Mode churn-in per week | AI Mode retention per week | Domains per answer |
|---|---|---|---|
| United States | 54% | 49% | 16.1 |
| United Kingdom | 54% | 49% | 15.7 |
| Germany | 56% | 46% | 13.5 |
| Spain | 56% | 44% | 11.8 |
| France | 57% | 43% | 12.3 |
| Italy | 59% | 43% | 13.7 |
ChatGPT behaves differently. Its churn swung from 42% in France to 60% in the UK to 74% in Germany, a spread SISTRIX attributes to uneven crawl coverage across languages. Teams running multi-market AI visibility programs should expect a stable Google surface and an unpredictable ChatGPT one.
Google's newest markets inherit the same volatility. AI Overviews and AI Mode launched in France on 22 July 2026, closing the last major market gap, which means French B2B teams are now competing inside a surface that reshuffles more than half its sources weekly.
Match your refresh interval to each platform's half-life, then set it slightly tighter.
| Platform | Half-life | Refresh cadence to stay in rotation |
|---|---|---|
| ChatGPT | 3.4 weeks | Every 2 to 3 weeks |
| Google AI Mode | 4.3 weeks | Monthly |
| Google AI Overviews | 4.7 weeks | Monthly to 5 weeks |
| Google Gemini | 4.8 weeks | Monthly to 5 weeks |
| Perplexity | 5.8 weeks | Every 6 weeks |
A refresh does not mean a rewrite. Swapping in a current statistic from a 2026 source, updating an example, correcting a claim that has aged, and moving the visible last-updated date all reset the freshness signal. Scrunch and Stacker's data shows metadata freshness carrying more weight than most on-page factors.
Pick your priority surface before you build the calendar. SISTRIX found that AI Overviews, AI Mode, and ChatGPT cite different domains more than 80% of the time for the same question, and they reward opposite content profiles: product and shop pages in AI Mode, documentation and institutional sources in ChatGPT, where core domains scored 16/20 on E-E-A-T signals against 14/20 for peripheral ones. Optimizing all three at once spreads effort thin. Choose one target and monitor the rest.
Report presence over time instead of citation counts at a point in time. On a surface that swaps 56% of its sources weekly, one citation tells you about one week. Three metrics carry more signal:
Read third-party studies with the measurement unit in view. SISTRIX counts distinct domains with consistent source attribution and reports 3 to 4 sources per ChatGPT response. Semrush counts retrieved sources across 126 million prompts and reports an average of 15 for ChatGPT and 3 for Gemini. Neither is wrong. They count different objects, and stacking their numbers side by side produces nonsense.
Four moves, in order:
Nobori's chunk-first content framework covers the page-level structure that gets you into the citation set in the first place. Retention work only pays off once you are in it.
SISTRIX found no stabilization across 17 weeks and six countries. Churn rates held between 54% and 59% from December 2025 through April 2026 with no downward trend, which points to citation drift as a structural property of retrieval-augmented systems rather than a launch-phase wobble.
Two 2026 developments could shift the picture. The European Commission adopted a decision on 16 July 2026 under Article 6(7) of the Digital Markets Act requiring Google to give third-party AI services access to its data, and issued an €890 million DMA fine over self-preferencing in Search and the Play Store. More AI services drawing on Google's index would widen the source pool and likely raise churn rather than lower it. Meanwhile Google's new Search Console controls let publishers exit AI answers while staying indexed, which shrinks the eligible pool on the pages that opt out.
Plan for volatility to persist. Put a refresh date on every page you want cited, audit retention monthly, and treat the calendar as maintenance rather than production. Teams running that discipline through 2027 will hold visibility while their competitors cycle in and out.
What is AI citation half-life? Citation half-life is the number of weeks it takes for 50% of a cohort of AI-cited sources to stop appearing in AI answers. Scrunch and Stacker measured a 4.5-week median across 3.5 million citation events, 120,000+ domains, eight industries, and six platforms between September 2025 and March 2026, using survival-curve analysis with 200 bootstrap resamples.
Which AI platform churns citations fastest? ChatGPT, at a 3.4-week half-life. SISTRIX separately measured 74% weekly domain churn on ChatGPT Search, rising to 85% at the URL level. Perplexity holds citations longest at a 5.8-week half-life and a 44% 28-day retention rate.
What is a good Citation Retention Rate? The cross-platform average is 33% over 28 days, per Digital Authority Partners' June 2026 study of 1,127 URLs. Perplexity leads at 44% and Gemini trails at 11%. Anything above 33% on your priority engines puts you ahead of the benchmark.
How often should you refresh content for AI citations? Match the cadence to your priority platform: every 2 to 3 weeks for ChatGPT, monthly for Google AI Mode and AI Overviews, monthly to five weeks for Gemini, and every 6 weeks for Perplexity. Substantive updates count. A new statistic, a corrected claim, and an updated visible timestamp reset the freshness signal without a rewrite.
Does earned media make AI citations last longer? Yes. Content distributed through editorial networks held a half-life close to 10 weeks against 4.5 weeks for single-domain content, a 2.1x advantage that held across six platforms and eight industries. Perplexity plus editorial distribution reached 12.3 weeks.
Do Google rankings predict AI citation durability? No. Moz found 88% of Google AI Mode citations fall outside the organic top 10 across 40,000 queries, and Digital Authority Partners found 60% of AI-cited URLs absent from the top 20 on both Google and Bing. The two systems decay independently.
Ready to see whether your AI citations are holding or churning? Nobori tracks your brand's visibility across ChatGPT, Gemini, Perplexity, Google AI Overviews, and Claude, refreshed daily. See which citations survived the week, which competitors replaced you, and which pages need a refresh. Get daily AI visibility alerts for your brand → nobori.ai
Sources: Scrunch and Stacker, half-life of AI citations, 3.5M citation events (March 2026); Digital Authority Partners, The AI Visibility Gap Study, 1,127 URLs across five platforms (June 2026); SISTRIX AI Research Index, 82,619 prompts and 1,548,213 snapshots across six countries and 17 weeks (May 2026); Seer Interactive, content recency and AI visibility (July 2026); Semrush 2026 AI Visibility Index, 126M U.S. prompts (July 2026); Previsible 2026 State of AI Discovery Report, 6.77M LLM sessions (July 2026); Moz, AI Mode citations across 40,000 queries (2026); European Commission DMA decisions (July 2026).
