Five independent studies — spanning more than 4 million tracked citation events across six AI engines — now agree on a counterintuitive finding: AI search citations are simultaneously volatile and stable, depending on what you measure. At the URL level, engines replace 44–88% of their sources daily. At the domain level, 96.8% show zero change week over week. The difference between those two numbers defines how citation authority actually works.
How much do AI citations actually change? #
The answer depends on the unit of measurement and the time window.
GetMentions analyzed 530,875 citations across 2,398 queries and four engines over seven consecutive days in June 2026. Their finding: day-over-day source churn ranges from 44.4% (Perplexity) to 88.3% (Gemini). Only 0.4% of Gemini's Day 1 sources survived all seven days. Perplexity retained 11.1%.
BrightEdge measured the same phenomenon differently using their AI Catalyst platform during the week of February 1, 2026. At the domain level, 96.8% of cited domains showed zero change. Among the 3.2% that moved, 87% were losses — only 0.4% of all tracked domains gained new citations.
The underlying refresh rate explains the churn. FogTrail's continuous monitoring across hundreds of queries in B2B SaaS, developer tools, and fintech found that all five major engines refresh their citation pools approximately every 48 hours — making weekly monitoring blind to roughly 73% of refresh cycles. Separately, Cite Solutions tracked 37,230 AI answers across 297 daily editions and found the top-cited brand for a given category changed between consecutive editions 23.8% of the time, drawing from a pool of 5,239 distinct domains.
The reconciliation: engines keep going back to the same domains, but they rotate which specific pages from those domains they cite. The domain layer is sticky. The URL layer churns.
The 4.5-week half-life #
A meta-analysis published by Nobori, drawing on the Scrunch/Stacker longitudinal study of 3.5 million citation events across 120,000+ domains and six platforms (September 2025 through March 2026), established the median citation half-life at 4.5 weeks. That means a page cited today has a roughly 50% chance of still being cited a month from now.
The 28-day retention rate varies sharply by engine:
| Engine | 28-Day Retention | Half-Life | Daily Churn (GetMentions) |
|---|---|---|---|
| Perplexity | 44% | 5.8 weeks | 44.4% |
| Microsoft Copilot | 34% | — | — |
| ChatGPT | 31% | 3.4 weeks | 79.2% |
| Google AI Overviews | 27% | 4.7 weeks | — |
| Google AI Mode | — | — | 75.9% |
| Gemini | 11% | 4.8 weeks | 88.3% |
Perplexity retains roughly four times more citations than Gemini over the same period. That gap has practical consequences for content strategy — a page optimized for Perplexity needs updating every six weeks, while the same page targeting Gemini operates on a two-to-three-week refresh cycle.
Why some citations last and others disappear #
Three factors separate durable citations from transient ones.
Citation volume. BrightEdge identified a stability inflection point at 50 citations: domains cited fewer than 50 times experienced roughly 50% weekly volatility. Above 50 citations, volatility dropped to 8%. The gap between the most- and least-volatile domains was 70x.
Content type. The LLMO Framework's 90-day tracking study found that content type dominates engine choice in determining half-life:
| Content Type | ChatGPT Half-Life | Claude Half-Life | Perplexity Half-Life |
|---|---|---|---|
| Evergreen how-to | 6.8 weeks | 7.4 weeks | 9.1 weeks |
| Methodology post | 5.1 weeks | 5.9 weeks | 6.7 weeks |
| Experience report | 3.2 weeks | 3.6 weeks | 4.4 weeks |
Evergreen content persists roughly twice as long as experience reports across every engine measured.
Content freshness. A Lureon analysis aggregating multiple citation-freshness datasets — including the Lantern AI Citation Report (200M+ citations) and an Ahrefs study of 17 million cited URLs — found that 76% of AI citations come from content updated within the previous 30 days. Pages older than 12 months largely exit the citation pool. Established URLs that receive substantive updates integrate into answers 40% faster than newly published pages, suggesting that refresh beats replacement.
Distribution breadth. The Nobori meta-analysis found that content syndicated through editorial networks showed a 2.1x durability advantage: approximately 10 weeks versus 4.5 weeks for single-domain sources. Substantive updates recovered 60–75% of peak citation rates, while cosmetic edits yielded no measurable improvement.
Cross-engine fragmentation makes single-engine strategies unreliable #
The GetMentions study found that 84% of cited sources appear in only one engine. Just 1.1% appeared across all four engines studied over a seven-day window. A separate Digital Authority Partners study of 1,127 URLs confirmed the pattern: maximum overlap between any two platforms was 17%.
This fragmentation means a brand visible in ChatGPT may be invisible in Gemini, and vice versa. The LLMPulse Tremor tracker, which measures week-over-week citation churn across 170,000+ repeated prompts on five engines, currently reads 6.0 on a 10-point scale — "moderate" but persistent, with citation-domain Jaccard distances varying across engine pairs. Multi-engine measurement is not a monitoring luxury — it is a prerequisite for understanding actual AI visibility.
AI search is consolidating, not redistributing #
BrightEdge's week-over-week data revealed a structural pattern: citation losses do not flow to new competitors. When engines drop a source, they tighten their citation radius around fewer trusted domains rather than swapping preferences. Of all citation-level changes in the February 2026 dataset, 87% were losses and only 13% were gains.
Industry-level volatility reinforces this pattern. Finance domains showed 51% volatility, with 91% of those changes being declines. Government and institutional domains experienced only 3.6% volatility. The engines are not experimenting with new sources — they are pruning toward a smaller, more stable citation set.
What this means for Machine Relations #
Citation volatility data from five independent research teams converges on a single operational reality: AI citations are a flow, not a stock. Winning a citation is the beginning of a maintenance obligation, not the end of an optimization project.
Three implications follow directly from the data:
-
Domain authority compounds; page-level optimization churns. The 96.8% domain stability against 44–88% URL churn means institutional credibility — built through consistent, multi-surface publishing — outlasts any individual page's citation window.
-
Refresh cadence is a competitive variable. The 4.5-week median half-life sets a floor: content updated less than monthly loses roughly half its AI visibility each cycle. The engine-specific data suggests ChatGPT requires bi-weekly attention, Google surfaces monthly, and Perplexity every six weeks.
-
Cross-engine fragmentation demands multi-engine measurement. With 84% of citations engine-exclusive, a single-engine tracking approach misses the majority of a brand's AI visibility surface. The Machine Relations Index measures across six engines for this reason — single-engine snapshots systematically undercount or overcount real visibility.
FAQ #
How often do AI search engines change their citations? #
At the URL level, AI engines replace between 44% (Perplexity) and 88% (Gemini) of their cited sources daily. At the domain level, 96.8% remain stable week over week. The median citation half-life across all engines is 4.5 weeks.
Which AI engine has the most stable citations? #
Perplexity consistently shows the highest citation retention across studies: 44% of sources persist at 28 days, compared to 31% for ChatGPT and 11% for Gemini. Perplexity's daily churn rate of 44.4% is roughly half that of other engines.
How often should content be updated to maintain AI citations? #
Engine-specific data suggests different cadences: every two to three weeks for ChatGPT, monthly for Google surfaces, and every six weeks for Perplexity. Content updated less than quarterly loses most of its AI visibility. Substantive updates recover 60–75% of peak citation rates, while cosmetic changes show no measurable improvement.
Does being cited by one AI engine mean you are cited by others? #
No. 84% of cited sources appear in only one engine, and the maximum overlap between any two platforms is 17%. AI engines select sources independently, making cross-engine visibility measurement necessary for accurate assessment.